forza61/academic-rag-data · Datasets at Hugging Face

chunk_id string	paper_id string	title string	section_title string	section_path list	para_index int64	text string	reference_ids list	inline_citations list	year int64	venue null	url string	authors list	references list
Artificial Intelligence_1_abstract_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	Abstract	[ "Abstract" ]	0	Biosignals offer valuable insights into the physiological states of the human body. Although biosignal modalities differ in functionality, signal fidelity, sensor comfort, and cost, they are often intercorrelated, reflecting the holistic and interconnected nature of human physiology. This opens up the possibility of performing the same tasks using alternative biosignal modalities, thereby improving the accessibility, usability, and adaptability of health monitoring systems. However, the limited availability of large labeled datasets presents challenges for training models tailored to specific tasks and modalities of interest. Unsupervised cross-modal knowledge transfer offers a promising solution by leveraging knowledge from an existing modality to support model training for a new modality. Existing methods are typically based on knowledge distillation, which requires running a teacher model alongside student model training, resulting in high computational and memory overhead. This challenge is further exacerbated by the recent development of foundation models that demonstrate superior performance and generalization across tasks at the cost of large model sizes. To this end, we explore a new framework for unsupervised cross-modal knowledge transfer of biosignals by training a lightweight bridge network to align the intermediate representations and enable information flow between foundation models and across modalities. Specifically, we introduce an efficient strategy for selecting alignment positions where the bridge should be constructed, along with a flexible prototype network as the bridge architecture. Extensive experiments across multiple biosignal modalities, tasks, and datasets show that BioX-Bridge reduces the number of trainable parameters by 88-99% while maintaining or even improving transfer performance compared to state-of-the-art methods.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec0_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	INTRODUCTION	[ "INTRODUCTION" ]	0	Biosignals, such as electrocardiogram (ECG), electroencephalogram (EEG), and photoplethysmography (PPG), provide critical insights into the underlying physiological states of individuals. They are essential tools in modern healthcare and have often been considered the gold standard for diagnostics (Rosenberg & Van Hout, 2013;Stracina et al., 2022). In the past decade, the advancement of artificial intelligence (AI) has enabled remarkable capabilities in automated diagnostics and monitoring, such as stress assessment (Mentis et al., 2024), sleep stage classification (Mostafa et al., 2019), and arrhythmia detection (Parvaneh et al., 2019). However, many biosignal sensors are not suitable for use outside clinical settings due to factors such as user discomfort, high manufacturing costs, and excessive power consumption.	[]	[ { "raw": "Rosenberg & Van Hout, 2013", "authors": "Rosenberg & Van Hout", "year": "2013" }, { "raw": "Stracina et al., 2022", "authors": "Stracina et al.", "year": "2022" }, { "raw": "Mentis et al., 2024", "authors": "Mentis et al.", "year": "2024" }, { "raw": "Mostafa et al., 2019", "authors": "Mostafa et al.", "year": "2019" }, { "raw": "Parvaneh et al., 2019", "authors": "Parvaneh et al.", "year": "2019" } ]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec0_p1	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	INTRODUCTION	[ "INTRODUCTION" ]	1	A promising direction is to harness the correlations between different biosignal modalities and perform the same tasks using alternative modalities, making health monitoring systems more accessible, practical, and flexible (Wang et al., 2023;Yang et al., 2023). For example, being able to perform the same tasks using single-lead PPG data from a wearable smartwatch, instead of relying on 12-lead ECGs, would greatly reduce hardware complexity and cost, while enabling continuous, user-friendly monitoring in everyday environments. Unfortunately, training such models requires large-scale labeled datasets, which are often difficult to obtain in biosignal applications due to the high cost and domain-specific expertise required for data collection and annotation. This highlights the need for effective knowledge transfer between biosignal modalities, leveraging models trained in old or well-established modalities to support the development of models for new or underrepresented biosignal modalities.	[]	[ { "raw": "Wang et al., 2023", "authors": "Wang et al.", "year": "2023" }, { "raw": "Yang et al., 2023", "authors": "Yang et al.", "year": "2023" } ]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec0_p2	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	INTRODUCTION	[ "INTRODUCTION" ]	2	Unsupervised cross-modal knowledge transfer stands out as a practical solution to address the aforementioned needs. Existing methods can be divided into two categories: data translation and knowledge distillation. As illustrated in Figure 1(a), data translation directly translates data from the new modality to the old modality, enabling the direct reuse of existing models from the old modality (Sarkar & Etemad, 2021). However, the exploration of data translation has been limited to a certain pair of modalities, such as PPG and ECG. Figure 1(b) illustrates knowledge distillation, which seeks to train a student model for the new modality to mimic the output of a pre-trained teacher model from the old modality (Abbaspourazad et al., 2024b;Zhang et al., 2024). The distillation process is memory intensive, as it requires forward inference with both the student and teacher models, in addition to backpropagation with the student model. The computational burden is further exacerbated by the emergence of large-scale biosignal foundation models (Coppola et al., 2024;Jiang et al., 2024;Pillai et al., 2025), which are mostly trained on specific modalities and have demonstrated exceptional performance across a wide range of tasks. Although these models offer tremendous performance gains, their use in cross-modal knowledge transfer is hindered by their size, which makes traditional knowledge distillation solutions computationally prohibitive for users without access to high-end GPUs. For example, distilling knowledge from PaPaGei, a PPG-based foundation model (Pillai et al., 2025), to the ECG-FM student model (McKeen et al., 2024) on the WESAD dataset (Schmidt et al., 2018), with a batch size of eight, requires more than 32GB of VRAM. Furthermore, because of data sharing regulations and privacy concerns, such a distillation process often needs to be performed locally where the data resides, under low-resource conditions. These constraints call for the development of an effective and efficient cross-modal transfer framework that can fully leverage the representation capability and embedded knowledge of the foundation models.	[]	[ { "raw": "Sarkar & Etemad, 2021", "authors": "Sarkar & Etemad", "year": "2021" }, { "raw": "Abbaspourazad et al., 2024b", "authors": "Abbaspourazad et al.", "year": "2024b" }, { "raw": "Zhang et al., 2024", "authors": "Zhang et al.", "year": "2024" }, { "raw": "Coppola et al., 2024", "authors": "Coppola et al.", "year": "2024" }, { "raw": "Jiang et al., 2024", "authors": "Jiang et al.", "year": "2024" }, { "raw": "Pillai et al., 2025", "authors": "Pillai et al.", "year": "2025" }, { "raw": "Pillai et al., 2025", "authors": "Pillai et al.", "year": "2025" }, { "raw": "McKeen et al., 2024", "authors": "McKeen et al.", "year": "2024" }, { "raw": "Schmidt et al., 2018", "authors": "Schmidt et al.", "year": "2018" } ]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec0_p3	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	INTRODUCTION	[ "INTRODUCTION" ]	3	To this end, we propose BioX-Bridge, a new framework for unsupervised cross-modal knowledge transfer via model bridging, as illustrated in Figure 1(c). The core idea is to construct a bridge that projects intermediate representations from one biosignal model to another, leveraging the powerful representational capability and the rich embedded knowledge of foundation models. 1 The framework comprises two key components: bridge position selection and bridge architecture design. Specifically, we introduce an efficient two-stage strategy for selecting optimal input and output positions by evaluating the quality and similarity of intermediate representations between two biosignal models.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec0_p4	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	INTRODUCTION	[ "INTRODUCTION" ]	4	To enable effective projection between high-dimensional spaces, we design a prototype network composed of a learnable prototype set and a low-rank approximation module to compute aggregation weights. Notably, only the bridge network requires training to enable interoperability between models of different modalities. We evaluate the effectiveness of BioX-Bridge in three biosignal datasets involving different modalities, demonstrating superior efficiency compared to existing methods.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec0_p5	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	INTRODUCTION	[ "INTRODUCTION" ]	5	Extensive ablation studies further confirm the robustness of the proposed framework under various conditions. Our contributions can be summarized as follows:	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec0_p6	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	INTRODUCTION	[ "INTRODUCTION" ]	6	• We propose BioX-Bridge, a novel unsupervised model bridging framework that enables crossmodal knowledge transfer through information flow between biosignal models.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec0_p7	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	INTRODUCTION	[ "INTRODUCTION" ]	7	• We introduce key components to support the framework, including an efficient two-stage strategy for selecting bridge positions and a prototype network with low-rank approximation for effective high-dimensional projection.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec0_p8	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	INTRODUCTION	[ "INTRODUCTION" ]	8	• We demonstrate the efficiency of BioX-Bridge through experiments on three biosignal datasets, four modalities, and six transfer directions, demonstrating robustness through comprehensive ablation studies.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec1_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	RELATED WORKS	[ "RELATED WORKS" ]	0	Unsupervised Cross-modal Knowledge Transfer Existing methods can be divided into two categories: knowledge distillation and data translation.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec1_p1	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	RELATED WORKS	[ "RELATED WORKS" ]	1	Knowledge distillation was introduced as a model compression technique, where a smaller student model learns to mimic a larger and high-performing teacher model by matching its output distributions (Hinton et al., 2015). The concept has since been extended to cross-modal knowledge transfer. Early efforts focused on computer vision applications across a variety of sensor modalities, such as vision to depth images (Garcia et al., 2018;Gupta et al., 2016;Hoffman et al., 2016;Tian et al., 2020), to radio frequency heatmaps (Zhao et al., 2018), and to sound (Aytar et al., 2016;Xue et al., 2021). The core idea is to leverage unlabeled but semantically aligned data pairs to bridge the modality gap and transfer relevant knowledge to the corresponding tasks (Gou et al., 2021;Moslemi et al., 2024). Recent efforts have also investigated cross-modal knowledge distillation for biosignals. For example, Brant-X (Zhang et al., 2024) introduced a unified biosignal alignment framework that transfers knowledge from EEG to other biosignal modalities through a two-level semantic alignment strategy, such that the student model can provide complementary representations to the teacher model and improve downstream task performance. In another work (Abbaspourazad et al., 2024b), the distillation of knowledge from PPG to accelerometer signals was used to accurately predict physiological states such as heart rate. However, the aforementioned methods require training a full-size student model from scratch, which becomes increasingly impractical as model sizes grow, especially for resource-constrained settings.	[]	[ { "raw": "Hinton et al., 2015", "authors": "Hinton et al.", "year": "2015" }, { "raw": "Garcia et al., 2018", "authors": "Garcia et al.", "year": "2018" }, { "raw": "Gupta et al., 2016", "authors": "Gupta et al.", "year": "2016" }, { "raw": "Hoffman et al., 2016", "authors": "Hoffman et al.", "year": "2016" }, { "raw": "Tian et al., 2020", "authors": "Tian et al.", "year": "2020" }, { "raw": "Zhao et al., 2018", "authors": "Zhao et al.", "year": "2018" }, { "raw": "Aytar et al., 2016", "authors": "Aytar et al.", "year": "2016" }, { "raw": "Xue et al., 2021", "authors": "Xue et al.", "year": "2021" }, { "raw": "Gou et al., 2021", "authors": "Gou et al.", "year": "2021" }, { "raw": "Moslemi et al., 2024", "authors": "Moslemi et al.", "year": "2024" }, { "raw": "Zhang et al., 2024", "authors": "Zhang et al.", "year": "2024" }, { "raw": "Abbaspourazad et al., 2024b", "authors": "Abbaspourazad et al.", "year": "2024b" } ]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec1_p2	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	RELATED WORKS	[ "RELATED WORKS" ]	2	Data translation aims to achieve unsupervised cross-modal knowledge transfer by directly translating raw data from one modality to another. Generative adversarial networks (GAN) (Goodfellow et al., 2020) and their variants (Mirza & Osindero, 2014;Zhu et al., 2017) have been widely adopted for modality translation tasks in the visual and signal processing domains (Duan et al., 2021;Sikka et al., 2021;Yang et al., 2020). A recent work (Wang et al., 2023) leveraged knowledge graphs to learn transformations between independently trained foundation models for proteins, drugs, and text. Nevertheless, reliance on structured knowledge graphs limits their applicability to biosignal scenarios, where such structured relationships are scarce or nonexistent. In the biosignal domain, cross-modal translation efforts have been largely limited to translation from PPG to ECG (Sarkar & Etemad, 2021;Zhu et al., 2021). Extending such translation to other modalities, such as EEG to ECG, remains largely underexplored.	[]	[ { "raw": "Goodfellow et al., 2020", "authors": "Goodfellow et al.", "year": "2020" }, { "raw": "Mirza & Osindero, 2014", "authors": "Mirza & Osindero", "year": "2014" }, { "raw": "Zhu et al., 2017", "authors": "Zhu et al.", "year": "2017" }, { "raw": "Duan et al., 2021", "authors": "Duan et al.", "year": "2021" }, { "raw": "Sikka et al., 2021", "authors": "Sikka et al.", "year": "2021" }, { "raw": "Yang et al., 2020", "authors": "Yang et al.", "year": "2020" }, { "raw": "Wang et al., 2023", "authors": "Wang et al.", "year": "2023" }, { "raw": "Sarkar & Etemad, 2021", "authors": "Sarkar & Etemad", "year": "2021" }, { "raw": "Zhu et al., 2021", "authors": "Zhu et al.", "year": "2021" } ]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec1_p3	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	RELATED WORKS	[ "RELATED WORKS" ]	3	Biosignals Foundation Models Inspired by the recent success of large-scale pre-training in natural language processing (Achiam et al., 2023) and computer vision (Dosovitskiy et al., 2020), the development of foundation models for biosignals has garnered much interest (Han et al., 2024;Lai et al., 2025). Through large-scale self-supervised training on public and private biosignal datasets, several biosignal foundation models have been developed to capture rich and transferable representations, enabling more robust and efficient downstream adaptation. These models span a variety of modalities, including EEG (Chen et al., 2025a;2024;Cui et al., 2023;Jiang et al., 2024;Wang et al., 2024), ECG (Coppola et al., 2024;Li et al., 2024;McKeen et al., 2024), PPG (Chen et al., 2025b;Pillai et al., 2025;Saha et al., 2025), accelerometer (Abbaspourazad et al., 2024b), and general-purpose biosignal models (Yang et al., 2023). In addition to unimodal models, recent work has explored multimodal foundation models for various applications such as health monitoring (Abbaspourazad et al., 2024a;Luo et al., 2024), sleep (Thapa et al., 2024), and activity recognition (Narayanswamy et al., 2024). Despite their strong performance on data from modalities seen during At the inference stage, the bridge has been constructed and enables the flow of information between the two models in order to make predictions on data from the new modality. (c) The bridge consists of a low-rank approximation module and a prototype set. The low-rank approximation module generates aggregation weights for the prototype vectors.	[]	[ { "raw": "Achiam et al., 2023", "authors": "Achiam et al.", "year": "2023" }, { "raw": "Dosovitskiy et al., 2020", "authors": "Dosovitskiy et al.", "year": "2020" }, { "raw": "Han et al., 2024", "authors": "Han et al.", "year": "2024" }, { "raw": "Lai et al., 2025", "authors": "Lai et al.", "year": "2025" }, { "raw": "Chen et al., 2025a", "authors": "Chen et al.", "year": "2025a" }, { "raw": "Cui et al., 2023", "authors": "Cui et al.", "year": "2023" }, { "raw": "Jiang et al., 2024", "authors": "Jiang et al.", "year": "2024" }, { "raw": "Wang et al., 2024", "authors": "Wang et al.", "year": "2024" }, { "raw": "Coppola et al., 2024", "authors": "Coppola et al.", "year": "2024" }, { "raw": "Li et al., 2024", "authors": "Li et al.", "year": "2024" }, { "raw": "McKeen et al., 2024", "authors": "McKeen et al.", "year": "2024" }, { "raw": "Chen et al., 2025b", "authors": "Chen et al.", "year": "2025b" }, { "raw": "Pillai et al., 2025", "authors": "Pillai et al.", "year": "2025" }, { "raw": "Saha et al., 2025", "authors": "Saha et al.", "year": "2025" }, { "raw": "Abbaspourazad et al., 2024b", "authors": "Abbaspourazad et al.", "year": "2024b" }, { "raw": "Yang et al., 2023", "authors": "Yang et al.", "year": "2023" }, { "raw": "Abbaspourazad et al., 2024a", "authors": "Abbaspourazad et al.", "year": "2024a" }, { "raw": "Luo et al., 2024", "authors": "Luo et al.", "year": "2024" }, { "raw": "Thapa et al., 2024", "authors": "Thapa et al.", "year": "2024" }, { "raw": "Narayanswamy et al., 2024", "authors": "Narayanswamy et al.", "year": "2024" } ]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec1_p4	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	RELATED WORKS	[ "RELATED WORKS" ]	4	pre-training, these models struggle with generalization to unseen modalities due to mismatches in input dimensions and data distributions (Liu et al., 2024).	[]	[ { "raw": "Liu et al., 2024", "authors": "Liu et al.", "year": "2024" } ]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec1_p5	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	RELATED WORKS	[ "RELATED WORKS" ]	5	We provide further discussion on model stitching and domain adaptation in Appendix C.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec2_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	METHODS	[ "METHODS" ]	0	The core concept of our proposed BioX-Bridge framework is to build a bridging network that facilitates efficient and effective projection between intermediate representations of biosignal models. This allows the framework to harness the strong representational power of one model while integrating the task-specific knowledge contained in another. We define the problem in Section 3.1 and introduce the idea of model bridging in Section 3.2. We detail the position of the bridge, its architecture and training in Sections 3.3-3.5. An overview of BioX-Bridge is presented in Figure 2.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec3_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	PROBLEM DEFINITION	[ "PROBLEM DEFINITION" ]	0	Assume that we are given an annotated dataset from an old biosignal modality, old) \| labeled samples for a specific task, and a corresponding model,	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec3_p1	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	PROBLEM DEFINITION	[ "PROBLEM DEFINITION" ]	1	, where f (old) θ is a pre-trained encoder parametrized by θ followed by a task head g (old) ω parametrized by ω. We also have an un-annotated dataset from a new modality,	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec3_p2	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	PROBLEM DEFINITION	[ "PROBLEM DEFINITION" ]	2	, which shares the same underlying label set with D (old) . We further have a disjoint, un-annotated paired dataset D (pair) = {(x	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec3_p3	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	PROBLEM DEFINITION	[ "PROBLEM DEFINITION" ]	3	. The unsupervised cross-modal knowledge transfer problem aims to obtain a model f , such that f can make predictions on D (new) .	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec4_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	MODEL BRIDGING	[ "MODEL BRIDGING" ]	0	Let f (old) θ be the model for the old modality, parametrized by θ of L layers, and let f (new) ϕ be the model for the new modality, parametrized by ϕ of M layers. The intermediate representations from the m-th layer of the new modality model can then be extracted as:	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec4_p1	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	MODEL BRIDGING	[ "MODEL BRIDGING" ]	1	where x (new) denotes a biosignal time series sample from the new modality. f (new) ϕ ≤m denotes the subset of the new modality model consisting of its first m layers, subject to the constraint Next, we introduce a bridge network to enable the information flow between the new and old modality models by projecting representations from the new modality into the representation space of the old modality:	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec4_p2	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	MODEL BRIDGING	[ "MODEL BRIDGING" ]	2	where b ψ denotes the bridge network parametrized by ψ. h(old) l denotes the projected representation from the new modality to the old modality. Note that the projected representation is designed to mimic the intermediate representation from the l-th layer of the old modality model, defined as h (old) old) , where x (old) is the paired input signal from the old modality. Thus,	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec4_p3	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	MODEL BRIDGING	[ "MODEL BRIDGING" ]	3	Finally, we can obtain predictions using the old modality model starting from the (l + 1)-th layer:	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec4_p4	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	MODEL BRIDGING	[ "MODEL BRIDGING" ]	4	where m and l are also known as the bridge input and output positions, • denotes function composition.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec5_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	BRIDGE POSITION SELECTION	[ "BRIDGE POSITION SELECTION" ]	0	There are L × M possible locations where the bridge can be constructed between the layers of the two models. Although a brute-force search would yield the optimal bridge position, it is computationally expensive. In particular, the choice of the bridge position is one of the most influential factors affecting transfer performance, as we will show in ablation studies. To this end, we propose a two-stage strategy for efficient bridge position selection, as illustrated in Figure 3.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec5_p1	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	BRIDGE POSITION SELECTION	[ "BRIDGE POSITION SELECTION" ]	1	Stage 1: Bridge Input Position (m) Selection The bridge serves to project new modality representations to the old modality representation space, enabling the bridged model to mimic the behavior of the old modality model. As the saying "garbage in, garbage out" suggests, it is important to select discriminative new modality representations that can effectively distinguish among the predictions produced by the old modality model, also known as pseudo labels. We propose to select the input position of the bridge by linear probing, which has been widely used to evaluate the quality of intermediate representations (Alain & Bengio, 2016). The bridge input position selection can be formulated as:	[]	[ { "raw": "Alain & Bengio, 2016", "authors": "Alain & Bengio", "year": "2016" } ]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec5_p2	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	BRIDGE POSITION SELECTION	[ "BRIDGE POSITION SELECTION" ]	2	where ) denotes the pseudo label. L probe denotes the empirical loss for the linear prober.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec5_p3	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	BRIDGE POSITION SELECTION	[ "BRIDGE POSITION SELECTION" ]	3	, we can ease the transformation process by selecting h (old) l to be as similar as possible to h (new) m . We select linear CKA (Kornblith et al., 2019)	[]	[ { "raw": "Kornblith et al., 2019", "authors": "Kornblith et al.", "year": "2019" } ]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec5_p4	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	BRIDGE POSITION SELECTION	[ "BRIDGE POSITION SELECTION" ]	4	For detailed formulation of CKA linear , please refer to the appendix.