Selected work

Representative Papers

Adaptive Visual Token Reduction for Accelerated Image Understanding

Oct 06, 2026

This study addresses the high computational overhead and loss of spatial structural information in large vision-language models when processing high-resolution images by proposing the ReFIT framework. This framework introduces a pioneering adaptive window reshaping mechanism that overcomes the limitations of conventional fixed cropping. Through Correlation-guided Window Reshaping (RWR) and Instruction-guided Token Refinement (ITR), it achieves dynamic optimization and efficient reduction of visual tokens, effectively preserving critical spatial structural features such as text. Experiments across four VQA benchmarks demonstrate that the proposed method significantly improves answer accuracy while reducing computational costs, validating the effectiveness of regional localization and information deduplication.

0 citationsRead paper

GRC-Net: Global Representation Consistency Network for Unsupervised Multimodal Anomaly Detection

Oct 06, 2026

Existing local patch-based anomaly detection methods are susceptible to reconstruction noise, resulting in unstable errors within normal regions and hindering the effective identification of multimodal structural anomalies. To address this, this work proposes GRC-Net, which introduces a global representation consistency mechanism. By leveraging global tokens and attention MLPs to capture holistic context, the method suppresses reconstruction noise and integrates a stable reconstruction module for unsupervised multimodal anomaly detection, fundamentally resolving the reconstruction instability inherent in local representations. Extensive experiments on the MVTec 3D-AD and Eyecandies datasets demonstrate that GRC-Net significantly outperforms existing methods in both image-level and pixel-level detection performance, substantially enhancing overall detection robustness.

0 citationsRead paper

Attenuated in-context identification in time-series foundation models: diagnosis under counterfactual inputs and repair by synthetic forced-system fine-tuning

Oct 06, 2026

This study addresses the dynamic response attenuation exhibited by time series foundation models under counterfactual inputs, which impedes accurate prediction of controlled system variations. We propose an evaluation framework based on paired counterfactual inputs that reveals amplitude attenuation deficiencies stemming from pretraining priors. To mitigate this issue, we perform short-horizon fine-tuning on models such as Chronos-2 using synthetically generated forced system data. Experimental results demonstrate that fine-tuning restores sensitivity to 0.83–0.96 and, for specific nonlinear systems, surpasses classical structure-agnostic identification methods. This work provides an effective pathway for enhancing the physical consistency of time series foundation models.

0 citationsRead paper

M2P-AD: Memory-to-Prototype Learning with Boundary-aware Score Refinement for 3D Anomaly Detection

Jul 15, 2026

This work addresses the challenges of over-response in normal regions and false positives near object boundaries in 3D anomaly detection by proposing a novel approach that integrates a Memory-to-Prototype (M2P) module with a Boundary-aware Score Refinement (BSR) strategy. The M2P module leverages a memory bank to learn representative prototypes of normal features, enabling precise modeling of the normal distribution. Concurrently, a boundary extraction module is introduced to facilitate structure-aware correction of anomaly scores through BSR, effectively preserving geometric integrity while suppressing boundary artifacts. Evaluated on Real3D-AD, Anomaly-ShapeNet, and MulSen-AD datasets, the proposed method significantly reduces false alarms in normal regions and boundary-related false positives, achieving more accurate and robust anomaly localization and outperforming current state-of-the-art methods.

0 citationsRead paper

Two Confounds in Cross-Model Value Comparison: Response Determinism and the Access Harness

Jul 11, 2026

This study addresses the confounding of genuine value differences with variations in response determinism in cross-model value comparisons, which is further exacerbated by interference from evaluation harnesses, leading to mischaracterizations of models' individualized values. To disentangle these effects, the authors propose a determinism-corrected decomposition framework that leverages rule-free value dilemmas, repeated forced-choice experiments, and a determinism index to isolate true value divergence from determinism-related artifacts. The approach also systematically evaluates the impact of deployment interfaces—such as APIs and client-side implementations—on value expression. Experiments across nine mainstream language models reveal substantial inter-model differences in determinism (ranging from 0.66 to 0.95) and show that correcting for determinism markedly reduces apparent individualization. Notably, different interfaces induce value profile shifts of up to 0.31 and can even reverse specific moral judgments, thereby uncovering—for the first time—the formative role of the deployment layer in shaping model-expressed values.

