Selected work

Representative Papers

Vista: Scene-Aware Optimization for Streaming Video Question Answering under Post-Hoc Queries

Feb 09, 2026

This work proposes a scene-aware streaming video question answering framework to address context loss and memory overflow caused by the continuous arrival of video frames and arbitrary-time queries. By integrating dynamic scene segmentation, GPU/CPU-coordinated compressed storage, and a query-driven selective recall mechanism, the approach enables efficient, low-latency comprehension of long videos. Key innovations include dynamic video segmentation based on scene clustering, heterogeneous memory management, index-based retrieval, and a model-agnostic architecture. Evaluated on StreamingBench, the method achieves state-of-the-art performance, significantly enhancing system scalability and inference completeness.

1 citationsRead paper

DIST-FL: Enhancing Security for TEE-based Aggregation in Federated Learning

Jun 03, 2026

This work addresses critical security vulnerabilities in existing Trusted Execution Environment (TEE)-based federated learning systems, which remain susceptible to server-side state rollback and I/O manipulation attacks that compromise privacy and robustness. To mitigate these threats, the authors propose a novel multi-TEE collaborative architecture that leverages an append-only ledger and an operation linearizability mechanism to effectively prevent state rollback within TEEs for the first time. Additionally, the design incorporates trusted input handling to counteract I/O manipulation. The resulting system enables secure and efficient distributed aggregation over wide-area networks, achieving a sixfold increase in throughput compared to prior approaches while preserving the performance of single-TEE deployments, thereby significantly expanding the security guarantees of TEE-based federated learning.

0 citationsRead paper

DIVA: Harnessing the Representation Divergence in Unified Multimodal Models for Mutual Reinforcement

May 24, 2026

This work addresses the challenge of representation interference in unified multimodal models, where conflicting objectives between generation and understanding tasks hinder joint optimization. To resolve this, the authors propose DIVA, a framework that explicitly disentangles visual representations within a unified architecture into shared and task-specific components. By leveraging mutual information estimation, DIVA enables complementary information flow and decoupled training between understanding and generation branches. This approach reveals, for the first time, the complementary internal structure of representations in unified multimodal models. Combined with a self-improving post-training strategy, DIVA significantly enhances collaborative multimodal performance—achieving gains of 7.82% on visual understanding tasks and 8.46% on generation tasks—outperforming all existing baselines comprehensively.

0 citationsRead paper

WindowQuant: Mixed-Precision KV Cache Quantization based on Window-Level Similarity for VLMs Inference Optimization

May 04, 2026

This work addresses the high inference latency and substantial GPU memory consumption of video language models (VLMs), which stem from excessively long visual token sequences. While existing token-level mixed-precision KV cache quantization methods aim to mitigate these issues, they suffer from large search overhead and suboptimal hardware efficiency. To overcome these limitations, we propose WindowQuant, a novel window-level adaptive mixed-precision quantization framework that rapidly determines the optimal bit-width for each visual token window based on its similarity to the textual prompt. Furthermore, WindowQuant enhances hardware utilization through KV cache window reordering. Experimental results demonstrate that WindowQuant consistently outperforms state-of-the-art approaches across multiple benchmarks, achieving significant reductions in both inference latency and memory footprint with negligible accuracy degradation.

0 citationsRead paper

VLA-InfoEntropy: A Training-Free Vision-Attention Information Entropy Approach for Vision-Language-Action Models Inference Acceleration and Success

Apr 06, 2026

Vision-Language-Action models face significant computational overhead and low inference efficiency due to their joint processing of high-dimensional visual inputs, complex language instructions, and continuous action spaces, hindering real-time deployment. To address this, this work proposes a training-free dynamic focusing strategy that dynamically guides the model during inference from global features to local critical regions. The approach leverages image entropy to quantify the textural informativeness of visual tokens and attention entropy to assess textual semantic relevance, integrating spatial, semantic, and temporal cues to identify and prioritize task-relevant regions. This method substantially reduces redundant computation and the number of active parameters during inference, achieving markedly higher inference speed while maintaining or even improving task performance compared to existing approaches.