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec5_p5	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	BRIDGE POSITION SELECTION	[ "BRIDGE POSITION SELECTION" ]	5	Algorithm 1: BioX-Bridge learning procedure	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec5_p6	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	BRIDGE POSITION SELECTION	[ "BRIDGE POSITION SELECTION" ]	6	Input: Old modality model f (old) θ ; New modality model f (new) ϕ ; Task head g (old) ω ; Paired dataset	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec5_p7	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	BRIDGE POSITION SELECTION	[ "BRIDGE POSITION SELECTION" ]	7	Select bridge input position, m, using Eq. ( 4) Select bridge output position, l, using Eq. ( 5)	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec5_p8	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	BRIDGE POSITION SELECTION	[ "BRIDGE POSITION SELECTION" ]	8	using Eq. ( 7) whose intermediate representation exhibits the strongest linear association with the pseudolabels. For bridge output position, we select the layer from f (old) θ whose representation is most similar to that of the bridge input layer.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec6_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	BRIDGE ARCHITECTURE	[ "BRIDGE ARCHITECTURE" ]	0	Models of different modalities operate in distinct representational spaces. The bridge network should be sufficiently parametrized to enable the projection and alignment of the two spaces. A naive bridge architecture is a full-rank linear layer, but this is prohibitively expensive because of the highdimensional projection from the new modality to the old modality. For example, using LaBraM (Jiang et al., 2024) as f (new) ϕ and HuBERT-ECG (Coppola et al., 2024) as f (old) θ , the projection would require	[]	[ { "raw": "Jiang et al., 2024", "authors": "Jiang et al.", "year": "2024" }, { "raw": "Coppola et al., 2024", "authors": "Coppola et al.", "year": "2024" } ]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec6_p1	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	BRIDGE ARCHITECTURE	[ "BRIDGE ARCHITECTURE" ]	1	7 billion parameters. To address the challenge of high-dimensional projection, we propose a prototype network. The prototype network consists of two modules, a prototype set, and a low-rank approximation module. Specifically, the prototype set, P ∈ R Np×d (old) l , consisting of N p prototype vectors with embedding dimension d (new) m , introduces the flexibility to incorporate prior knowledge from f Np , reduces the number of trainable parameters through a low-rank factorization, while generating aggregation weights for prototype vectors, as illustrated in Figure 2.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec6_p2	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	BRIDGE ARCHITECTURE	[ "BRIDGE ARCHITECTURE" ]	2	where Pool(•) denotes a pooling operation along the N (new) m dimension, and Reshape N (old) l ×Np (•) denotes the reshape operation to the specified output dimensions.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec7_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	BRIDGE TRAINING	[ "BRIDGE TRAINING" ]	0	As the difference between h (old) l and h(old) l approaches zero, the bridged model yields predictions identical to those of the old modality model. Formally:	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec7_p1	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	BRIDGE TRAINING	[ "BRIDGE TRAINING" ]	1	Naturally, the training objective for the bridge network is to align the intermediate representations in the L-th layer2 :	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec7_p2	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	BRIDGE TRAINING	[ "BRIDGE TRAINING" ]	2	where L align denotes the loss function, such as cosine loss and mean absolute error loss. The learning process of BioX-Bridge is presented in Algorithm 1. Given the unbalanced nature of the datasets, we report Balanced Accuracy, F1-Weighted, and F1-Macro, following (Jiang et al., 2024;Pillai et al., 2025).	[]	[ { "raw": "Jiang et al., 2024", "authors": "Jiang et al.", "year": "2024" }, { "raw": "Pillai et al., 2025", "authors": "Pillai et al.", "year": "2025" } ]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec8_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	EXPERIMENTS	[ "EXPERIMENTS" ]	0	Backbone Foundation Models For EEG, we adopt the base version of the LaBraM architecture with 5.8M parameters (Jiang et al., 2024). For ECG, we adopt the small version of the HuBERT-ECG architecture with 30.4M parameters (Coppola et al., 2024). For PPG, we adopt the small version of the PaPaGei architecture with 5.7M parameters (Pillai et al., 2025). For EMG, we adopt NormWear with 136.1M parameters (Luo et al., 2024). Note that LaBraM, HuBERT-ECG, and NormWear adopt a CNN-transformer architecture, while PaPaGei adopts a CNN architecture. All models are initialized with the pre-trained weights provided by the original publications. Note that biosignal foundation models are still early in their development, in comparison to foundation models for language and vision. While current models contain a relatively small number of parameters, our method for efficient cross-modal knowledge transfer would be even more valuable as they scale up.	[]	[ { "raw": "Jiang et al., 2024", "authors": "Jiang et al.", "year": "2024" }, { "raw": "Coppola et al., 2024", "authors": "Coppola et al.", "year": "2024" }, { "raw": "Pillai et al., 2025", "authors": "Pillai et al.", "year": "2025" }, { "raw": "Luo et al., 2024", "authors": "Luo et al.", "year": "2024" } ]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec8_p1	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	EXPERIMENTS	[ "EXPERIMENTS" ]	1	Baselines We compare our method with the following baselines evaluated on D (new) : (i) Random denotes a model that produces predictions at random. (ii) CardioGAN uses GAN to synthesize ECG from PPG (Sarkar & Etemad, 2021), and we translate the new modality data (PPG) to the old modality (ECG) for evaluation. (iii) KD (Hinton et al., 2015) is the baseline of knowledge distillation. (iv) KD-contrast (Abbaspourazad et al., 2024b) is a variant of knowledge distillation with contrast loss (Zhang et al., 2024). (v) Oracle denotes the absolute best performance that can be achieved, which is simply the performance of the old modality model using old modality data. Please refer to the appendix for implementation details and further discussions.	[]	[ { "raw": "Sarkar & Etemad, 2021", "authors": "Sarkar & Etemad", "year": "2021" }, { "raw": "Hinton et al., 2015", "authors": "Hinton et al.", "year": "2015" }, { "raw": "Abbaspourazad et al., 2024b", "authors": "Abbaspourazad et al.", "year": "2024b" }, { "raw": "Zhang et al., 2024", "authors": "Zhang et al.", "year": "2024" } ]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec9_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER PERFORMANCE	[ "UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER PERFORMANCE" ]	0	Experiment results on the ISRUC, FOG, and WESAD dataset are presented in Table 1. We observe that BioX-Bridge significantly reduces the number of trainable parameters by 87.9-99.1% and continues to achieve performance comparable to or better than that of the baseline methods. For example, for WESAD (PPG → ECG), BioX-Bridge requires merely 1.3% of trainable parameters while outperforming the baseline methods by around 1-2% across all metrics.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec9_p1	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER PERFORMANCE	[ "UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER PERFORMANCE" ]	1	We also observe that the knowledge transfer performance gap compared to Oracle varies across datasets and knowledge transfer directions. For example, on the ISRUC dataset, we observe approximately 20% balanced accuracy gap between BioX-Bridge (60.11%) and Oracle ( On another note, BioX-Bridge and KD-Contrast achieved higher balanced accuracy than Oracle on several occasions, respectively. This is possible because balanced accuracy is simply an average of recall across classes. Higher balanced accuracy scores and lower F1 scores reflect that knowledge transfer methods achieved better recall but worse precision than Oracle.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec10_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	ABLATION STUDIES	[ "ABLATION STUDIES" ]	0	We conduct ablation studies on the WESAD dataset and the direction of knowledge transfer (PPG → ECG). Additional results are presented in the appendix.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec11_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	Bridge Rank and Prototype Set	[ "Bridge Rank and Prototype Set" ]	0	We study the impact of different hyperparameters for the prototype network in Figures 4a and4b. A performance drop is observed when the approximation rank and prototype set size are too small or too large, likely due to under-/over-parameterization of the bridge network. In particular, the performance peaks at around 0.75M parameters in both cases.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec11_p1	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	Bridge Rank and Prototype Set	[ "Bridge Rank and Prototype Set" ]	1	Dataset Size We reduce the size of the paired dataset for bridge training. We observe in Figure 4c that the transfer performance slowly decays by around 2% at 20% dataset size, showcasing the robustness of the bridge under the low-data regime. Bridge Position Selection To show that the bridge position selection strategy proposed in Section 3.3 is effective, we compare the unsupervised performance of cross-modal knowledge transfer at various positions in Table 2. For "Fixed", we train nine bridges in predefined positions, which is a combination of the first, middle, and last layers of the old and new modality models. The results are the average over all nine positions in five seeds (see more results in the appendix).	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec12_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	Foundation Model	[ "Foundation Model" ]	0	We further analyze the impact of using different foundation models for crossmodal knowledge transfer. In Table 3, we replace the HuBERT-ECG foundation model (Coppola et al., 2024) with ECG-FM (McKeen et al., 2024). Notably, due to the large number of trainable parameters (90M), the knowledge distillation methods with ECG-FM could only be performed with a batch size of 4 on a V100 GPU. As a result, training for more than 50 epochs requires 6.5 hours for knowledge distillation methods and 1.9 hours for BioX-Bridge. Moreover, the performance gap between knowledge distillation methods and BioX-Bridge is much pronounced at 10-17%.	[]	[ { "raw": "Coppola et al., 2024", "authors": "Coppola et al.", "year": "2024" }, { "raw": "McKeen et al., 2024", "authors": "McKeen et al.", "year": "2024" } ]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec13_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	CONCLUSION	[ "CONCLUSION" ]	0	We present BioX-Bridge as an efficient framework for unsupervised cross-modal knowledge transfer across biosignals. To address the challenges of high-dimensional projection between biosignal foundation models, we design a prototype-based architecture for parameter-efficient learning of transformations between representation spaces. Our proposed two-stage bridge position selection strategy further identifies connection points that enable more effective alignment of intermediate representations. Through extensive experiments on diverse biosignal datasets and tasks, we demonstrated that BioX-Bridge achieves performance comparable to or superior to that of state-of-the-art methods while drastically reducing the number of trainable parameters. This work highlights the potential of model bridging as a powerful alternative to conventional cross-modal knowledge transfer techniques, offering a pathway to more accessible, adaptable, modality-agnostic, and resource-efficient biosignal applications in real-world settings, where computing resources and labelled data are often limited.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec14_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	ETHICAL AND REPRODUCIBILITY STATEMENT	[ "ETHICAL AND REPRODUCIBILITY STATEMENT" ]	0	This study makes use of datasets involving human subjects (ISRUC, WESAD, and FOG). All datasets employed are publicly available, and we follow the usage terms and ethical guidelines specified by the original data providers. No new data were collected for this work, and all analyses were conducted on de-identified, previously published datasets.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec14_p1	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	ETHICAL AND REPRODUCIBILITY STATEMENT	[ "ETHICAL AND REPRODUCIBILITY STATEMENT" ]	1	To ensure reproducibility, we provide detailed descriptions of our experimental setups, including data preprocessing steps, model architectures, hyperparameters, and training procedures, in Section 4.1 and Appendix D. Our code implementation is available in the supplementary materials and will be available in a dedicated repository upon publication.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec15_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	APPENDIX TABLE OF CONTENTS	[ "APPENDIX TABLE OF CONTENTS" ]	0	A l denote the matrix of old modality representations from the l-th layer. The CKA linear operator introduced in Eq. 5 is formulated as follows (Kornblith et al., 2019):	[]	[ { "raw": "Kornblith et al., 2019", "authors": "Kornblith et al.", "year": "2019" } ]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec15_p1	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	APPENDIX TABLE OF CONTENTS	[ "APPENDIX TABLE OF CONTENTS" ]	1	HSIC is the Hilbert-Schmidt Independence Criterion. H is the centering matrix. Note that while this formulation uses the entire H (new) m and H (old) l to compute similarity between representations of the old and new modalities, it is also possible to improve efficiency by computing similarity using only a subset of their rows.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec16_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	B LIMITATIONS AND FUTURE WORK	[ "B LIMITATIONS AND FUTURE WORK" ]	0	Although BioX-Bridge greatly reduces training computational requirements and improves the efficiency of cross-modal knowledge transfer, it depends on the availability of pre-trained models for each biosignal modality, an assumption that may not hold for emerging or underexplored biosignals. Furthermore, depending on the position of the bridge, the inference time of the bridged model could Each foundation model specifies its preprocessing pipeline, and we follow these procedures accordingly. If a notch or bandpass filter has already been applied to the dataset, we skip that step during preprocessing.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec16_p1	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	B LIMITATIONS AND FUTURE WORK	[ "B LIMITATIONS AND FUTURE WORK" ]	1	WESAD In this dataset, ECG signals are sampled at 700Hz and PPG signals at 64Hz. For ECG, which uses HuBERT-ECG as its foundation model, we first downsample to 500Hz, apply a finite impulse response (FIR) bandpass filter between 0.05-47Hz, resample to 100Hz, and then perform channel-wise z-score normalization. For PPG, which uses PaPaGei as its foundation model, we upsample the signals to 125Hz, apply a 4th-order Chebyshev bandpass filter between 0.5-12Hz, and normalize using z-score normalization. Finally, all recordings are segmented into 60-second windows with a 5-second step size.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec16_p2	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	B LIMITATIONS AND FUTURE WORK	[ "B LIMITATIONS AND FUTURE WORK" ]	2	FOG In this dataset, both EEG and EMG signals were collected at 1000Hz, downsampled to 500Hz with a notch and bandpass filter already applied. For EEG, which uses LaBraM as its foundation model, we downsample to 200Hz and convert the unit to 0.1mV. For EMG, which uses NormWear as its foundation model, we downsample to 130Hz and normalize using z-score normalization. All recordings are segmented into 3-second windows with a sliding step size of 0.3 seconds.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec16_p3	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	B LIMITATIONS AND FUTURE WORK	[ "B LIMITATIONS AND FUTURE WORK" ]	3	ISRUC In this dataset, EEG signals are sampled at 200Hz with a notch and bandpass filter already applied, while PPG signals are also sampled at 200Hz with a notch filter applied. For EEG, which uses LaBraM as its foundation model, no resampling is required; we only convert the unit to 0.1mV. For ECG, which uses HuBERT-ECG as its foundation model, we upsample to 500Hz, apply a bandpass filter between 0.05-47Hz, resample to 100Hz, and normalize using z-score normalization.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec17_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	D.1.2 DATASET SPLIT	[ "D.1.2 DATASET SPLIT" ]	0	Dataset split is summarized in Figure A1. The datasets contain synchronized data from the old and new modalities. We perform a subject-wise split for WESAD and ISRUC and sample-wise split for FOG (Zhang et al., 2022) to obtain four subsets D (old) , D (new) , D (val) , and D (pair) , at a ratio of 33%, 22%, 11%, and 33%, respectively. We use old modality data from D (old) to train the linear prober g (old) ω . New modality data from D (new) is used to evaluate bridge performance in an unseen set. All data from D (pair) and D (val) are used to train and help select hyperparameters.	[]	[ { "raw": "Zhang et al., 2022", "authors": "Zhang et al.", "year": "2022" } ]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec18_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	New modality data	[ "New modality data" ]	0	Old modality data	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec20_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	≈33% ≈22%	[ "≈33% ≈22%" ]	0	Evaluate Unused	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec21_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	≈11%	[ "≈11%" ]	0	Hyperparameter Selection Unsupervised Training Figure A1: Illustration of Dataset Split. The dataset is divided into four subject-independent subsets. We first use the old modality data from D (old) to train the linear prober g (old) ω for experiment setup, followed by unsupervised training on D (pair) . The subsets D (val) and D (new) are used for hyperparameter selection and testing, respectively.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec22_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	D.1.3 BIOX-BRIDGE AND BASELINE IMPLEMENTATION DETAILS	[ "D.1.3 BIOX-BRIDGE AND BASELINE IMPLEMENTATION DETAILS" ]	0	BioX-Bridge Implementation Details To prepare f (old) θ and g (old) ω for evaluation, we adapt the pre-trained foundation models for the classification tasks using D (old) , we apply mean pooling to the last-layer representations and add a linear layer for classification. The weights of the foundation model are frozen, and the linear layer is trained for more than 50 epochs. We select a learning rate of 1e-4 for LaBraM, 1e-4 for PaPaGei, 1e-4 for NormWear, and 1e-5 for HuBERT-ECG, as suggested in the original publications. To select the input position of the bridge, we use logistic regression with L2 regularization as the linear prober. We select the layer with the best F1 macro score over five-fold cross-validation. We train the bridge over 50 epochs with cosine embedding loss on D (pair) . All experiments are conducted on V100 GPUs with 32GB VRAM. The training time for a single run ranges from 10 minutes to 4 hours, depending on the dataset and the backbone foundation model.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec22_p1	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	D.1.3 BIOX-BRIDGE AND BASELINE IMPLEMENTATION DETAILS	[ "D.1.3 BIOX-BRIDGE AND BASELINE IMPLEMENTATION DETAILS" ]	1	To select the bridge rank and number of prototypes, we perform a grid search over r = {4, 8, 16, 32} and N p = {50, 100, 150, 200, 250, 300}. For compatibility across various network architectures without modifying the forward functions, we retrieve and replace intermediate representations using forward hooks and forward pre-hooks3 . Please see the source code for detailed implementation.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec22_p2	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	D.1.3 BIOX-BRIDGE AND BASELINE IMPLEMENTATION DETAILS	[ "D.1.3 BIOX-BRIDGE AND BASELINE IMPLEMENTATION DETAILS" ]	2	Baseline Implementation Details For the Random baseline, we simulate a model that randomly assigns labels to samples uniformly across all classes. For CardioGAN, we adopt the pre-trained weights provided by the original publication. We do not perform any fine-tuning as CardioGAN has been trained with WESAD as one of its datasets. We adopt the preprocessing code provided by CargioGAN to prepare the PPG. For each recording, we generate the corresponding ECG using a sliding window of 4 seconds and a step size of 4 seconds. The generated ECG recording is then preprocessed following the HuBERT-ECG pipeline described in Section D.1.1.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec22_p3	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	D.1.3 BIOX-BRIDGE AND BASELINE IMPLEMENTATION DETAILS	[ "D.1.3 BIOX-BRIDGE AND BASELINE IMPLEMENTATION DETAILS" ]	3	For knowledge distillation (KD), we append a linear layer to map mean-pooled representations from the last layer of the foundation model to classwise probabilities. We then fine-tune the entire foundation model with a learning rate of 1e-4 for LaBraM, 1e-4 for PaPaGei, 1e-4 for NormWear, and 1e-5 for HuBERT-ECG, as suggested in the original publications. For contrastive knowledge distillation (KD-Contrast), we append a linear layer to map mean-pooled representations from the last layer of the new modality foundation model to those of the old modality foundation model. During inference, we adopt the linear prober from the old modality foundation model to produce classwise probabilities. The temperature for the InfoNCE loss is selected as 0.04 following (Abbaspourazad et al., 2024b).	[]	[ { "raw": "Abbaspourazad et al., 2024b", "authors": "Abbaspourazad et al.", "year": "2024b" } ]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec23_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	D.2 MAIN RESULTS WITH STANDARD DEVIATIONS	[ "D.2 MAIN RESULTS WITH STANDARD DEVIATIONS" ]	0	Tables 1-3 only report the average performance for five seeds due to space constraints. In Table A1-A5, we present the full results with standard deviation.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec24_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	D.3 ADDITIONAL ABLATION STUDIES D.3.1 BIOX-BRIDGE FRAMEWORK WITH TRADITIONAL BIOSIGNAL MODELS	[ "D.3 ADDITIONAL ABLATION STUDIES D.3.1 BIOX-BRIDGE FRAMEWORK WITH TRADITIONAL BIOSIGNAL MODELS" ]	0	All results presented in the main body focused on building a bridge between biosignal foundation models to better support ongoing research in this area. To show that BioX-Bridge is also compatible with traditional biosignal models, we replace HuBERT-ECG with a pre-trained ECG model, ECG-DualNet (Rohr et al., 2022), in Table A6. We notice that BioX-Bridge continues to outperform the baseline methods while significantly reducing the number of trainable parameters	[]	[ { "raw": "Rohr et al., 2022", "authors": "Rohr et al.", "year": "2022" } ]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_1__sec25_p0	Artificial Intelligence_1	BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS	D.3.2 IMPACT OF FOUNDATION MODEL SIZE	[ "D.3.2 IMPACT OF FOUNDATION MODEL SIZE" ]	0	The HubERT-ECG family of models is available in three sizes, with varying numbers of parameters. We replace the small version of HuBERT-ECG (30M) with base (93M) and large (183M) versions, and report both baseline and our results in Table A7. We observe that the best transfer performance for BioX-Bridge is achieved using the large version of HuBERT-ECG, thanks to the more powerful representational capabilities. Overall, using base and large model yields better transfer performance	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02276	[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]	[]