0 citationsRead paper

Recent publications

Latest Papers

Adaptive Visual Token Reduction for Accelerated Image Understanding

Oct 06, 2026

This study addresses the high computational overhead and loss of spatial structural information in large vision-language models when processing high-resolution images by proposing the ReFIT framework. This framework introduces a pioneering adaptive window reshaping mechanism that overcomes the limitations of conventional fixed cropping. Through Correlation-guided Window Reshaping (RWR) and Instruction-guided Token Refinement (ITR), it achieves dynamic optimization and efficient reduction of visual tokens, effectively preserving critical spatial structural features such as text. Experiments across four VQA benchmarks demonstrate that the proposed method significantly improves answer accuracy while reducing computational costs, validating the effectiveness of regional localization and information deduplication.

0 citationsRead paper

GRC-Net: Global Representation Consistency Network for Unsupervised Multimodal Anomaly Detection

Oct 06, 2026

Existing local patch-based anomaly detection methods are susceptible to reconstruction noise, resulting in unstable errors within normal regions and hindering the effective identification of multimodal structural anomalies. To address this, this work proposes GRC-Net, which introduces a global representation consistency mechanism. By leveraging global tokens and attention MLPs to capture holistic context, the method suppresses reconstruction noise and integrates a stable reconstruction module for unsupervised multimodal anomaly detection, fundamentally resolving the reconstruction instability inherent in local representations. Extensive experiments on the MVTec 3D-AD and Eyecandies datasets demonstrate that GRC-Net significantly outperforms existing methods in both image-level and pixel-level detection performance, substantially enhancing overall detection robustness.

0 citationsRead paper

Attenuated in-context identification in time-series foundation models: diagnosis under counterfactual inputs and repair by synthetic forced-system fine-tuning

Oct 06, 2026

This study addresses the dynamic response attenuation exhibited by time series foundation models under counterfactual inputs, which impedes accurate prediction of controlled system variations. We propose an evaluation framework based on paired counterfactual inputs that reveals amplitude attenuation deficiencies stemming from pretraining priors. To mitigate this issue, we perform short-horizon fine-tuning on models such as Chronos-2 using synthetically generated forced system data. Experimental results demonstrate that fine-tuning restores sensitivity to 0.83–0.96 and, for specific nonlinear systems, surpasses classical structure-agnostic identification methods. This work provides an effective pathway for enhancing the physical consistency of time series foundation models.

0 citationsRead paper

M2P-AD: Memory-to-Prototype Learning with Boundary-aware Score Refinement for 3D Anomaly Detection

Jul 15, 2026

This work addresses the challenges of over-response in normal regions and false positives near object boundaries in 3D anomaly detection by proposing a novel approach that integrates a Memory-to-Prototype (M2P) module with a Boundary-aware Score Refinement (BSR) strategy. The M2P module leverages a memory bank to learn representative prototypes of normal features, enabling precise modeling of the normal distribution. Concurrently, a boundary extraction module is introduced to facilitate structure-aware correction of anomaly scores through BSR, effectively preserving geometric integrity while suppressing boundary artifacts. Evaluated on Real3D-AD, Anomaly-ShapeNet, and MulSen-AD datasets, the proposed method significantly reduces false alarms in normal regions and boundary-related false positives, achieving more accurate and robust anomaly localization and outperforming current state-of-the-art methods.

0 citationsRead paper

Two Confounds in Cross-Model Value Comparison: Response Determinism and the Access Harness

Jul 11, 2026

This study addresses the confounding of genuine value differences with variations in response determinism in cross-model value comparisons, which is further exacerbated by interference from evaluation harnesses, leading to mischaracterizations of models' individualized values. To disentangle these effects, the authors propose a determinism-corrected decomposition framework that leverages rule-free value dilemmas, repeated forced-choice experiments, and a determinism index to isolate true value divergence from determinism-related artifacts. The approach also systematically evaluates the impact of deployment interfaces—such as APIs and client-side implementations—on value expression. Experiments across nine mainstream language models reveal substantial inter-model differences in determinism (ranging from 0.66 to 0.95) and show that correcting for determinism markedly reduces apparent individualization. Notably, different interfaces induce value profile shifts of up to 0.31 and can even reverse specific moral judgments, thereby uncovering—for the first time—the formative role of the deployment layer in shaping model-expressed values.

0 citationsRead paper