0 citationsRead paper

Recent publications

Latest Papers

DIST-FL: Enhancing Security for TEE-based Aggregation in Federated Learning

Jun 03, 2026

This work addresses critical security vulnerabilities in existing Trusted Execution Environment (TEE)-based federated learning systems, which remain susceptible to server-side state rollback and I/O manipulation attacks that compromise privacy and robustness. To mitigate these threats, the authors propose a novel multi-TEE collaborative architecture that leverages an append-only ledger and an operation linearizability mechanism to effectively prevent state rollback within TEEs for the first time. Additionally, the design incorporates trusted input handling to counteract I/O manipulation. The resulting system enables secure and efficient distributed aggregation over wide-area networks, achieving a sixfold increase in throughput compared to prior approaches while preserving the performance of single-TEE deployments, thereby significantly expanding the security guarantees of TEE-based federated learning.

0 citationsRead paper

DIVA: Harnessing the Representation Divergence in Unified Multimodal Models for Mutual Reinforcement

May 24, 2026

This work addresses the challenge of representation interference in unified multimodal models, where conflicting objectives between generation and understanding tasks hinder joint optimization. To resolve this, the authors propose DIVA, a framework that explicitly disentangles visual representations within a unified architecture into shared and task-specific components. By leveraging mutual information estimation, DIVA enables complementary information flow and decoupled training between understanding and generation branches. This approach reveals, for the first time, the complementary internal structure of representations in unified multimodal models. Combined with a self-improving post-training strategy, DIVA significantly enhances collaborative multimodal performance—achieving gains of 7.82% on visual understanding tasks and 8.46% on generation tasks—outperforming all existing baselines comprehensively.

0 citationsRead paper

WindowQuant: Mixed-Precision KV Cache Quantization based on Window-Level Similarity for VLMs Inference Optimization

May 04, 2026

This work addresses the high inference latency and substantial GPU memory consumption of video language models (VLMs), which stem from excessively long visual token sequences. While existing token-level mixed-precision KV cache quantization methods aim to mitigate these issues, they suffer from large search overhead and suboptimal hardware efficiency. To overcome these limitations, we propose WindowQuant, a novel window-level adaptive mixed-precision quantization framework that rapidly determines the optimal bit-width for each visual token window based on its similarity to the textual prompt. Furthermore, WindowQuant enhances hardware utilization through KV cache window reordering. Experimental results demonstrate that WindowQuant consistently outperforms state-of-the-art approaches across multiple benchmarks, achieving significant reductions in both inference latency and memory footprint with negligible accuracy degradation.

0 citationsRead paper

VLA-InfoEntropy: A Training-Free Vision-Attention Information Entropy Approach for Vision-Language-Action Models Inference Acceleration and Success

Apr 06, 2026

Vision-Language-Action models face significant computational overhead and low inference efficiency due to their joint processing of high-dimensional visual inputs, complex language instructions, and continuous action spaces, hindering real-time deployment. To address this, this work proposes a training-free dynamic focusing strategy that dynamically guides the model during inference from global features to local critical regions. The approach leverages image entropy to quantify the textural informativeness of visual tokens and attention entropy to assess textual semantic relevance, integrating spatial, semantic, and temporal cues to identify and prioritize task-relevant regions. This method substantially reduces redundant computation and the number of active parameters during inference, achieving markedly higher inference speed while maintaining or even improving task performance compared to existing approaches.

0 citationsRead paper

Confusion-Aware In-Context-Learning for Vision-Language Models in Robotic Manipulation

Mar 16, 2026

This work addresses the susceptibility of vision-language models in robotic manipulation to shortcut learning, which often leads to confusion between visually similar objects, undermining robustness and causing unpredictable failures. To mitigate this issue, the authors propose Confusion-Aware In-Context Learning (CAICL), a novel approach that integrates confusion analysis with in-context learning for the first time. CAICL identifies sources of confusion, dissects error-prone features, and incorporates these insights into prompt design to steer the model toward discriminative cues. The method establishes a confusion-aware learning framework tailored for robotic manipulation, achieving an 85.5% success rate on VIMA-Bench. It consistently alleviates shortcut learning across tasks of varying generalization difficulty, significantly enhancing the model’s ability to distinguish confusable objects and improving operational stability.

0 citationsRead paper