Artificial Intelligence_10_abstract_p0	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	Abstract	[ "Abstract" ]	0	In tabular anomaly detection (AD), textual semantics often carry critical signals, as the definition of an anomaly is closely tied to domain-specific context. However, existing benchmarks provide only raw data points without semantic context, overlooking rich textual metadata such as feature descriptions and domain knowledge that experts rely on in practice. This limitation restricts research flexibility and prevents models from fully leveraging domain knowledge for detection. ReTabAD addresses this gap by Restoring textual semantics to enable context-aware Tabular Anomaly Detection research. We provide (1) 20 carefully curated tabular datasets enriched with structured textual metadata, together with implementations of stateof-the-art AD algorithms-including classical, deep learning, and LLM-based approaches-and (2) a zero-shot LLM framework that leverages semantic context without task-specific training, establishing a strong baseline for future research. Furthermore, this work provides insights into the role and utility of textual metadata in AD through experiments and analysis. Results show that semantic context improves detection performance and enhances interpretability by supporting domain-aware reasoning. These findings establish ReTabAD as a benchmark for systematic exploration of context-aware AD. The resource is publicly released at https://yoonsanghyu.github.io/ReTabAD/.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec0_p0	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	INTRODUCTION	[ "INTRODUCTION" ]	0	Anomaly detection (AD), the task of identifying patterns that diverge from the distribution of normal data, represents a fundamental challenge across diverse domains such as finance (Al-Hashedi & Magalingam, 2021), cybersecurity (Landauer et al., 2023), manufacturing (Kharitonov et al., 2022), and healthcare (Fernando et al., 2021). In particular, anomaly detection in tabular data, i.e., tabular AD, is of critical importance, as structured datasets-including sensor logs, transaction records, and clinical measurements-form the backbone of real-world operational systems. Despite its practical importance, tabular AD remains a challenging problem due to the heterogeneous nature of tabular datasets, which often combine numerical, categorical, and even textual attributes with rich semantic meaning.	[]	[ { "raw": "Al-Hashedi & Magalingam, 2021", "authors": "Al-Hashedi & Magalingam", "year": "2021" }, { "raw": "Landauer et al., 2023", "authors": "Landauer et al.", "year": "2023" }, { "raw": "Kharitonov et al., 2022", "authors": "Kharitonov et al.", "year": "2022" }, { "raw": "Fernando et al., 2021", "authors": "Fernando et al.", "year": "2021" } ]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec0_p1	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	INTRODUCTION	[ "INTRODUCTION" ]	1	DAMI Repository (Campos et al., 2016) and ADBench (Han et al., 2022) have made significant contributions to tabular AD by establishing standardized benchmarks that enabled fair and rigorous comparisons across AD algorithms. Despite this contribution, these benchmarks were developed under earlier detection paradigms that prioritized only numerical features with conventional machine learning pipelines-for instance, converting categorical values into arbitrary integer codes, or discarding non-numerical fields (as shown in Figure 1). More critically, they excluded textual metadata that practitioners routinely rely on in real-world applications, e.g., feature descriptions, domain context, categorical semantics, measurement units, and operational constraints. This omission is particularly problematic because the definition of an anomaly is inherently "context-dependent" (Chandola et al., 2009;Pang et al., 2021)-without understanding feature semantics and domain constraints, models may misclassify benign deviations as anomalies or fail to detect subtle but critical abnormalities. Recent advances in Large language models (LLMs) demonstrate improved capabilities for handling textual representations of numerical data, particularly in contextual reasoning (Kojima et al., 2022;Fons et al., 2024). These advances suggest new opportunities for leveraging semantic metadata in tabular AD that has been missing from traditional approaches. In particular, such textual metadata can serve as valuable domain priors-enabling models to form more plausible decision boundaries, better interpret anomalous behavior, and generalize more effectively in small-data settings. Early efforts such as AnoLLM (Tsai et al., 2025) have begun to explore this direction, revealing the potential of LLM-based approaches in tabular settings. However, their application remains constrained to utilizing only column names, due to the lack of richer textual annotations in existing benchmarks. This disconnect underscores the need for a new benchmark that restores semantic context, allowing researchers to meaningfully evaluate and advance context-aware tabular AD.	[]	[ { "raw": "Campos et al., 2016", "authors": "Campos et al.", "year": "2016" }, { "raw": "Han et al., 2022", "authors": "Han et al.", "year": "2022" }, { "raw": "Chandola et al., 2009", "authors": "Chandola et al.", "year": "2009" }, { "raw": "Pang et al., 2021", "authors": "Pang et al.", "year": "2021" }, { "raw": "Kojima et al., 2022", "authors": "Kojima et al.", "year": "2022" }, { "raw": "Fons et al., 2024", "authors": "Fons et al.", "year": "2024" }, { "raw": "Tsai et al., 2025", "authors": "Tsai et al.", "year": "2025" } ]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec0_p2	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	INTRODUCTION	[ "INTRODUCTION" ]	2	Motivated by these considerations, we introduce ReTabAD, the first context-aware tabular AD benchmark that incorporates rich textual metadata into AD, enabling models to ground their predictions in semantic context rather than purely numerical patterns. As highlighted in Table 1, our benchmark prioritizes quality by curating 20 tabular datasets with rich textual metadata and carefully validated anomaly definitions. This design stands in contrast to ADBench (Han et al., 2022), which achieves broader coverage by simply transforming CV/NLP datasets into tabular format-an approach that increases dataset quantity but lacks the semantic context necessary for context-aware AD. Alongside these curated datasets, ReTabAD enables comprehensive evaluation across 17 algorithms, ranging from classical and deep learning methods to the latest LLM-based approach. Furthermore, we also propose a zero-shot LLM baseline that leverages the provided metadata without any task-specific training to establish a fair starting point for context-aware AD. Empirically, incorporating textual metadata improves zero-shot LLM performance by an average of +7.6% AUROC across multiple models (see Table 4). As a result, our zero-shot LLM approach achieves performance comparable to state-of-the-art training-based methods, despite never being trained on task-specific data. Together, these contributions position ReTabAD as a unified and forward-looking benchmark that enables systematic study of tabular AD in both traditional and context-aware paradigms.	[]	[ { "raw": "Han et al., 2022", "authors": "Han et al.", "year": "2022" } ]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec0_p3	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	INTRODUCTION	[ "INTRODUCTION" ]	3	To sum up, our contributions are summarized as follows:	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec0_p4	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	INTRODUCTION	[ "INTRODUCTION" ]	4	• The first context-aware tabular AD benchmark: We present ReTabAD, which provides 20 carefully curated tabular datasets enriched with textual metadata and faithful anomaly definitions, and enables standardized evaluation across 16 unsupervised algorithms, including classical, deep learning, and LLM-based approaches.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec0_p5	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	INTRODUCTION	[ "INTRODUCTION" ]	5	• Zero-shot tabular AD framework leveraging textual metadata: We introduce and evaluate a zero-shot LLM framework that establishes not only a strong and competitive baseline but also a new evaluation paradigm for context-aware tabular AD. By leveraging semantic context without task-specific training, ReTabAD provides a standardized reference point for future research.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec0_p6	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	INTRODUCTION	[ "INTRODUCTION" ]	6	• Insights into context-aware tabular AD: We systematically analyze how incorporating textual metadata affects not only model performance but also reasoning quality. Specifically, we evaluate alignment between predicted key features and supervised attributions, and qualitatively examine model-generated reasoning texts, providing insights into how context enhances interpretability and anomaly attribution. Overall, our work emphasizes the importance of semantic information (e.g., textual metadata) in AD for tabular data. We hope that this perspective, together with our new benchmark, will guide future research toward developing context-aware approaches.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec1_p0	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	LIMITATIONS ON PREVIOUS TABULAR AD BENCHMARKS	[ "LIMITATIONS ON PREVIOUS TABULAR AD BENCHMARKS" ]	0	Benchmarking efforts in tabular AD have primarily aimed to provide reproducible and systematic comparisons of algorithms across diverse datasets. The comparative studies (Goldstein & Uchida, 2016;Campos et al., 2016;Ruff et al., 2021b) established the first generation of evaluation resources, offering systematic comparisons of a broad range of algorithms on real-world datasets. In particular, the DAMI Repository (Campos et al., 2016) introduced semantically meaningful tabular datasets with well-defined anomaly classes, demonstrating the importance of defining anomalies in a way that reflects meaningful deviations. With the rise of deep learning, ADBench (Han et al., 2022) further advanced the field by integrating both classical and deep anomaly detection algorithms, collecting datasets from multiple domains, and analyzing them from diverse perspectives. Together with open-source libraries such as PyOD (Zhao et al., 2019) and DeepOD (Xu et al., 2023a), these resources have substantially advanced the field by enabling standardized protocols and large-scale experimental studies. Despite these contributions, existing benchmarks lack textual metadata that makes semantic context explicit. Although anomaly classes are often well defined, the datasets are typically provided only in numerical form, and the conversion processes-such as one-hot encoding, text embedding, or dimensionality reduction-inevitably discard important semantic information. Some benchmarks also incorporate datasets from non-tabular modalities like images and speech, where feature-level semantics cannot be clearly specified. As a result, most prior research has been confined to improving performance within these numeric-only settings, focusing on better modeling of given datasets rather than leveraging the rich textual information that practitioners use in real-world anomaly detection.	[]	[ { "raw": "Goldstein & Uchida, 2016", "authors": "Goldstein & Uchida", "year": "2016" }, { "raw": "Campos et al., 2016", "authors": "Campos et al.", "year": "2016" }, { "raw": "Ruff et al., 2021b", "authors": "Ruff et al.", "year": "2021b" }, { "raw": "Campos et al., 2016", "authors": "Campos et al.", "year": "2016" }, { "raw": "Han et al., 2022", "authors": "Han et al.", "year": "2022" }, { "raw": "Zhao et al., 2019", "authors": "Zhao et al.", "year": "2019" }, { "raw": "Xu et al., 2023a", "authors": "Xu et al.", "year": "2023a" } ]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec2_p0	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	RETABAD: TABULAR AD BENCHMARK RESTORING SEMANTIC CONTEXT	[ "RETABAD: TABULAR AD BENCHMARK RESTORING SEMANTIC CONTEXT" ]	0	To overcome these limitations as described in Section 2, we introduce ReTabAD, a next-generation benchmark that focuses exclusively on genuinely tabular datasets while systematically integrating rich semantic information. This design enables the systematic evaluation of context-aware AD. Label: Binary label (0: Normal, 1: Suspect/Pathologic)	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec3_p0	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	DATASETS	[ "DATASETS" ]	0	Data Collection and Annotation. ReTabAD covers 20 tabular datasets drawn from widely used repositories and real-world application domains. Building upon the comprehensive coverage of ADBench (Han et al., 2022), we curate a subset of datasets and enrich them with textual metadata according to the following criteria: (1) datasets must have a clearly defined domain and ground-truth anomaly labels, (2) dataset size should be manageable for diverse models while reflecting real-world class imbalance, (3) sufficient documentation should exist to recover reliable semantic descriptions, and (4) datasets that are too easy, where performance is already saturated, should be discarded. In addition, we incorporate new datasets by inspecting their class label definitions and class-wise sample ratios, and then annotating anomalies accordingly to construct well-posed AD tasks. A complete list of the datasets and their description is summarized in Table 7 in the Appendix A and an overview of the data collection and annotation process is illustrated in Figure 1.	[]	[ { "raw": "Han et al., 2022", "authors": "Han et al.", "year": "2022" } ]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec3_p1	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	DATASETS	[ "DATASETS" ]	1	Semantically-rich Tabular Data. To incorporate rich semantic information into tabular data, we consider the following two principles:	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec3_p2	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	DATASETS	[ "DATASETS" ]	2	• Raw Numerical Preservation: ReTabAD preserves numerical features in their original scales rather than applying unspecified normalization or standardization. Maintaining raw values retains domain-meaningful interpretations that support intuitive reasoning about anomalies. For instance, a resting heart rate of 200 bpm in an adult is directly recognizable as a critical medical anomaly, whereas its normalized counterpart only conveys statistical rarity without clinical significance.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec3_p3	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	DATASETS	[ "DATASETS" ]	3	• Categorical Text Restoration: Existing benchmarks often provide categorical features in inconsistent formats (e.g., integer-encoded, text-embeddings) or apply inappropriate normalization that discards semantic meaning. Contrary to this practice, ReTabAD restores categorical attributes to their original textual values instead of applying such transformations. This restoration enables models, particularly language-based approaches, to reason over human-interpretable categories rather than arbitrary numerical surrogates.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec3_p4	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	DATASETS	[ "DATASETS" ]	4	Structured Text Metadata. We provide each dataset supplemented with a JSON-formatted metadata file, containing structured semantic descriptions omitted in previous benchmarks. To ensure reliability, we provide direct links to the original data sources, enabling researchers to verify our interpretations.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec3_p5	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	DATASETS	[ "DATASETS" ]	5	The metadata is organized into three categories:	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec3_p6	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	DATASETS	[ "DATASETS" ]	6	• Dataset-Level Descriptions: High-level descriptors, including dataset name, purpose, origin, and collection methodology. Each entry includes links to original publications or repositories to enhance credibility.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec3_p7	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	DATASETS	[ "DATASETS" ]	7	• Column-Level Descriptions: For each column, ReTabAD provides its name, human-readable description, logical type (e.g., numerical, categorical, ordinal, binary), measurement units, if available. All descriptions are cross-referenced with the original dataset documentation.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec3_p8	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	DATASETS	[ "DATASETS" ]	8	• Label-Level Descriptions: For each dataset, ReTabAD clearly specifies which classes are considered normal and which are treated as anomalies. These anomaly definitions are derived directly from the original dataset documentation, with full source attribution.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec4_p0	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	ALGORITHMS	[ "ALGORITHMS" ]	0	To provide a comprehensive comparison, we evaluate 16 representative unsupervised AD models, with full algorithmic details provided in Appendix B. These models are grouped into three categories:	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec4_p1	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	ALGORITHMS	[ "ALGORITHMS" ]	1	(1) Classical methods: We include a diverse range of foundational algorithms that remain widely used. These encompass distance-based (e.g., KNN (Ramaswamy et al., 2000)), density-based (e.g., LOF (Breunig et al., 2000)), boundary-based (e.g., OCSVM (Schölkopf et al., 1999)), and ensemble techniques (e.g., Isolation Forest (Liu et al., 2008)). (2) Deep learning methods: We consider representative deep models designed to capture complex, non-linear representations of normality beyond classical heuristics. These include methods based on deep one-class classification (e.g., DeepSVDD (Ruff et al., 2018)), reconstruction (e.g., RCA (Liu et al., 2021), MCM (Yin et al., 2024)), and contrastive learning (e.g., GOAD (Bergman & Hoshen, 2020), NeuTraL (Qiu et al., 2021)). (3) LLM-based methods: We include AnoLLM (Tsai et al., 2025), which leverages semantic information but is constrained to using only column names as context. To further leverage rich and structured metadata, we propose a new zero-shot LLM framework that directly utilizes context data as input for AD. This investigation illustrates the potential of applying LLMs to AD tasks when comprehensive semantic cues are available. Implementation details are provided in Section 3.3.	[]	[ { "raw": "Ramaswamy et al., 2000", "authors": "Ramaswamy et al.", "year": "2000" }, { "raw": "Breunig et al., 2000", "authors": "Breunig et al.", "year": "2000" }, { "raw": "Schölkopf et al., 1999", "authors": "Schölkopf et al.", "year": "1999" }, { "raw": "Liu et al., 2008", "authors": "Liu et al.", "year": "2008" }, { "raw": "Ruff et al., 2018", "authors": "Ruff et al.", "year": "2018" }, { "raw": "Liu et al., 2021", "authors": "Liu et al.", "year": "2021" }, { "raw": "Yin et al., 2024", "authors": "Yin et al.", "year": "2024" }, { "raw": "Bergman & Hoshen, 2020", "authors": "Bergman & Hoshen", "year": "2020" }, { "raw": "Qiu et al., 2021", "authors": "Qiu et al.", "year": "2021" }, { "raw": "Tsai et al., 2025", "authors": "Tsai et al.", "year": "2025" } ]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec5_p0	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	ZERO-SHOT LLM FRAMEWORK FOR CONTEXT-AWARE AD	[ "ZERO-SHOT LLM FRAMEWORK FOR CONTEXT-AWARE AD" ]	0	Motivation. Despite the importance of textual metadata in real-world tabular datasets, there has been no systematic attempt to leverage such information. Conventional training-based AD models are designed to process data as numerical or categorical vectors, making it structurally difficult to directly integrate textual descriptions that contain the semantic information of features. Although there have been attempts to utilize textual metadata, most have remained limited-for instance, relying only on column name without capturing the broader context. As a consequence, such approaches fail to leverage critical signals for AD conveyed by metadata, including inter-feature relationships and domain-specific contextual cues. To address this limitation, we propose a zero-shot LLM-based AD framework that explores the potential of semantic context through context-aware reasoning over tabular data. By enabling the LLM to interpret feature semantics and relationships in natural language, our framework shifts the focus from purely statistical patterns to high-level contextual understanding and points toward a new paradigm for context-aware AD.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec5_p1	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	ZERO-SHOT LLM FRAMEWORK FOR CONTEXT-AWARE AD	[ "ZERO-SHOT LLM FRAMEWORK FOR CONTEXT-AWARE AD" ]	1	Our Design. To utilize LLMs for tabular AD, we propose a zero-shot LLM framework as a strong baseline for context-aware AD, as illustrated in Figure 2. We employ a carefully designed prompt that integrates domain knowledge and structured reasoning guidelines. The prompt for each data instance x i consists of three key components: a System Prompt (S), a Data Input Formatter (T (x i )), and a Structured Output Query (Q).	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec6_p0	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	System Prompt (S):	[ "System Prompt (S):" ]	0	The system prompt provides the necessary context for the task. It is constructed by concatenating the following information(detailed templates are in the Appendix D.3):	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec6_p1	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	System Prompt (S):	[ "System Prompt (S):" ]	1	• Role and Task Definition (S role ): Assigns the LLM an expert persona for each dataset's domain (e.g., "You are a domain expert in finance"). { "anomaly_score": s, "key_features": F, "reasoning": e } where s is the anomaly score in [0, 1], F is a list of key features, and e is the explanation in text.	[ 0, 1 ]	[]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[ { "id": 1, "text": "Financial fraud detection applying data mining techniques: A comprehensive review from 2009 to 2019\n\t\t\n\t\t\tAl-HashediKhaled Gubran\n\t\t\n\t\t\n\t\t\tPritheegaMagalingam\n\t\t\n\t\n\t\n\t\tComputer Science Review\n\t\t\n\t\t\t40\n\t\t\t100402\n\t\t\t2021" } ]
Artificial Intelligence_10__sec7_p0	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	EVALUATION	[ "EVALUATION" ]	0	Problem Setup. The concept of normality, a ground-truth law of normal behavior for a given task (Ruff et al., 2021a), is modeled through a probability distribution P with corresponding density function p(x). Conventional tabular anomaly detectors map each data instance x ∈ R K to a class label y ∈ {0, 1} by learning this distribution, where K denotes the feature dimension, and 0 and 1 indicate normal and anomalous status, respectively. ReTabAD extends this traditional setup by incorporating semantic metadata M, which encompasses domain knowledge, feature descriptions, and contextual information about normal behavior patterns. The primary objective is to model a conditional distribution p(x\|M) that leverages semantic context for enhanced AD. This work focuses on analyzing the impact of semantic context M on detection performance.	[]	[ { "raw": "Ruff et al., 2021a", "authors": "Ruff et al.", "year": "2021a" } ]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec7_p1	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	EVALUATION	[ "EVALUATION" ]	1	Hyperparameter Optimization. We follow the default model architectures and tune key hyperparameters, including preprocessing settings. Hyperparameters for training-based methods are faithfully searched, and detailed configurations are provided in the Appendix C for reproducibility.	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec7_p2	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	EVALUATION	[ "EVALUATION" ]	2	Evaluator LLMs. We evaluate a set of leading LLMs in a zero-shot AD setting to assess their general semantic reasoning capabilities. Specifically, we include state-of-the-art reasoning-oriented models-GPT-4.1 (Fachada et al., 2025), Claude-3.7-sonnet (Anthropic, 2025), Qwen3-235B (Yang et al., 2025), and Gemini-2.5-pro (Comanici et al., 2025)-as well as a lightweight yet competitive model, GPT-4o-mini (OpenAI, 2024). For consistency and clarity, we report detailed ablation and qualitative analyses with Gemini-2.5-pro.	[]	[ { "raw": "Fachada et al., 2025", "authors": "Fachada et al.", "year": "2025" }, { "raw": "Anthropic, 2025", "authors": "Anthropic", "year": "2025" }, { "raw": "Yang et al., 2025", "authors": "Yang et al.", "year": "2025" }, { "raw": "Comanici et al., 2025", "authors": "Comanici et al.", "year": "2025" }, { "raw": "OpenAI, 2024", "authors": "OpenAI", "year": "2024" } ]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[ { "id": 2, "text": "Anthropic\n\t\t\n\t\t\n\t\t\n\t\t\tAugust 2025" }, { "id": 23, "text": "GPT-4o, DALL-E, and Whisper\n\t\t\n\t\t\tShimonIfrah\n\t\t\n\t\t10.1007/979-8-8688-0599-8_6\n\t\t\n\t\n\t\n\t\tGetting Started with Azure OpenAI\n\t\t\n\t\t\tApress\n\t\t\t2024" } ]
Artificial Intelligence_10__sec7_p3	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	EVALUATION	[ "EVALUATION" ]	3	Evaluation Protocol. ReTabAD follows the one-class classification setting, where the training set consists of 50% of the normal samples, while the test set consists of the remaining normal samples and all anomalous samples, as in previous studies (Bergman & Hoshen, 2020;Livernoche et al., 2023). For performance evaluation, we primarily report the Area Under the Receiver Operating Characteristic Curve (AUROC), a standard metric in AD (Ruff et al., 2018). To mitigate randomness, each experiment is run five times with different seeds, and Full results with the Area Under the Precision-Recall Curve (AUPRC) (Saito & Rehmsmeier, 2015) and standard deviations are in the Appendix E, F.	[]	[ { "raw": "Bergman & Hoshen, 2020", "authors": "Bergman & Hoshen", "year": "2020" }, { "raw": "Livernoche et al., 2023", "authors": "Livernoche et al.", "year": "2023" }, { "raw": "Ruff et al., 2018", "authors": "Ruff et al.", "year": "2018" }, { "raw": "Saito & Rehmsmeier, 2015", "authors": "Saito & Rehmsmeier", "year": "2015" } ]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec8_p0	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	EXPERIMENTS ON RETABAD BENCHMARK	[ "EXPERIMENTS ON RETABAD BENCHMARK" ]	0	This section explores the potential of context-aware approaches in tabular AD. Specifically, we design our experiments to investigate the following research questions:	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]
Artificial Intelligence_10__sec8_p1	Artificial Intelligence_10	RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION	EXPERIMENTS ON RETABAD BENCHMARK	[ "EXPERIMENTS ON RETABAD BENCHMARK" ]	1	• Does semantic context M improve AD performance? ( §4.1)	[]	[]	2,025	null	https://arxiv.org//pdf/2510.02060	[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]	[]

