Selected work

Representative Papers

Structured Pruning of Large Language Models via Power Transformation and Sign-Preserving Score Aggregation with Adaptive Feature Retention

Jul 08, 2026

This work addresses the challenges of transferring adaptive feature retention (AFR) from unstructured to structured pruning, which include heterogeneous pruning score distributions, loss of sign information, and outlier interference. To bridge this gap, the authors propose a unified structured pruning framework that introduces power transformation to align score distributions, designs a sign-preserving aggregation mechanism to maintain consistent optimization directions, and incorporates a percentile-based outlier removal strategy. This approach effectively narrows the performance gap between structured and unstructured pruning, achieving accuracy on Llama-3-8B, Vicuna-v1.5-13B, and LLaVA-v1.5-13B models that closely matches unstructured pruning while delivering substantial real-world inference speedups.

0 citationsRead paper

SPG: Sparse-Projected Guides with Sparse Autoencoders for Zero-Shot Anomaly Detection

Apr 03, 2026

This work addresses zero-shot anomaly detection and segmentation—identifying anomalies on unseen categories without target-domain adaptation—by proposing SPG, a prompt-free framework. Departing from conventional prompting strategies, SPG introduces sparse projection guidance for the first time: leveraging frozen foundation models (e.g., DINOv3 or OpenCLIP ViT-L/14@336px), it learns sparse guidance coefficients in two stages—first training a sparse autoencoder (SAE) and then optimizing only the guidance coefficients to produce normal and anomaly guidance vectors. These coefficients map sparsely to a small set of SAE dictionary atoms, revealing both category-agnostic and category-specific factors. Experiments demonstrate that SPG achieves state-of-the-art image-level detection performance under cross-dataset zero-shot settings on MVTec AD and VisA, and sets new records in pixel-level segmentation AUROC, particularly with the DINOv3 backbone.

0 citationsRead paper

OD-RASE: Ontology-Driven Risk Assessment and Safety Enhancement for Autonomous Driving

Mar 06, 2026

This study addresses the limitations of existing road infrastructure, which is predominantly designed for human drivers and often fails to ensure safety in rare or complex scenarios encountered by autonomous vehicles, with safety enhancements typically lagging behind accident occurrences. To bridge this gap, the authors propose OD-RASE, a novel framework that integrates domain-specific traffic ontologies with large vision-language models (LVLMs) to accurately identify road structures contributing to accidents and automatically generate interpretable infrastructure improvement recommendations. The approach combines ontology-driven data filtering, LVLM-based reasoning, and diffusion-model-enhanced visualization, and introduces the first annotated dataset for this task. Experimental results demonstrate that OD-RASE achieves high precision in predicting high-risk road configurations and produces actionable retrofitting strategies, significantly enhancing the proactive safety of autonomous driving systems.

0 citationsRead paper

Bidirectional Action Sequence Learning for Long-term Action Anticipation with Large Language Models

Aug 01, 2025

Unidirectional modeling in long-horizon video action forecasting struggles to capture semantically heterogeneous sub-actions. Method: We propose BiAnt, the first framework that deeply integrates bidirectional action sequence learning with large language models (LLMs). BiAnt employs an encoder-decoder architecture that jointly performs forward future prediction and backward contextual reconstruction, leveraging LLMs to explicitly model semantic dependencies and temporal symmetry among actions—thereby overcoming representational limitations of conventional unidirectional models. Contribution/Results: On the Ego4D benchmark, BiAnt achieves significant improvements in edit distance over state-of-the-art baselines, empirically validating the efficacy of bidirectional collaborative reasoning for long-term action anticipation. This work establishes a novel, interpretable, and robust action forecasting paradigm, particularly beneficial for safety-critical applications such as autonomous driving and service robotics requiring early risk identification.

0 citationsRead paper

Multi-modal Integration Analysis of Alzheimer's Disease Using Large Language Models and Knowledge Graphs

May 21, 2025

Alzheimer’s disease (AD) research faces challenges in integrating heterogeneous, unpaired, and cross-cohort multimodal data—including MRI, gene expression, biomarkers, EEG, and clinical metrics—due to their distributed nature and lack of subject-level alignment. Method: We propose the first large language model (LLM)-driven knowledge graph reasoning framework enabling population-level, ID-agnostic, concept-level cross-modal association mining and natural language hypothesis generation. Our approach integrates multimodal statistical feature selection, cross-cohort cross-validation, and expert consensus evaluation (Cohen’s κ = 0.82). Contribution/Results: We identify a novel pathological cascade—“metabolic risk → neuroinflammation → tau dysregulation”—and robust frontal EEG–gene expression associations (r = 0.42–0.58, p < 0.01; high-significance links: r > 0.6, p < 0.001), with effect sizes stable across cohorts (variance < 15%). These findings yield testable, mechanistically grounded hypotheses for AD pathogenesis and therapeutic target discovery.

