Selected work

Representative Papers

End-to-end Optimization of Belief and Policy Learning in Shared Autonomy Paradigms

Jan 30, 2026

This work addresses the fundamental challenge in shared autonomy of enhancing task success while preserving user autonomy by efficiently inferring user intent and dynamically adjusting assistance levels. The authors propose BRACE, a novel framework that achieves, for the first time, end-to-end joint optimization of Bayesian intention inference and context-adaptive assistance policies, conditioning the policy on both environmental context and the full belief distribution over goals. Theoretical analysis reveals that optimal assistance levels should decrease with reduced goal uncertainty and increase under stronger environmental constraints, and that incorporating belief information yields a quadratic reduction in expected regret. Empirical evaluations across three task domains demonstrate BRACE’s superiority over state-of-the-art methods, improving task success rates by 6.3% and path efficiency by 41% compared to IDA and DQN, respectively, and achieving gains of 36.3% in success rate and 87% in path efficiency over unassisted control.

1 citationsRead paper

Adapting Vision Foundation Models to Acoustics for Pose-Free 3D Sonar Reconstruction

Sep 05, 2026

研究通过利用视觉基础模型与声学之间的几何关系及物理噪声模型生成合成数据，解决了缺乏大规模声纳数据集的问题，实现了基于声纳的无姿态3D重建。

0 citationsRead paper

The Fallacy of Independent Ceilings: Characterizing Coupled Load-Branch Stall Interaction

Aug 11, 2026

This work challenges the conventional practice of treating branch misprediction and cache miss penalties as independent performance bottlenecks, which overlooks their coupling effects under irregular workloads and consequently underestimates the potential of joint optimization. The study is the first to reveal and quantify the phenomenon of symbiotic stall latency (SSL), introduces the Symbiotic Stall Opportunity (SSO) metric to efficiently identify highly coupled workloads, and proposes Joint Speedup Synergy (JSS) to evaluate the benefits of coordinated optimization. Through lightweight MPKI analysis, full-system simulation, and oracle experiments across 53 workloads, the authors find that 70% exhibit significant coupling, with 40% surpassing the theoretical upper bound of independent speedups by over 6%; notably, high-SSO workloads achieve JSS values of 1.23–3.29. These results expose the limitations of traditional decoupled analysis and lead to the identification of four canonical software coupling patterns.

0 citationsRead paper

LLM-Based Embeddings for Program Analysis and Optimization

Aug 07, 2026

This work proposes a novel approach to program analysis and optimization leveraging large language models (LLMs). Addressing the challenge of effectively integrating source code and intermediate representation (IR) information—a limitation in existing methods—it introduces LLMCompiler, pre-trained on IR, and employs a chunked embedding and aggregation strategy to produce unified program-level embeddings. By innovatively unifying the semantics of source code and IR, the method achieves a 1.54% error rate on algorithm classification, representing a 12% improvement over the current state of the art. It also attains competitive accuracy in heterogeneous device mapping, significantly advancing the application of LLMs in program understanding and optimization.

0 citationsRead paper

AdaptoNet: Modular Foundation-Adaptive Neural Networks for Cyber-Physical Attack Detection in Power Grids

Aug 02, 2026

This work addresses cyber-physical attacks on power systems that combine physical disturbances with data manipulation, particularly data denial attacks that undermine conventional detection methods. To tackle this challenge, the authors propose AdaptoNet, a modular neural network architecture that dynamically activates trainable adaptive modules based on a binary measurement availability vector, while keeping the base modules frozen and thus avoiding retraining. This design enables effective discrimination between genuine anomalies and denied measurements even under partial observability. Experimental results on four IEEE benchmark systems demonstrate that, under measurement blocking attacks affecting up to 20% of sensors, AdaptoNet restores the F1 score from below 12% to over 81%, closely approaching the baseline performance of 89%–99% achieved with complete data.

0 citationsRead paper

Recent publications

Latest Papers

Adapting Vision Foundation Models to Acoustics for Pose-Free 3D Sonar Reconstruction

Sep 05, 2026

研究通过利用视觉基础模型与声学之间的几何关系及物理噪声模型生成合成数据，解决了缺乏大规模声纳数据集的问题，实现了基于声纳的无姿态3D重建。

0 citationsRead paper

The Fallacy of Independent Ceilings: Characterizing Coupled Load-Branch Stall Interaction

Aug 11, 2026

This work challenges the conventional practice of treating branch misprediction and cache miss penalties as independent performance bottlenecks, which overlooks their coupling effects under irregular workloads and consequently underestimates the potential of joint optimization. The study is the first to reveal and quantify the phenomenon of symbiotic stall latency (SSL), introduces the Symbiotic Stall Opportunity (SSO) metric to efficiently identify highly coupled workloads, and proposes Joint Speedup Synergy (JSS) to evaluate the benefits of coordinated optimization. Through lightweight MPKI analysis, full-system simulation, and oracle experiments across 53 workloads, the authors find that 70% exhibit significant coupling, with 40% surpassing the theoretical upper bound of independent speedups by over 6%; notably, high-SSO workloads achieve JSS values of 1.23–3.29. These results expose the limitations of traditional decoupled analysis and lead to the identification of four canonical software coupling patterns.

0 citationsRead paper

LLM-Based Embeddings for Program Analysis and Optimization

Aug 07, 2026

This work proposes a novel approach to program analysis and optimization leveraging large language models (LLMs). Addressing the challenge of effectively integrating source code and intermediate representation (IR) information—a limitation in existing methods—it introduces LLMCompiler, pre-trained on IR, and employs a chunked embedding and aggregation strategy to produce unified program-level embeddings. By innovatively unifying the semantics of source code and IR, the method achieves a 1.54% error rate on algorithm classification, representing a 12% improvement over the current state of the art. It also attains competitive accuracy in heterogeneous device mapping, significantly advancing the application of LLMs in program understanding and optimization.

0 citationsRead paper

AdaptoNet: Modular Foundation-Adaptive Neural Networks for Cyber-Physical Attack Detection in Power Grids

Aug 02, 2026

This work addresses cyber-physical attacks on power systems that combine physical disturbances with data manipulation, particularly data denial attacks that undermine conventional detection methods. To tackle this challenge, the authors propose AdaptoNet, a modular neural network architecture that dynamically activates trainable adaptive modules based on a binary measurement availability vector, while keeping the base modules frozen and thus avoiding retraining. This design enables effective discrimination between genuine anomalies and denied measurements even under partial observability. Experimental results on four IEEE benchmark systems demonstrate that, under measurement blocking attacks affecting up to 20% of sensors, AdaptoNet restores the F1 score from below 12% to over 81%, closely approaching the baseline performance of 89%–99% achieved with complete data.

0 citationsRead paper

Lights, Camera, Carbon: Architectural Scaling Laws for Video Generation Energy Consumption

Jul 05, 2026

This work addresses the lack of interpretable and standardized methods for evaluating the energy consumption of text-to-video generation models under unknown architectures and parameter counts. The authors propose a weight- and implementation-agnostic bidirectional estimation framework grounded in first principles and observable generation parameters—such as output resolution and duration—to establish, for the first time, a theoretical scaling law linking model energy consumption to architectural complexity. By decomposing the quadratic and linear energy components inherent in diffusion model inference and integrating empirical GPU measurements with observed inference times, the framework retrodicts architectural efficiency and enables a unified sustainability benchmark across diverse models. Validated on six open-source models (8.3B–27B parameters) across three GPU types, the approach achieves an average absolute percentage error below 3% in energy prediction, accurately capturing architectural differences.

0 citationsRead paper