Artificial Intelligence_1_abstract_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

Abstract

[ "Abstract" ]

0

Biosignals offer valuable insights into the physiological states of the human body. Although biosignal modalities differ in functionality, signal fidelity, sensor comfort, and cost, they are often intercorrelated, reflecting the holistic and interconnected nature of human physiology. This opens up the possibility of performing the same tasks using alternative biosignal modalities, thereby improving the accessibility, usability, and adaptability of health monitoring systems. However, the limited availability of large labeled datasets presents challenges for training models tailored to specific tasks and modalities of interest. Unsupervised cross-modal knowledge transfer offers a promising solution by leveraging knowledge from an existing modality to support model training for a new modality. Existing methods are typically based on knowledge distillation, which requires running a teacher model alongside student model training, resulting in high computational and memory overhead. This challenge is further exacerbated by the recent development of foundation models that demonstrate superior performance and generalization across tasks at the cost of large model sizes. To this end, we explore a new framework for unsupervised cross-modal knowledge transfer of biosignals by training a lightweight bridge network to align the intermediate representations and enable information flow between foundation models and across modalities. Specifically, we introduce an efficient strategy for selecting alignment positions where the bridge should be constructed, along with a flexible prototype network as the bridge architecture. Extensive experiments across multiple biosignal modalities, tasks, and datasets show that BioX-Bridge reduces the number of trainable parameters by 88-99% while maintaining or even improving transfer performance compared to state-of-the-art methods.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec0_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

INTRODUCTION

[ "INTRODUCTION" ]

0

Biosignals, such as electrocardiogram (ECG), electroencephalogram (EEG), and photoplethysmography (PPG), provide critical insights into the underlying physiological states of individuals. They are essential tools in modern healthcare and have often been considered the gold standard for diagnostics (Rosenberg & Van Hout, 2013;Stracina et al., 2022). In the past decade, the advancement of artificial intelligence (AI) has enabled remarkable capabilities in automated diagnostics and monitoring, such as stress assessment (Mentis et al., 2024), sleep stage classification (Mostafa et al., 2019), and arrhythmia detection (Parvaneh et al., 2019). However, many biosignal sensors are not suitable for use outside clinical settings due to factors such as user discomfort, high manufacturing costs, and excessive power consumption.