0 citationsRead paper

Recent publications

Latest Papers

Structured Pruning of Large Language Models via Power Transformation and Sign-Preserving Score Aggregation with Adaptive Feature Retention

Jul 08, 2026

This work addresses the challenges of transferring adaptive feature retention (AFR) from unstructured to structured pruning, which include heterogeneous pruning score distributions, loss of sign information, and outlier interference. To bridge this gap, the authors propose a unified structured pruning framework that introduces power transformation to align score distributions, designs a sign-preserving aggregation mechanism to maintain consistent optimization directions, and incorporates a percentile-based outlier removal strategy. This approach effectively narrows the performance gap between structured and unstructured pruning, achieving accuracy on Llama-3-8B, Vicuna-v1.5-13B, and LLaVA-v1.5-13B models that closely matches unstructured pruning while delivering substantial real-world inference speedups.

0 citationsRead paper

SPG: Sparse-Projected Guides with Sparse Autoencoders for Zero-Shot Anomaly Detection

Apr 03, 2026

This work addresses zero-shot anomaly detection and segmentation—identifying anomalies on unseen categories without target-domain adaptation—by proposing SPG, a prompt-free framework. Departing from conventional prompting strategies, SPG introduces sparse projection guidance for the first time: leveraging frozen foundation models (e.g., DINOv3 or OpenCLIP ViT-L/14@336px), it learns sparse guidance coefficients in two stages—first training a sparse autoencoder (SAE) and then optimizing only the guidance coefficients to produce normal and anomaly guidance vectors. These coefficients map sparsely to a small set of SAE dictionary atoms, revealing both category-agnostic and category-specific factors. Experiments demonstrate that SPG achieves state-of-the-art image-level detection performance under cross-dataset zero-shot settings on MVTec AD and VisA, and sets new records in pixel-level segmentation AUROC, particularly with the DINOv3 backbone.

0 citationsRead paper

OD-RASE: Ontology-Driven Risk Assessment and Safety Enhancement for Autonomous Driving

Mar 06, 2026

This study addresses the limitations of existing road infrastructure, which is predominantly designed for human drivers and often fails to ensure safety in rare or complex scenarios encountered by autonomous vehicles, with safety enhancements typically lagging behind accident occurrences. To bridge this gap, the authors propose OD-RASE, a novel framework that integrates domain-specific traffic ontologies with large vision-language models (LVLMs) to accurately identify road structures contributing to accidents and automatically generate interpretable infrastructure improvement recommendations. The approach combines ontology-driven data filtering, LVLM-based reasoning, and diffusion-model-enhanced visualization, and introduces the first annotated dataset for this task. Experimental results demonstrate that OD-RASE achieves high precision in predicting high-risk road configurations and produces actionable retrofitting strategies, significantly enhancing the proactive safety of autonomous driving systems.

0 citationsRead paper

Bidirectional Action Sequence Learning for Long-term Action Anticipation with Large Language Models

Aug 01, 2025

Unidirectional modeling in long-horizon video action forecasting struggles to capture semantically heterogeneous sub-actions. Method: We propose BiAnt, the first framework that deeply integrates bidirectional action sequence learning with large language models (LLMs). BiAnt employs an encoder-decoder architecture that jointly performs forward future prediction and backward contextual reconstruction, leveraging LLMs to explicitly model semantic dependencies and temporal symmetry among actions—thereby overcoming representational limitations of conventional unidirectional models. Contribution/Results: On the Ego4D benchmark, BiAnt achieves significant improvements in edit distance over state-of-the-art baselines, empirically validating the efficacy of bidirectional collaborative reasoning for long-term action anticipation. This work establishes a novel, interpretable, and robust action forecasting paradigm, particularly beneficial for safety-critical applications such as autonomous driving and service robotics requiring early risk identification.

0 citationsRead paper

Multi-modal Integration Analysis of Alzheimer's Disease Using Large Language Models and Knowledge Graphs

May 21, 2025

Alzheimer’s disease (AD) research faces challenges in integrating heterogeneous, unpaired, and cross-cohort multimodal data—including MRI, gene expression, biomarkers, EEG, and clinical metrics—due to their distributed nature and lack of subject-level alignment. Method: We propose the first large language model (LLM)-driven knowledge graph reasoning framework enabling population-level, ID-agnostic, concept-level cross-modal association mining and natural language hypothesis generation. Our approach integrates multimodal statistical feature selection, cross-cohort cross-validation, and expert consensus evaluation (Cohen’s κ = 0.82). Contribution/Results: We identify a novel pathological cascade—“metabolic risk → neuroinflammation → tau dysregulation”—and robust frontal EEG–gene expression associations (r = 0.42–0.58, p < 0.01; high-significance links: r > 0.6, p < 0.001), with effect sizes stable across cohorts (variance < 15%). These findings yield testable, mechanistically grounded hypotheses for AD pathogenesis and therapeutic target discovery.

0 citationsRead paper