[]

[ { "raw": "Rosenberg & Van Hout, 2013", "authors": "Rosenberg & Van Hout", "year": "2013" }, { "raw": "Stracina et al., 2022", "authors": "Stracina et al.", "year": "2022" }, { "raw": "Mentis et al., 2024", "authors": "Mentis et al.", "year": "2024" }, { "raw": "Mostafa et al., 2019", "authors": "Mostafa et al.", "year": "2019" }, { "raw": "Parvaneh et al., 2019", "authors": "Parvaneh et al.", "year": "2019" } ]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec0_p1

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

INTRODUCTION

[ "INTRODUCTION" ]

1

A promising direction is to harness the correlations between different biosignal modalities and perform the same tasks using alternative modalities, making health monitoring systems more accessible, practical, and flexible (Wang et al., 2023;Yang et al., 2023). For example, being able to perform the same tasks using single-lead PPG data from a wearable smartwatch, instead of relying on 12-lead ECGs, would greatly reduce hardware complexity and cost, while enabling continuous, user-friendly monitoring in everyday environments. Unfortunately, training such models requires large-scale labeled datasets, which are often difficult to obtain in biosignal applications due to the high cost and domain-specific expertise required for data collection and annotation. This highlights the need for effective knowledge transfer between biosignal modalities, leveraging models trained in old or well-established modalities to support the development of models for new or underrepresented biosignal modalities.

[]

[ { "raw": "Wang et al., 2023", "authors": "Wang et al.", "year": "2023" }, { "raw": "Yang et al., 2023", "authors": "Yang et al.", "year": "2023" } ]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec0_p2

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

INTRODUCTION

[ "INTRODUCTION" ]

2

Unsupervised cross-modal knowledge transfer stands out as a practical solution to address the aforementioned needs. Existing methods can be divided into two categories: data translation and knowledge distillation. As illustrated in Figure 1(a), data translation directly translates data from the new modality to the old modality, enabling the direct reuse of existing models from the old modality (Sarkar & Etemad, 2021). However, the exploration of data translation has been limited to a certain pair of modalities, such as PPG and ECG. Figure 1(b) illustrates knowledge distillation, which seeks to train a student model for the new modality to mimic the output of a pre-trained teacher model from the old modality (Abbaspourazad et al., 2024b;Zhang et al., 2024). The distillation process is memory intensive, as it requires forward inference with both the student and teacher models, in addition to backpropagation with the student model. The computational burden is further exacerbated by the emergence of large-scale biosignal foundation models (Coppola et al., 2024;Jiang et al., 2024;Pillai et al., 2025), which are mostly trained on specific modalities and have demonstrated exceptional performance across a wide range of tasks. Although these models offer tremendous performance gains, their use in cross-modal knowledge transfer is hindered by their size, which makes traditional knowledge distillation solutions computationally prohibitive for users without access to high-end GPUs. For example, distilling knowledge from PaPaGei, a PPG-based foundation model (Pillai et al., 2025), to the ECG-FM student model (McKeen et al., 2024) on the WESAD dataset (Schmidt et al., 2018), with a batch size of eight, requires more than 32GB of VRAM. Furthermore, because of data sharing regulations and privacy concerns, such a distillation process often needs to be performed locally where the data resides, under low-resource conditions. These constraints call for the development of an effective and efficient cross-modal transfer framework that can fully leverage the representation capability and embedded knowledge of the foundation models.

[]

[ { "raw": "Sarkar & Etemad, 2021", "authors": "Sarkar & Etemad", "year": "2021" }, { "raw": "Abbaspourazad et al., 2024b", "authors": "Abbaspourazad et al.", "year": "2024b" }, { "raw": "Zhang et al., 2024", "authors": "Zhang et al.", "year": "2024" }, { "raw": "Coppola et al., 2024", "authors": "Coppola et al.", "year": "2024" }, { "raw": "Jiang et al., 2024", "authors": "Jiang et al.", "year": "2024" }, { "raw": "Pillai et al., 2025", "authors": "Pillai et al.", "year": "2025" }, { "raw": "Pillai et al., 2025", "authors": "Pillai et al.", "year": "2025" }, { "raw": "McKeen et al., 2024", "authors": "McKeen et al.", "year": "2024" }, { "raw": "Schmidt et al., 2018", "authors": "Schmidt et al.", "year": "2018" } ]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec0_p3

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

INTRODUCTION

[ "INTRODUCTION" ]

3

To this end, we propose BioX-Bridge, a new framework for unsupervised cross-modal knowledge transfer via model bridging, as illustrated in Figure 1(c). The core idea is to construct a bridge that projects intermediate representations from one biosignal model to another, leveraging the powerful representational capability and the rich embedded knowledge of foundation models. 1 The framework comprises two key components: bridge position selection and bridge architecture design. Specifically, we introduce an efficient two-stage strategy for selecting optimal input and output positions by evaluating the quality and similarity of intermediate representations between two biosignal models.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec0_p4

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

INTRODUCTION

[ "INTRODUCTION" ]

4

To enable effective projection between high-dimensional spaces, we design a prototype network composed of a learnable prototype set and a low-rank approximation module to compute aggregation weights. Notably, only the bridge network requires training to enable interoperability between models of different modalities. We evaluate the effectiveness of BioX-Bridge in three biosignal datasets involving different modalities, demonstrating superior efficiency compared to existing methods.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec0_p5

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

INTRODUCTION

[ "INTRODUCTION" ]

5

Extensive ablation studies further confirm the robustness of the proposed framework under various conditions. Our contributions can be summarized as follows:

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec0_p6

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

INTRODUCTION

[ "INTRODUCTION" ]

6

• We propose BioX-Bridge, a novel unsupervised model bridging framework that enables crossmodal knowledge transfer through information flow between biosignal models.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec0_p7

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

INTRODUCTION

[ "INTRODUCTION" ]

7

• We introduce key components to support the framework, including an efficient two-stage strategy for selecting bridge positions and a prototype network with low-rank approximation for effective high-dimensional projection.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec0_p8

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

INTRODUCTION

[ "INTRODUCTION" ]

8

• We demonstrate the efficiency of BioX-Bridge through experiments on three biosignal datasets, four modalities, and six transfer directions, demonstrating robustness through comprehensive ablation studies.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec1_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

RELATED WORKS

[ "RELATED WORKS" ]

0

Unsupervised Cross-modal Knowledge Transfer Existing methods can be divided into two categories: knowledge distillation and data translation.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec1_p1

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

RELATED WORKS

[ "RELATED WORKS" ]

1

Knowledge distillation was introduced as a model compression technique, where a smaller student model learns to mimic a larger and high-performing teacher model by matching its output distributions (Hinton et al., 2015). The concept has since been extended to cross-modal knowledge transfer. Early efforts focused on computer vision applications across a variety of sensor modalities, such as vision to depth images (Garcia et al., 2018;Gupta et al., 2016;Hoffman et al., 2016;Tian et al., 2020), to radio frequency heatmaps (Zhao et al., 2018), and to sound (Aytar et al., 2016;Xue et al., 2021). The core idea is to leverage unlabeled but semantically aligned data pairs to bridge the modality gap and transfer relevant knowledge to the corresponding tasks (Gou et al., 2021;Moslemi et al., 2024). Recent efforts have also investigated cross-modal knowledge distillation for biosignals. For example, Brant-X (Zhang et al., 2024) introduced a unified biosignal alignment framework that transfers knowledge from EEG to other biosignal modalities through a two-level semantic alignment strategy, such that the student model can provide complementary representations to the teacher model and improve downstream task performance. In another work (Abbaspourazad et al., 2024b), the distillation of knowledge from PPG to accelerometer signals was used to accurately predict physiological states such as heart rate. However, the aforementioned methods require training a full-size student model from scratch, which becomes increasingly impractical as model sizes grow, especially for resource-constrained settings.

[]

[ { "raw": "Hinton et al., 2015", "authors": "Hinton et al.", "year": "2015" }, { "raw": "Garcia et al., 2018", "authors": "Garcia et al.", "year": "2018" }, { "raw": "Gupta et al., 2016", "authors": "Gupta et al.", "year": "2016" }, { "raw": "Hoffman et al., 2016", "authors": "Hoffman et al.", "year": "2016" }, { "raw": "Tian et al., 2020", "authors": "Tian et al.", "year": "2020" }, { "raw": "Zhao et al., 2018", "authors": "Zhao et al.", "year": "2018" }, { "raw": "Aytar et al., 2016", "authors": "Aytar et al.", "year": "2016" }, { "raw": "Xue et al., 2021", "authors": "Xue et al.", "year": "2021" }, { "raw": "Gou et al., 2021", "authors": "Gou et al.", "year": "2021" }, { "raw": "Moslemi et al., 2024", "authors": "Moslemi et al.", "year": "2024" }, { "raw": "Zhang et al., 2024", "authors": "Zhang et al.", "year": "2024" }, { "raw": "Abbaspourazad et al., 2024b", "authors": "Abbaspourazad et al.", "year": "2024b" } ]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec1_p2

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

RELATED WORKS

[ "RELATED WORKS" ]

2

Data translation aims to achieve unsupervised cross-modal knowledge transfer by directly translating raw data from one modality to another. Generative adversarial networks (GAN) (Goodfellow et al., 2020) and their variants (Mirza & Osindero, 2014;Zhu et al., 2017) have been widely adopted for modality translation tasks in the visual and signal processing domains (Duan et al., 2021;Sikka et al., 2021;Yang et al., 2020). A recent work (Wang et al., 2023) leveraged knowledge graphs to learn transformations between independently trained foundation models for proteins, drugs, and text. Nevertheless, reliance on structured knowledge graphs limits their applicability to biosignal scenarios, where such structured relationships are scarce or nonexistent. In the biosignal domain, cross-modal translation efforts have been largely limited to translation from PPG to ECG (Sarkar & Etemad, 2021;Zhu et al., 2021). Extending such translation to other modalities, such as EEG to ECG, remains largely underexplored.

[]

[ { "raw": "Goodfellow et al., 2020", "authors": "Goodfellow et al.", "year": "2020" }, { "raw": "Mirza & Osindero, 2014", "authors": "Mirza & Osindero", "year": "2014" }, { "raw": "Zhu et al., 2017", "authors": "Zhu et al.", "year": "2017" }, { "raw": "Duan et al., 2021", "authors": "Duan et al.", "year": "2021" }, { "raw": "Sikka et al., 2021", "authors": "Sikka et al.", "year": "2021" }, { "raw": "Yang et al., 2020", "authors": "Yang et al.", "year": "2020" }, { "raw": "Wang et al., 2023", "authors": "Wang et al.", "year": "2023" }, { "raw": "Sarkar & Etemad, 2021", "authors": "Sarkar & Etemad", "year": "2021" }, { "raw": "Zhu et al., 2021", "authors": "Zhu et al.", "year": "2021" } ]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec1_p3

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

RELATED WORKS

[ "RELATED WORKS" ]

3

Biosignals Foundation Models Inspired by the recent success of large-scale pre-training in natural language processing (Achiam et al., 2023) and computer vision (Dosovitskiy et al., 2020), the development of foundation models for biosignals has garnered much interest (Han et al., 2024;Lai et al., 2025). Through large-scale self-supervised training on public and private biosignal datasets, several biosignal foundation models have been developed to capture rich and transferable representations, enabling more robust and efficient downstream adaptation. These models span a variety of modalities, including EEG (Chen et al., 2025a;2024;Cui et al., 2023;Jiang et al., 2024;Wang et al., 2024), ECG (Coppola et al., 2024;Li et al., 2024;McKeen et al., 2024), PPG (Chen et al., 2025b;Pillai et al., 2025;Saha et al., 2025), accelerometer (Abbaspourazad et al., 2024b), and general-purpose biosignal models (Yang et al., 2023). In addition to unimodal models, recent work has explored multimodal foundation models for various applications such as health monitoring (Abbaspourazad et al., 2024a;Luo et al., 2024), sleep (Thapa et al., 2024), and activity recognition (Narayanswamy et al., 2024). Despite their strong performance on data from modalities seen during At the inference stage, the bridge has been constructed and enables the flow of information between the two models in order to make predictions on data from the new modality. (c) The bridge consists of a low-rank approximation module and a prototype set. The low-rank approximation module generates aggregation weights for the prototype vectors.

[]

[ { "raw": "Achiam et al., 2023", "authors": "Achiam et al.", "year": "2023" }, { "raw": "Dosovitskiy et al., 2020", "authors": "Dosovitskiy et al.", "year": "2020" }, { "raw": "Han et al., 2024", "authors": "Han et al.", "year": "2024" }, { "raw": "Lai et al., 2025", "authors": "Lai et al.", "year": "2025" }, { "raw": "Chen et al., 2025a", "authors": "Chen et al.", "year": "2025a" }, { "raw": "Cui et al., 2023", "authors": "Cui et al.", "year": "2023" }, { "raw": "Jiang et al., 2024", "authors": "Jiang et al.", "year": "2024" }, { "raw": "Wang et al., 2024", "authors": "Wang et al.", "year": "2024" }, { "raw": "Coppola et al., 2024", "authors": "Coppola et al.", "year": "2024" }, { "raw": "Li et al., 2024", "authors": "Li et al.", "year": "2024" }, { "raw": "McKeen et al., 2024", "authors": "McKeen et al.", "year": "2024" }, { "raw": "Chen et al., 2025b", "authors": "Chen et al.", "year": "2025b" }, { "raw": "Pillai et al., 2025", "authors": "Pillai et al.", "year": "2025" }, { "raw": "Saha et al., 2025", "authors": "Saha et al.", "year": "2025" }, { "raw": "Abbaspourazad et al., 2024b", "authors": "Abbaspourazad et al.", "year": "2024b" }, { "raw": "Yang et al., 2023", "authors": "Yang et al.", "year": "2023" }, { "raw": "Abbaspourazad et al., 2024a", "authors": "Abbaspourazad et al.", "year": "2024a" }, { "raw": "Luo et al., 2024", "authors": "Luo et al.", "year": "2024" }, { "raw": "Thapa et al., 2024", "authors": "Thapa et al.", "year": "2024" }, { "raw": "Narayanswamy et al., 2024", "authors": "Narayanswamy et al.", "year": "2024" } ]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec1_p4

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

RELATED WORKS

[ "RELATED WORKS" ]

4

pre-training, these models struggle with generalization to unseen modalities due to mismatches in input dimensions and data distributions (Liu et al., 2024).

[]

[ { "raw": "Liu et al., 2024", "authors": "Liu et al.", "year": "2024" } ]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec1_p5

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

RELATED WORKS

[ "RELATED WORKS" ]

5

We provide further discussion on model stitching and domain adaptation in Appendix C.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec2_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

METHODS

[ "METHODS" ]

0

The core concept of our proposed BioX-Bridge framework is to build a bridging network that facilitates efficient and effective projection between intermediate representations of biosignal models. This allows the framework to harness the strong representational power of one model while integrating the task-specific knowledge contained in another. We define the problem in Section 3.1 and introduce the idea of model bridging in Section 3.2. We detail the position of the bridge, its architecture and training in Sections 3.3-3.5. An overview of BioX-Bridge is presented in Figure 2.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec3_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

PROBLEM DEFINITION

[ "PROBLEM DEFINITION" ]

0

Assume that we are given an annotated dataset from an old biosignal modality, old) | labeled samples for a specific task, and a corresponding model,

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec3_p1

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

PROBLEM DEFINITION

[ "PROBLEM DEFINITION" ]

1

, where f (old) θ is a pre-trained encoder parametrized by θ followed by a task head g (old) ω parametrized by ω. We also have an un-annotated dataset from a new modality,

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec3_p2

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

PROBLEM DEFINITION

[ "PROBLEM DEFINITION" ]

2

, which shares the same underlying label set with D (old) . We further have a disjoint, un-annotated paired dataset D (pair) = {(x

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec3_p3

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

PROBLEM DEFINITION

[ "PROBLEM DEFINITION" ]

3

. The unsupervised cross-modal knowledge transfer problem aims to obtain a model f , such that f can make predictions on D (new) .

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec4_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

MODEL BRIDGING

[ "MODEL BRIDGING" ]

0

Let f (old) θ be the model for the old modality, parametrized by θ of L layers, and let f (new) ϕ be the model for the new modality, parametrized by ϕ of M layers. The intermediate representations from the m-th layer of the new modality model can then be extracted as:

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec4_p1

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

MODEL BRIDGING

[ "MODEL BRIDGING" ]

1

where x (new) denotes a biosignal time series sample from the new modality. f (new) ϕ ≤m denotes the subset of the new modality model consisting of its first m layers, subject to the constraint Next, we introduce a bridge network to enable the information flow between the new and old modality models by projecting representations from the new modality into the representation space of the old modality:

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec4_p2

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

MODEL BRIDGING

[ "MODEL BRIDGING" ]

2

where b ψ denotes the bridge network parametrized by ψ. h(old) l denotes the projected representation from the new modality to the old modality. Note that the projected representation is designed to mimic the intermediate representation from the l-th layer of the old modality model, defined as h (old) old) , where x (old) is the paired input signal from the old modality. Thus,

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec4_p3

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

MODEL BRIDGING

[ "MODEL BRIDGING" ]

3

Finally, we can obtain predictions using the old modality model starting from the (l + 1)-th layer:

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec4_p4

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

MODEL BRIDGING

[ "MODEL BRIDGING" ]

4

where m and l are also known as the bridge input and output positions, • denotes function composition.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec5_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

BRIDGE POSITION SELECTION

[ "BRIDGE POSITION SELECTION" ]

0

There are L × M possible locations where the bridge can be constructed between the layers of the two models. Although a brute-force search would yield the optimal bridge position, it is computationally expensive. In particular, the choice of the bridge position is one of the most influential factors affecting transfer performance, as we will show in ablation studies. To this end, we propose a two-stage strategy for efficient bridge position selection, as illustrated in Figure 3.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec5_p1

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

BRIDGE POSITION SELECTION

[ "BRIDGE POSITION SELECTION" ]

1

Stage 1: Bridge Input Position (m) Selection The bridge serves to project new modality representations to the old modality representation space, enabling the bridged model to mimic the behavior of the old modality model. As the saying "garbage in, garbage out" suggests, it is important to select discriminative new modality representations that can effectively distinguish among the predictions produced by the old modality model, also known as pseudo labels. We propose to select the input position of the bridge by linear probing, which has been widely used to evaluate the quality of intermediate representations (Alain & Bengio, 2016). The bridge input position selection can be formulated as:

[]

[ { "raw": "Alain & Bengio, 2016", "authors": "Alain & Bengio", "year": "2016" } ]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec5_p2

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

BRIDGE POSITION SELECTION

[ "BRIDGE POSITION SELECTION" ]

2

where ) denotes the pseudo label. L probe denotes the empirical loss for the linear prober.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec5_p3

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

BRIDGE POSITION SELECTION

[ "BRIDGE POSITION SELECTION" ]

3

, we can ease the transformation process by selecting h (old) l to be as similar as possible to h (new) m . We select linear CKA (Kornblith et al., 2019)

[]

[ { "raw": "Kornblith et al., 2019", "authors": "Kornblith et al.", "year": "2019" } ]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec5_p4

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

BRIDGE POSITION SELECTION

[ "BRIDGE POSITION SELECTION" ]

4

For detailed formulation of CKA linear , please refer to the appendix.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec5_p5

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

BRIDGE POSITION SELECTION

[ "BRIDGE POSITION SELECTION" ]

5

Algorithm 1: BioX-Bridge learning procedure

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec5_p6

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

BRIDGE POSITION SELECTION

[ "BRIDGE POSITION SELECTION" ]

6

Input: Old modality model f (old) θ ; New modality model f (new) ϕ ; Task head g (old) ω ; Paired dataset

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec5_p7

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

BRIDGE POSITION SELECTION

[ "BRIDGE POSITION SELECTION" ]

7

Select bridge input position, m, using Eq. ( 4) Select bridge output position, l, using Eq. ( 5)

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec5_p8

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

BRIDGE POSITION SELECTION

[ "BRIDGE POSITION SELECTION" ]

8

using Eq. ( 7) whose intermediate representation exhibits the strongest linear association with the pseudolabels. For bridge output position, we select the layer from f (old) θ whose representation is most similar to that of the bridge input layer.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec6_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

BRIDGE ARCHITECTURE

[ "BRIDGE ARCHITECTURE" ]

0

Models of different modalities operate in distinct representational spaces. The bridge network should be sufficiently parametrized to enable the projection and alignment of the two spaces. A naive bridge architecture is a full-rank linear layer, but this is prohibitively expensive because of the highdimensional projection from the new modality to the old modality. For example, using LaBraM (Jiang et al., 2024) as f (new) ϕ and HuBERT-ECG (Coppola et al., 2024) as f (old) θ , the projection would require

[]

[ { "raw": "Jiang et al., 2024", "authors": "Jiang et al.", "year": "2024" }, { "raw": "Coppola et al., 2024", "authors": "Coppola et al.", "year": "2024" } ]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec6_p1

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

BRIDGE ARCHITECTURE

[ "BRIDGE ARCHITECTURE" ]

1

7 billion parameters. To address the challenge of high-dimensional projection, we propose a prototype network. The prototype network consists of two modules, a prototype set, and a low-rank approximation module. Specifically, the prototype set, P ∈ R Np×d (old) l , consisting of N p prototype vectors with embedding dimension d (new) m , introduces the flexibility to incorporate prior knowledge from f Np , reduces the number of trainable parameters through a low-rank factorization, while generating aggregation weights for prototype vectors, as illustrated in Figure 2.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec6_p2

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

BRIDGE ARCHITECTURE

[ "BRIDGE ARCHITECTURE" ]

2

where Pool(•) denotes a pooling operation along the N (new) m dimension, and Reshape N (old) l ×Np (•) denotes the reshape operation to the specified output dimensions.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec7_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

BRIDGE TRAINING

[ "BRIDGE TRAINING" ]

0

As the difference between h (old) l and h(old) l approaches zero, the bridged model yields predictions identical to those of the old modality model. Formally:

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec7_p1

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

BRIDGE TRAINING

[ "BRIDGE TRAINING" ]

1

Naturally, the training objective for the bridge network is to align the intermediate representations in the L-th layer2 :

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec7_p2

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

BRIDGE TRAINING

[ "BRIDGE TRAINING" ]

2

where L align denotes the loss function, such as cosine loss and mean absolute error loss. The learning process of BioX-Bridge is presented in Algorithm 1. Given the unbalanced nature of the datasets, we report Balanced Accuracy, F1-Weighted, and F1-Macro, following (Jiang et al., 2024;Pillai et al., 2025).

[]

[ { "raw": "Jiang et al., 2024", "authors": "Jiang et al.", "year": "2024" }, { "raw": "Pillai et al., 2025", "authors": "Pillai et al.", "year": "2025" } ]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec8_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

EXPERIMENTS

[ "EXPERIMENTS" ]

0

Backbone Foundation Models For EEG, we adopt the base version of the LaBraM architecture with 5.8M parameters (Jiang et al., 2024). For ECG, we adopt the small version of the HuBERT-ECG architecture with 30.4M parameters (Coppola et al., 2024). For PPG, we adopt the small version of the PaPaGei architecture with 5.7M parameters (Pillai et al., 2025). For EMG, we adopt NormWear with 136.1M parameters (Luo et al., 2024). Note that LaBraM, HuBERT-ECG, and NormWear adopt a CNN-transformer architecture, while PaPaGei adopts a CNN architecture. All models are initialized with the pre-trained weights provided by the original publications. Note that biosignal foundation models are still early in their development, in comparison to foundation models for language and vision. While current models contain a relatively small number of parameters, our method for efficient cross-modal knowledge transfer would be even more valuable as they scale up.

[]

[ { "raw": "Jiang et al., 2024", "authors": "Jiang et al.", "year": "2024" }, { "raw": "Coppola et al., 2024", "authors": "Coppola et al.", "year": "2024" }, { "raw": "Pillai et al., 2025", "authors": "Pillai et al.", "year": "2025" }, { "raw": "Luo et al., 2024", "authors": "Luo et al.", "year": "2024" } ]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec8_p1

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

EXPERIMENTS

[ "EXPERIMENTS" ]

1

Baselines We compare our method with the following baselines evaluated on D (new) : (i) Random denotes a model that produces predictions at random. (ii) CardioGAN uses GAN to synthesize ECG from PPG (Sarkar & Etemad, 2021), and we translate the new modality data (PPG) to the old modality (ECG) for evaluation. (iii) KD (Hinton et al., 2015) is the baseline of knowledge distillation. (iv) KD-contrast (Abbaspourazad et al., 2024b) is a variant of knowledge distillation with contrast loss (Zhang et al., 2024). (v) Oracle denotes the absolute best performance that can be achieved, which is simply the performance of the old modality model using old modality data. Please refer to the appendix for implementation details and further discussions.

[]

[ { "raw": "Sarkar & Etemad, 2021", "authors": "Sarkar & Etemad", "year": "2021" }, { "raw": "Hinton et al., 2015", "authors": "Hinton et al.", "year": "2015" }, { "raw": "Abbaspourazad et al., 2024b", "authors": "Abbaspourazad et al.", "year": "2024b" }, { "raw": "Zhang et al., 2024", "authors": "Zhang et al.", "year": "2024" } ]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec9_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER PERFORMANCE

[ "UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER PERFORMANCE" ]

0

Experiment results on the ISRUC, FOG, and WESAD dataset are presented in Table 1. We observe that BioX-Bridge significantly reduces the number of trainable parameters by 87.9-99.1% and continues to achieve performance comparable to or better than that of the baseline methods. For example, for WESAD (PPG → ECG), BioX-Bridge requires merely 1.3% of trainable parameters while outperforming the baseline methods by around 1-2% across all metrics.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec9_p1

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER PERFORMANCE

[ "UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER PERFORMANCE" ]

1

We also observe that the knowledge transfer performance gap compared to Oracle varies across datasets and knowledge transfer directions. For example, on the ISRUC dataset, we observe approximately 20% balanced accuracy gap between BioX-Bridge (60.11%) and Oracle ( On another note, BioX-Bridge and KD-Contrast achieved higher balanced accuracy than Oracle on several occasions, respectively. This is possible because balanced accuracy is simply an average of recall across classes. Higher balanced accuracy scores and lower F1 scores reflect that knowledge transfer methods achieved better recall but worse precision than Oracle.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec10_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

ABLATION STUDIES

[ "ABLATION STUDIES" ]

0

We conduct ablation studies on the WESAD dataset and the direction of knowledge transfer (PPG → ECG). Additional results are presented in the appendix.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec11_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

Bridge Rank and Prototype Set

[ "Bridge Rank and Prototype Set" ]

0

We study the impact of different hyperparameters for the prototype network in Figures 4a and4b. A performance drop is observed when the approximation rank and prototype set size are too small or too large, likely due to under-/over-parameterization of the bridge network. In particular, the performance peaks at around 0.75M parameters in both cases.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec11_p1

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

Bridge Rank and Prototype Set

[ "Bridge Rank and Prototype Set" ]

1

Dataset Size We reduce the size of the paired dataset for bridge training. We observe in Figure 4c that the transfer performance slowly decays by around 2% at 20% dataset size, showcasing the robustness of the bridge under the low-data regime. Bridge Position Selection To show that the bridge position selection strategy proposed in Section 3.3 is effective, we compare the unsupervised performance of cross-modal knowledge transfer at various positions in Table 2. For "Fixed", we train nine bridges in predefined positions, which is a combination of the first, middle, and last layers of the old and new modality models. The results are the average over all nine positions in five seeds (see more results in the appendix).

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec12_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

Foundation Model

[ "Foundation Model" ]

0

We further analyze the impact of using different foundation models for crossmodal knowledge transfer. In Table 3, we replace the HuBERT-ECG foundation model (Coppola et al., 2024) with ECG-FM (McKeen et al., 2024). Notably, due to the large number of trainable parameters (90M), the knowledge distillation methods with ECG-FM could only be performed with a batch size of 4 on a V100 GPU. As a result, training for more than 50 epochs requires 6.5 hours for knowledge distillation methods and 1.9 hours for BioX-Bridge. Moreover, the performance gap between knowledge distillation methods and BioX-Bridge is much pronounced at 10-17%.

[]

[ { "raw": "Coppola et al., 2024", "authors": "Coppola et al.", "year": "2024" }, { "raw": "McKeen et al., 2024", "authors": "McKeen et al.", "year": "2024" } ]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec13_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

CONCLUSION

[ "CONCLUSION" ]

0

We present BioX-Bridge as an efficient framework for unsupervised cross-modal knowledge transfer across biosignals. To address the challenges of high-dimensional projection between biosignal foundation models, we design a prototype-based architecture for parameter-efficient learning of transformations between representation spaces. Our proposed two-stage bridge position selection strategy further identifies connection points that enable more effective alignment of intermediate representations. Through extensive experiments on diverse biosignal datasets and tasks, we demonstrated that BioX-Bridge achieves performance comparable to or superior to that of state-of-the-art methods while drastically reducing the number of trainable parameters. This work highlights the potential of model bridging as a powerful alternative to conventional cross-modal knowledge transfer techniques, offering a pathway to more accessible, adaptable, modality-agnostic, and resource-efficient biosignal applications in real-world settings, where computing resources and labelled data are often limited.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec14_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

ETHICAL AND REPRODUCIBILITY STATEMENT

[ "ETHICAL AND REPRODUCIBILITY STATEMENT" ]

0

This study makes use of datasets involving human subjects (ISRUC, WESAD, and FOG). All datasets employed are publicly available, and we follow the usage terms and ethical guidelines specified by the original data providers. No new data were collected for this work, and all analyses were conducted on de-identified, previously published datasets.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec14_p1

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

ETHICAL AND REPRODUCIBILITY STATEMENT

[ "ETHICAL AND REPRODUCIBILITY STATEMENT" ]

1

To ensure reproducibility, we provide detailed descriptions of our experimental setups, including data preprocessing steps, model architectures, hyperparameters, and training procedures, in Section 4.1 and Appendix D. Our code implementation is available in the supplementary materials and will be available in a dedicated repository upon publication.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec15_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

APPENDIX TABLE OF CONTENTS

[ "APPENDIX TABLE OF CONTENTS" ]

0

A l denote the matrix of old modality representations from the l-th layer. The CKA linear operator introduced in Eq. 5 is formulated as follows (Kornblith et al., 2019):

[]

[ { "raw": "Kornblith et al., 2019", "authors": "Kornblith et al.", "year": "2019" } ]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec15_p1

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

APPENDIX TABLE OF CONTENTS

[ "APPENDIX TABLE OF CONTENTS" ]

1

HSIC is the Hilbert-Schmidt Independence Criterion. H is the centering matrix. Note that while this formulation uses the entire H (new) m and H (old) l to compute similarity between representations of the old and new modalities, it is also possible to improve efficiency by computing similarity using only a subset of their rows.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec16_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

B LIMITATIONS AND FUTURE WORK

[ "B LIMITATIONS AND FUTURE WORK" ]

0

Although BioX-Bridge greatly reduces training computational requirements and improves the efficiency of cross-modal knowledge transfer, it depends on the availability of pre-trained models for each biosignal modality, an assumption that may not hold for emerging or underexplored biosignals. Furthermore, depending on the position of the bridge, the inference time of the bridged model could Each foundation model specifies its preprocessing pipeline, and we follow these procedures accordingly. If a notch or bandpass filter has already been applied to the dataset, we skip that step during preprocessing.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec16_p1

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

B LIMITATIONS AND FUTURE WORK

[ "B LIMITATIONS AND FUTURE WORK" ]

1

WESAD In this dataset, ECG signals are sampled at 700Hz and PPG signals at 64Hz. For ECG, which uses HuBERT-ECG as its foundation model, we first downsample to 500Hz, apply a finite impulse response (FIR) bandpass filter between 0.05-47Hz, resample to 100Hz, and then perform channel-wise z-score normalization. For PPG, which uses PaPaGei as its foundation model, we upsample the signals to 125Hz, apply a 4th-order Chebyshev bandpass filter between 0.5-12Hz, and normalize using z-score normalization. Finally, all recordings are segmented into 60-second windows with a 5-second step size.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec16_p2

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

B LIMITATIONS AND FUTURE WORK

[ "B LIMITATIONS AND FUTURE WORK" ]

2

FOG In this dataset, both EEG and EMG signals were collected at 1000Hz, downsampled to 500Hz with a notch and bandpass filter already applied. For EEG, which uses LaBraM as its foundation model, we downsample to 200Hz and convert the unit to 0.1mV. For EMG, which uses NormWear as its foundation model, we downsample to 130Hz and normalize using z-score normalization. All recordings are segmented into 3-second windows with a sliding step size of 0.3 seconds.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec16_p3

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

B LIMITATIONS AND FUTURE WORK

[ "B LIMITATIONS AND FUTURE WORK" ]

3

ISRUC In this dataset, EEG signals are sampled at 200Hz with a notch and bandpass filter already applied, while PPG signals are also sampled at 200Hz with a notch filter applied. For EEG, which uses LaBraM as its foundation model, no resampling is required; we only convert the unit to 0.1mV. For ECG, which uses HuBERT-ECG as its foundation model, we upsample to 500Hz, apply a bandpass filter between 0.05-47Hz, resample to 100Hz, and normalize using z-score normalization.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec17_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

D.1.2 DATASET SPLIT

[ "D.1.2 DATASET SPLIT" ]

0

Dataset split is summarized in Figure A1. The datasets contain synchronized data from the old and new modalities. We perform a subject-wise split for WESAD and ISRUC and sample-wise split for FOG (Zhang et al., 2022) to obtain four subsets D (old) , D (new) , D (val) , and D (pair) , at a ratio of 33%, 22%, 11%, and 33%, respectively. We use old modality data from D (old) to train the linear prober g (old) ω . New modality data from D (new) is used to evaluate bridge performance in an unseen set. All data from D (pair) and D (val) are used to train and help select hyperparameters.

[]

[ { "raw": "Zhang et al., 2022", "authors": "Zhang et al.", "year": "2022" } ]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec18_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

New modality data

[ "New modality data" ]

0

Old modality data

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec20_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

≈33% ≈22%

[ "≈33% ≈22%" ]

0

Evaluate Unused

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec21_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

≈11%

[ "≈11%" ]

0

Hyperparameter Selection Unsupervised Training Figure A1: Illustration of Dataset Split. The dataset is divided into four subject-independent subsets. We first use the old modality data from D (old) to train the linear prober g (old) ω for experiment setup, followed by unsupervised training on D (pair) . The subsets D (val) and D (new) are used for hyperparameter selection and testing, respectively.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec22_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

D.1.3 BIOX-BRIDGE AND BASELINE IMPLEMENTATION DETAILS

[ "D.1.3 BIOX-BRIDGE AND BASELINE IMPLEMENTATION DETAILS" ]

0

BioX-Bridge Implementation Details To prepare f (old) θ and g (old) ω for evaluation, we adapt the pre-trained foundation models for the classification tasks using D (old) , we apply mean pooling to the last-layer representations and add a linear layer for classification. The weights of the foundation model are frozen, and the linear layer is trained for more than 50 epochs. We select a learning rate of 1e-4 for LaBraM, 1e-4 for PaPaGei, 1e-4 for NormWear, and 1e-5 for HuBERT-ECG, as suggested in the original publications. To select the input position of the bridge, we use logistic regression with L2 regularization as the linear prober. We select the layer with the best F1 macro score over five-fold cross-validation. We train the bridge over 50 epochs with cosine embedding loss on D (pair) . All experiments are conducted on V100 GPUs with 32GB VRAM. The training time for a single run ranges from 10 minutes to 4 hours, depending on the dataset and the backbone foundation model.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec22_p1

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

D.1.3 BIOX-BRIDGE AND BASELINE IMPLEMENTATION DETAILS

[ "D.1.3 BIOX-BRIDGE AND BASELINE IMPLEMENTATION DETAILS" ]

1

To select the bridge rank and number of prototypes, we perform a grid search over r = {4, 8, 16, 32} and N p = {50, 100, 150, 200, 250, 300}. For compatibility across various network architectures without modifying the forward functions, we retrieve and replace intermediate representations using forward hooks and forward pre-hooks3 . Please see the source code for detailed implementation.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec22_p2

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

D.1.3 BIOX-BRIDGE AND BASELINE IMPLEMENTATION DETAILS

[ "D.1.3 BIOX-BRIDGE AND BASELINE IMPLEMENTATION DETAILS" ]

2

Baseline Implementation Details For the Random baseline, we simulate a model that randomly assigns labels to samples uniformly across all classes. For CardioGAN, we adopt the pre-trained weights provided by the original publication. We do not perform any fine-tuning as CardioGAN has been trained with WESAD as one of its datasets. We adopt the preprocessing code provided by CargioGAN to prepare the PPG. For each recording, we generate the corresponding ECG using a sliding window of 4 seconds and a step size of 4 seconds. The generated ECG recording is then preprocessed following the HuBERT-ECG pipeline described in Section D.1.1.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec22_p3

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

D.1.3 BIOX-BRIDGE AND BASELINE IMPLEMENTATION DETAILS

[ "D.1.3 BIOX-BRIDGE AND BASELINE IMPLEMENTATION DETAILS" ]

3

For knowledge distillation (KD), we append a linear layer to map mean-pooled representations from the last layer of the foundation model to classwise probabilities. We then fine-tune the entire foundation model with a learning rate of 1e-4 for LaBraM, 1e-4 for PaPaGei, 1e-4 for NormWear, and 1e-5 for HuBERT-ECG, as suggested in the original publications. For contrastive knowledge distillation (KD-Contrast), we append a linear layer to map mean-pooled representations from the last layer of the new modality foundation model to those of the old modality foundation model. During inference, we adopt the linear prober from the old modality foundation model to produce classwise probabilities. The temperature for the InfoNCE loss is selected as 0.04 following (Abbaspourazad et al., 2024b).

[]

[ { "raw": "Abbaspourazad et al., 2024b", "authors": "Abbaspourazad et al.", "year": "2024b" } ]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec23_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

D.2 MAIN RESULTS WITH STANDARD DEVIATIONS

[ "D.2 MAIN RESULTS WITH STANDARD DEVIATIONS" ]

0

Tables 1-3 only report the average performance for five seeds due to space constraints. In Table A1-A5, we present the full results with standard deviation.

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec24_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

D.3 ADDITIONAL ABLATION STUDIES D.3.1 BIOX-BRIDGE FRAMEWORK WITH TRADITIONAL BIOSIGNAL MODELS

[ "D.3 ADDITIONAL ABLATION STUDIES D.3.1 BIOX-BRIDGE FRAMEWORK WITH TRADITIONAL BIOSIGNAL MODELS" ]

0

All results presented in the main body focused on building a bridge between biosignal foundation models to better support ongoing research in this area. To show that BioX-Bridge is also compatible with traditional biosignal models, we replace HuBERT-ECG with a pre-trained ECG model, ECG-DualNet (Rohr et al., 2022), in Table A6. We notice that BioX-Bridge continues to outperform the baseline methods while significantly reducing the number of trainable parameters

[]

[ { "raw": "Rohr et al., 2022", "authors": "Rohr et al.", "year": "2022" } ]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_1__sec25_p0

Artificial Intelligence_1

BIOX-BRIDGE: MODEL BRIDGING FOR UNSUPERVISED CROSS-MODAL KNOWLEDGE TRANSFER ACROSS BIOSIGNALS

D.3.2 IMPACT OF FOUNDATION MODEL SIZE

[ "D.3.2 IMPACT OF FOUNDATION MODEL SIZE" ]

0

The HubERT-ECG family of models is available in three sizes, with varying numbers of parameters. We replace the small version of HuBERT-ECG (30M) with base (93M) and large (183M) versions, and report both baseline and our results in Table A7. We observe that the best transfer performance for BioX-Bridge is achieved using the large version of HuBERT-ECG, thanks to the more powerful representational capabilities. Overall, using base and large model yields better transfer performance

[]

2,025

null

https://arxiv.org//pdf/2510.02276

[ { "firstname": "Chenqi", "surname": "Li", "email": "[email protected]" }, { "firstname": "Yu", "surname": "Liu", "email": "[email protected]" }, { "firstname": "Timothy", "surname": "Denison", "email": "[email protected]" }, { "firstname": "Tingting", "surname": "Zhu", "email": "[email protected]" } ]

[]

Artificial Intelligence_10_abstract_p0

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

Abstract

[ "Abstract" ]

0

In tabular anomaly detection (AD), textual semantics often carry critical signals, as the definition of an anomaly is closely tied to domain-specific context. However, existing benchmarks provide only raw data points without semantic context, overlooking rich textual metadata such as feature descriptions and domain knowledge that experts rely on in practice. This limitation restricts research flexibility and prevents models from fully leveraging domain knowledge for detection. ReTabAD addresses this gap by Restoring textual semantics to enable context-aware Tabular Anomaly Detection research. We provide (1) 20 carefully curated tabular datasets enriched with structured textual metadata, together with implementations of stateof-the-art AD algorithms-including classical, deep learning, and LLM-based approaches-and (2) a zero-shot LLM framework that leverages semantic context without task-specific training, establishing a strong baseline for future research. Furthermore, this work provides insights into the role and utility of textual metadata in AD through experiments and analysis. Results show that semantic context improves detection performance and enhances interpretability by supporting domain-aware reasoning. These findings establish ReTabAD as a benchmark for systematic exploration of context-aware AD. The resource is publicly released at https://yoonsanghyu.github.io/ReTabAD/.

[]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec0_p0

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

INTRODUCTION

[ "INTRODUCTION" ]

0

Anomaly detection (AD), the task of identifying patterns that diverge from the distribution of normal data, represents a fundamental challenge across diverse domains such as finance (Al-Hashedi & Magalingam, 2021), cybersecurity (Landauer et al., 2023), manufacturing (Kharitonov et al., 2022), and healthcare (Fernando et al., 2021). In particular, anomaly detection in tabular data, i.e., tabular AD, is of critical importance, as structured datasets-including sensor logs, transaction records, and clinical measurements-form the backbone of real-world operational systems. Despite its practical importance, tabular AD remains a challenging problem due to the heterogeneous nature of tabular datasets, which often combine numerical, categorical, and even textual attributes with rich semantic meaning.

[]

[ { "raw": "Al-Hashedi & Magalingam, 2021", "authors": "Al-Hashedi & Magalingam", "year": "2021" }, { "raw": "Landauer et al., 2023", "authors": "Landauer et al.", "year": "2023" }, { "raw": "Kharitonov et al., 2022", "authors": "Kharitonov et al.", "year": "2022" }, { "raw": "Fernando et al., 2021", "authors": "Fernando et al.", "year": "2021" } ]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec0_p1

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

INTRODUCTION

[ "INTRODUCTION" ]

1

DAMI Repository (Campos et al., 2016) and ADBench (Han et al., 2022) have made significant contributions to tabular AD by establishing standardized benchmarks that enabled fair and rigorous comparisons across AD algorithms. Despite this contribution, these benchmarks were developed under earlier detection paradigms that prioritized only numerical features with conventional machine learning pipelines-for instance, converting categorical values into arbitrary integer codes, or discarding non-numerical fields (as shown in Figure 1). More critically, they excluded textual metadata that practitioners routinely rely on in real-world applications, e.g., feature descriptions, domain context, categorical semantics, measurement units, and operational constraints. This omission is particularly problematic because the definition of an anomaly is inherently "context-dependent" (Chandola et al., 2009;Pang et al., 2021)-without understanding feature semantics and domain constraints, models may misclassify benign deviations as anomalies or fail to detect subtle but critical abnormalities. Recent advances in Large language models (LLMs) demonstrate improved capabilities for handling textual representations of numerical data, particularly in contextual reasoning (Kojima et al., 2022;Fons et al., 2024). These advances suggest new opportunities for leveraging semantic metadata in tabular AD that has been missing from traditional approaches. In particular, such textual metadata can serve as valuable domain priors-enabling models to form more plausible decision boundaries, better interpret anomalous behavior, and generalize more effectively in small-data settings. Early efforts such as AnoLLM (Tsai et al., 2025) have begun to explore this direction, revealing the potential of LLM-based approaches in tabular settings. However, their application remains constrained to utilizing only column names, due to the lack of richer textual annotations in existing benchmarks. This disconnect underscores the need for a new benchmark that restores semantic context, allowing researchers to meaningfully evaluate and advance context-aware tabular AD.

[]

[ { "raw": "Campos et al., 2016", "authors": "Campos et al.", "year": "2016" }, { "raw": "Han et al., 2022", "authors": "Han et al.", "year": "2022" }, { "raw": "Chandola et al., 2009", "authors": "Chandola et al.", "year": "2009" }, { "raw": "Pang et al., 2021", "authors": "Pang et al.", "year": "2021" }, { "raw": "Kojima et al., 2022", "authors": "Kojima et al.", "year": "2022" }, { "raw": "Fons et al., 2024", "authors": "Fons et al.", "year": "2024" }, { "raw": "Tsai et al., 2025", "authors": "Tsai et al.", "year": "2025" } ]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec0_p2

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

INTRODUCTION

[ "INTRODUCTION" ]

2

Motivated by these considerations, we introduce ReTabAD, the first context-aware tabular AD benchmark that incorporates rich textual metadata into AD, enabling models to ground their predictions in semantic context rather than purely numerical patterns. As highlighted in Table 1, our benchmark prioritizes quality by curating 20 tabular datasets with rich textual metadata and carefully validated anomaly definitions. This design stands in contrast to ADBench (Han et al., 2022), which achieves broader coverage by simply transforming CV/NLP datasets into tabular format-an approach that increases dataset quantity but lacks the semantic context necessary for context-aware AD. Alongside these curated datasets, ReTabAD enables comprehensive evaluation across 17 algorithms, ranging from classical and deep learning methods to the latest LLM-based approach. Furthermore, we also propose a zero-shot LLM baseline that leverages the provided metadata without any task-specific training to establish a fair starting point for context-aware AD. Empirically, incorporating textual metadata improves zero-shot LLM performance by an average of +7.6% AUROC across multiple models (see Table 4). As a result, our zero-shot LLM approach achieves performance comparable to state-of-the-art training-based methods, despite never being trained on task-specific data. Together, these contributions position ReTabAD as a unified and forward-looking benchmark that enables systematic study of tabular AD in both traditional and context-aware paradigms.

[]

[ { "raw": "Han et al., 2022", "authors": "Han et al.", "year": "2022" } ]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec0_p3

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

INTRODUCTION

[ "INTRODUCTION" ]

3

To sum up, our contributions are summarized as follows:

[]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec0_p4

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

INTRODUCTION

[ "INTRODUCTION" ]

4

• The first context-aware tabular AD benchmark: We present ReTabAD, which provides 20 carefully curated tabular datasets enriched with textual metadata and faithful anomaly definitions, and enables standardized evaluation across 16 unsupervised algorithms, including classical, deep learning, and LLM-based approaches.

[]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec0_p5

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

INTRODUCTION

[ "INTRODUCTION" ]

5

• Zero-shot tabular AD framework leveraging textual metadata: We introduce and evaluate a zero-shot LLM framework that establishes not only a strong and competitive baseline but also a new evaluation paradigm for context-aware tabular AD. By leveraging semantic context without task-specific training, ReTabAD provides a standardized reference point for future research.

[]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec0_p6

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

INTRODUCTION

[ "INTRODUCTION" ]

6

• Insights into context-aware tabular AD: We systematically analyze how incorporating textual metadata affects not only model performance but also reasoning quality. Specifically, we evaluate alignment between predicted key features and supervised attributions, and qualitatively examine model-generated reasoning texts, providing insights into how context enhances interpretability and anomaly attribution. Overall, our work emphasizes the importance of semantic information (e.g., textual metadata) in AD for tabular data. We hope that this perspective, together with our new benchmark, will guide future research toward developing context-aware approaches.

[]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec1_p0

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

LIMITATIONS ON PREVIOUS TABULAR AD BENCHMARKS

[ "LIMITATIONS ON PREVIOUS TABULAR AD BENCHMARKS" ]

0

Benchmarking efforts in tabular AD have primarily aimed to provide reproducible and systematic comparisons of algorithms across diverse datasets. The comparative studies (Goldstein & Uchida, 2016;Campos et al., 2016;Ruff et al., 2021b) established the first generation of evaluation resources, offering systematic comparisons of a broad range of algorithms on real-world datasets. In particular, the DAMI Repository (Campos et al., 2016) introduced semantically meaningful tabular datasets with well-defined anomaly classes, demonstrating the importance of defining anomalies in a way that reflects meaningful deviations. With the rise of deep learning, ADBench (Han et al., 2022) further advanced the field by integrating both classical and deep anomaly detection algorithms, collecting datasets from multiple domains, and analyzing them from diverse perspectives. Together with open-source libraries such as PyOD (Zhao et al., 2019) and DeepOD (Xu et al., 2023a), these resources have substantially advanced the field by enabling standardized protocols and large-scale experimental studies. Despite these contributions, existing benchmarks lack textual metadata that makes semantic context explicit. Although anomaly classes are often well defined, the datasets are typically provided only in numerical form, and the conversion processes-such as one-hot encoding, text embedding, or dimensionality reduction-inevitably discard important semantic information. Some benchmarks also incorporate datasets from non-tabular modalities like images and speech, where feature-level semantics cannot be clearly specified. As a result, most prior research has been confined to improving performance within these numeric-only settings, focusing on better modeling of given datasets rather than leveraging the rich textual information that practitioners use in real-world anomaly detection.

[]

[ { "raw": "Goldstein & Uchida, 2016", "authors": "Goldstein & Uchida", "year": "2016" }, { "raw": "Campos et al., 2016", "authors": "Campos et al.", "year": "2016" }, { "raw": "Ruff et al., 2021b", "authors": "Ruff et al.", "year": "2021b" }, { "raw": "Campos et al., 2016", "authors": "Campos et al.", "year": "2016" }, { "raw": "Han et al., 2022", "authors": "Han et al.", "year": "2022" }, { "raw": "Zhao et al., 2019", "authors": "Zhao et al.", "year": "2019" }, { "raw": "Xu et al., 2023a", "authors": "Xu et al.", "year": "2023a" } ]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec2_p0

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

RETABAD: TABULAR AD BENCHMARK RESTORING SEMANTIC CONTEXT

[ "RETABAD: TABULAR AD BENCHMARK RESTORING SEMANTIC CONTEXT" ]

0

To overcome these limitations as described in Section 2, we introduce ReTabAD, a next-generation benchmark that focuses exclusively on genuinely tabular datasets while systematically integrating rich semantic information. This design enables the systematic evaluation of context-aware AD. Label: Binary label (0: Normal, 1: Suspect/Pathologic)

[]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec3_p0

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

DATASETS

[ "DATASETS" ]

0

Data Collection and Annotation. ReTabAD covers 20 tabular datasets drawn from widely used repositories and real-world application domains. Building upon the comprehensive coverage of ADBench (Han et al., 2022), we curate a subset of datasets and enrich them with textual metadata according to the following criteria: (1) datasets must have a clearly defined domain and ground-truth anomaly labels, (2) dataset size should be manageable for diverse models while reflecting real-world class imbalance, (3) sufficient documentation should exist to recover reliable semantic descriptions, and (4) datasets that are too easy, where performance is already saturated, should be discarded. In addition, we incorporate new datasets by inspecting their class label definitions and class-wise sample ratios, and then annotating anomalies accordingly to construct well-posed AD tasks. A complete list of the datasets and their description is summarized in Table 7 in the Appendix A and an overview of the data collection and annotation process is illustrated in Figure 1.

[]

[ { "raw": "Han et al., 2022", "authors": "Han et al.", "year": "2022" } ]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec3_p1

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

DATASETS

[ "DATASETS" ]

1

Semantically-rich Tabular Data. To incorporate rich semantic information into tabular data, we consider the following two principles:

[]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec3_p2

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

DATASETS

[ "DATASETS" ]

2

• Raw Numerical Preservation: ReTabAD preserves numerical features in their original scales rather than applying unspecified normalization or standardization. Maintaining raw values retains domain-meaningful interpretations that support intuitive reasoning about anomalies. For instance, a resting heart rate of 200 bpm in an adult is directly recognizable as a critical medical anomaly, whereas its normalized counterpart only conveys statistical rarity without clinical significance.

[]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec3_p3

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

DATASETS

[ "DATASETS" ]

3

• Categorical Text Restoration: Existing benchmarks often provide categorical features in inconsistent formats (e.g., integer-encoded, text-embeddings) or apply inappropriate normalization that discards semantic meaning. Contrary to this practice, ReTabAD restores categorical attributes to their original textual values instead of applying such transformations. This restoration enables models, particularly language-based approaches, to reason over human-interpretable categories rather than arbitrary numerical surrogates.

[]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec3_p4

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

DATASETS

[ "DATASETS" ]

4

Structured Text Metadata. We provide each dataset supplemented with a JSON-formatted metadata file, containing structured semantic descriptions omitted in previous benchmarks. To ensure reliability, we provide direct links to the original data sources, enabling researchers to verify our interpretations.

[]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec3_p5

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

DATASETS

[ "DATASETS" ]

5

The metadata is organized into three categories:

[]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec3_p6

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

DATASETS

[ "DATASETS" ]

6

• Dataset-Level Descriptions: High-level descriptors, including dataset name, purpose, origin, and collection methodology. Each entry includes links to original publications or repositories to enhance credibility.

[]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec3_p7

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

DATASETS

[ "DATASETS" ]

7

• Column-Level Descriptions: For each column, ReTabAD provides its name, human-readable description, logical type (e.g., numerical, categorical, ordinal, binary), measurement units, if available. All descriptions are cross-referenced with the original dataset documentation.

[]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec3_p8

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

DATASETS

[ "DATASETS" ]

8

• Label-Level Descriptions: For each dataset, ReTabAD clearly specifies which classes are considered normal and which are treated as anomalies. These anomaly definitions are derived directly from the original dataset documentation, with full source attribution.

[]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec4_p0

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

ALGORITHMS

[ "ALGORITHMS" ]

0

To provide a comprehensive comparison, we evaluate 16 representative unsupervised AD models, with full algorithmic details provided in Appendix B. These models are grouped into three categories:

[]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec4_p1

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

ALGORITHMS

[ "ALGORITHMS" ]

1

(1) Classical methods: We include a diverse range of foundational algorithms that remain widely used. These encompass distance-based (e.g., KNN (Ramaswamy et al., 2000)), density-based (e.g., LOF (Breunig et al., 2000)), boundary-based (e.g., OCSVM (Schölkopf et al., 1999)), and ensemble techniques (e.g., Isolation Forest (Liu et al., 2008)). (2) Deep learning methods: We consider representative deep models designed to capture complex, non-linear representations of normality beyond classical heuristics. These include methods based on deep one-class classification (e.g., DeepSVDD (Ruff et al., 2018)), reconstruction (e.g., RCA (Liu et al., 2021), MCM (Yin et al., 2024)), and contrastive learning (e.g., GOAD (Bergman & Hoshen, 2020), NeuTraL (Qiu et al., 2021)). (3) LLM-based methods: We include AnoLLM (Tsai et al., 2025), which leverages semantic information but is constrained to using only column names as context. To further leverage rich and structured metadata, we propose a new zero-shot LLM framework that directly utilizes context data as input for AD. This investigation illustrates the potential of applying LLMs to AD tasks when comprehensive semantic cues are available. Implementation details are provided in Section 3.3.

[]

[ { "raw": "Ramaswamy et al., 2000", "authors": "Ramaswamy et al.", "year": "2000" }, { "raw": "Breunig et al., 2000", "authors": "Breunig et al.", "year": "2000" }, { "raw": "Schölkopf et al., 1999", "authors": "Schölkopf et al.", "year": "1999" }, { "raw": "Liu et al., 2008", "authors": "Liu et al.", "year": "2008" }, { "raw": "Ruff et al., 2018", "authors": "Ruff et al.", "year": "2018" }, { "raw": "Liu et al., 2021", "authors": "Liu et al.", "year": "2021" }, { "raw": "Yin et al., 2024", "authors": "Yin et al.", "year": "2024" }, { "raw": "Bergman & Hoshen, 2020", "authors": "Bergman & Hoshen", "year": "2020" }, { "raw": "Qiu et al., 2021", "authors": "Qiu et al.", "year": "2021" }, { "raw": "Tsai et al., 2025", "authors": "Tsai et al.", "year": "2025" } ]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec5_p0

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

ZERO-SHOT LLM FRAMEWORK FOR CONTEXT-AWARE AD

[ "ZERO-SHOT LLM FRAMEWORK FOR CONTEXT-AWARE AD" ]

0

Motivation. Despite the importance of textual metadata in real-world tabular datasets, there has been no systematic attempt to leverage such information. Conventional training-based AD models are designed to process data as numerical or categorical vectors, making it structurally difficult to directly integrate textual descriptions that contain the semantic information of features. Although there have been attempts to utilize textual metadata, most have remained limited-for instance, relying only on column name without capturing the broader context. As a consequence, such approaches fail to leverage critical signals for AD conveyed by metadata, including inter-feature relationships and domain-specific contextual cues. To address this limitation, we propose a zero-shot LLM-based AD framework that explores the potential of semantic context through context-aware reasoning over tabular data. By enabling the LLM to interpret feature semantics and relationships in natural language, our framework shifts the focus from purely statistical patterns to high-level contextual understanding and points toward a new paradigm for context-aware AD.

[]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec5_p1

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

ZERO-SHOT LLM FRAMEWORK FOR CONTEXT-AWARE AD

[ "ZERO-SHOT LLM FRAMEWORK FOR CONTEXT-AWARE AD" ]

1

Our Design. To utilize LLMs for tabular AD, we propose a zero-shot LLM framework as a strong baseline for context-aware AD, as illustrated in Figure 2. We employ a carefully designed prompt that integrates domain knowledge and structured reasoning guidelines. The prompt for each data instance x i consists of three key components: a System Prompt (S), a Data Input Formatter (T (x i )), and a Structured Output Query (Q).

[]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec6_p0

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

System Prompt (S):

[ "System Prompt (S):" ]

0

The system prompt provides the necessary context for the task. It is constructed by concatenating the following information(detailed templates are in the Appendix D.3):

[]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec6_p1

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

System Prompt (S):

[ "System Prompt (S):" ]

1

• Role and Task Definition (S role ): Assigns the LLM an expert persona for each dataset's domain (e.g., "You are a domain expert in finance"). { "anomaly_score": s, "key_features": F, "reasoning": e } where s is the anomaly score in [0, 1], F is a list of key features, and e is the explanation in text.

[ 0, 1 ]

[]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[ { "id": 1, "text": "Financial fraud detection applying data mining techniques: A comprehensive review from 2009 to 2019\n\t\t\n\t\t\tAl-HashediKhaled Gubran\n\t\t\n\t\t\n\t\t\tPritheegaMagalingam\n\t\t\n\t\n\t\n\t\tComputer Science Review\n\t\t\n\t\t\t40\n\t\t\t100402\n\t\t\t2021" } ]

Artificial Intelligence_10__sec7_p0

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

EVALUATION

[ "EVALUATION" ]

0

Problem Setup. The concept of normality, a ground-truth law of normal behavior for a given task (Ruff et al., 2021a), is modeled through a probability distribution P with corresponding density function p(x). Conventional tabular anomaly detectors map each data instance x ∈ R K to a class label y ∈ {0, 1} by learning this distribution, where K denotes the feature dimension, and 0 and 1 indicate normal and anomalous status, respectively. ReTabAD extends this traditional setup by incorporating semantic metadata M, which encompasses domain knowledge, feature descriptions, and contextual information about normal behavior patterns. The primary objective is to model a conditional distribution p(x|M) that leverages semantic context for enhanced AD. This work focuses on analyzing the impact of semantic context M on detection performance.

[]

[ { "raw": "Ruff et al., 2021a", "authors": "Ruff et al.", "year": "2021a" } ]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec7_p1

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

EVALUATION

[ "EVALUATION" ]

1

Hyperparameter Optimization. We follow the default model architectures and tune key hyperparameters, including preprocessing settings. Hyperparameters for training-based methods are faithfully searched, and detailed configurations are provided in the Appendix C for reproducibility.

[]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec7_p2

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

EVALUATION

[ "EVALUATION" ]

2

Evaluator LLMs. We evaluate a set of leading LLMs in a zero-shot AD setting to assess their general semantic reasoning capabilities. Specifically, we include state-of-the-art reasoning-oriented models-GPT-4.1 (Fachada et al., 2025), Claude-3.7-sonnet (Anthropic, 2025), Qwen3-235B (Yang et al., 2025), and Gemini-2.5-pro (Comanici et al., 2025)-as well as a lightweight yet competitive model, GPT-4o-mini (OpenAI, 2024). For consistency and clarity, we report detailed ablation and qualitative analyses with Gemini-2.5-pro.

[]

[ { "raw": "Fachada et al., 2025", "authors": "Fachada et al.", "year": "2025" }, { "raw": "Anthropic, 2025", "authors": "Anthropic", "year": "2025" }, { "raw": "Yang et al., 2025", "authors": "Yang et al.", "year": "2025" }, { "raw": "Comanici et al., 2025", "authors": "Comanici et al.", "year": "2025" }, { "raw": "OpenAI, 2024", "authors": "OpenAI", "year": "2024" } ]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[ { "id": 2, "text": "Anthropic\n\t\t\n\t\t\n\t\t\n\t\t\tAugust 2025" }, { "id": 23, "text": "GPT-4o, DALL-E, and Whisper\n\t\t\n\t\t\tShimonIfrah\n\t\t\n\t\t10.1007/979-8-8688-0599-8_6\n\t\t\n\t\n\t\n\t\tGetting Started with Azure OpenAI\n\t\t\n\t\t\tApress\n\t\t\t2024" } ]

Artificial Intelligence_10__sec7_p3

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

EVALUATION

[ "EVALUATION" ]

3

Evaluation Protocol. ReTabAD follows the one-class classification setting, where the training set consists of 50% of the normal samples, while the test set consists of the remaining normal samples and all anomalous samples, as in previous studies (Bergman & Hoshen, 2020;Livernoche et al., 2023). For performance evaluation, we primarily report the Area Under the Receiver Operating Characteristic Curve (AUROC), a standard metric in AD (Ruff et al., 2018). To mitigate randomness, each experiment is run five times with different seeds, and Full results with the Area Under the Precision-Recall Curve (AUPRC) (Saito & Rehmsmeier, 2015) and standard deviations are in the Appendix E, F.

[]

[ { "raw": "Bergman & Hoshen, 2020", "authors": "Bergman & Hoshen", "year": "2020" }, { "raw": "Livernoche et al., 2023", "authors": "Livernoche et al.", "year": "2023" }, { "raw": "Ruff et al., 2018", "authors": "Ruff et al.", "year": "2018" }, { "raw": "Saito & Rehmsmeier, 2015", "authors": "Saito & Rehmsmeier", "year": "2015" } ]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec8_p0

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

EXPERIMENTS ON RETABAD BENCHMARK

[ "EXPERIMENTS ON RETABAD BENCHMARK" ]

0

This section explores the potential of context-aware approaches in tabular AD. Specifically, we design our experiments to investigate the following research questions:

[]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]

Artificial Intelligence_10__sec8_p1

Artificial Intelligence_10

RETABAD: A BENCHMARK FOR RESTORING SEMAN-TIC CONTEXT IN TABULAR ANOMALY DETECTION

EXPERIMENTS ON RETABAD BENCHMARK

[ "EXPERIMENTS ON RETABAD BENCHMARK" ]

1

• Does semantic context M improve AD performance? ( §4.1)

[]

2,025

null

https://arxiv.org//pdf/2510.02060

[ { "firstname": "Sanghyu", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Dongmin", "surname": "Kim", "email": "[email protected]" }, { "firstname": "Suhee", "surname": "Yoon", "email": "[email protected]" }, { "firstname": "Ye", "surname": "Sim", "email": "" }, { "firstname": "Seungdong", "surname": "Yoa", "email": "[email protected]" }, { "firstname": "Hye-Seung", "surname": "Cho", "email": "[email protected]" }, { "firstname": "Soonyoung", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Hankook", "surname": "Lee", "email": "[email protected]" }, { "firstname": "Woohyung", "surname": "Lim", "email": "[email protected]" } ]

[]