Selected work

Representative Papers

X-OPM: Explainable Automatic Digital On-Chip Power Modeling for Enhanced Robustness

Oct 06, 2026

This study addresses the lack of physical interpretability and insufficient cross-workload generalization in existing on-chip power consumption models by proposing a robust feature engineering framework grounded in VLSI design principles. Methodologically, it introduces a human-in-the-loop workflow to balance modeling accuracy against overhead, integrating tree-based models to capture nonlinear interactions with linear models for prediction, while leveraging EDA tools for layout verification. Experimental results demonstrate that the proposed model achieves an R² exceeding 0.93 with an area overhead below 0.1%. It outperforms state-of-the-art methods such as APOLLO in predictive accuracy and maintains stable generalization performance under previously unseen workloads.

0 citationsRead paper

Rewarding Novel Deductions: Solver-guided Process Rewards for Logical Reasoning

Sep 28, 2026

This study addresses the lack of process supervision in the logical reasoning of large language models and the tendency of smaller models to produce inconsistent and redundant steps by proposing the SPRING framework. This method introduces the first formal definition of a "novel reasoning step" and employs an SMT solver to verify the logical validity of intermediate deductions. Based on this verification, it constructs a process-level reward signal that balances consistency with information gain, guiding the model via reinforcement learning to generate high-quality reasoning chains. Experimental results demonstrate that SPRING significantly outperforms existing baselines on benchmarks such as ZebraLogic, achieving accuracy improvements of up to 49.71 percentage points and reaching 93.14% accuracy on the Knights and Knaves task.

0 citationsRead paper

Safe Greenhouse Climate Control Using Lagrangian-Constrained PPO with Kolmogorov-Arnold Networks

Sep 28, 2026

This study addresses the limitation of fixed penalties in traditional reinforcement learning for greenhouse control, which struggle to constrain long-term climate violations. We formulate greenhouse regulation as a Constrained Markov Decision Process (CMDP) and propose a Lagrangian safe reinforcement learning framework based on RCPO-PPO. The core innovation lies in pioneering the integration of Kolmogorov-Arnold Networks (KANs) in place of standard MLPs to enhance nonlinear representation, combined with sinusoidal temporal features to capture diurnal cycles, thereby achieving decoupled optimization of economic returns and climate safety. Experimental results demonstrate that, compared to baseline methods, the proposed approach reduces cumulative climate violations by 18.65% while increasing lettuce profit by 2.91%, effectively balancing production efficiency with long-term risk mitigation.

0 citationsRead paper

RoboChrono: A Real Robot Benchmark for Streaming Task Understanding

Sep 28, 2026

This study addresses the persistent challenge of correlating visual observations with interaction histories and task progress in robotic manipulation. To this end, it constructs a streaming task understanding benchmark grounded in real-world robot and human demonstrations, encompassing seven task categories such as recognition, alignment, and temporal localization. Moving beyond aggregate scoring, the work introduces a diagnostic evaluation perspective to conduct zero-shot assessments and input ablation studies across eighteen vision-language models. The findings reveal that strong visual matching capabilities do not inherently translate to temporal ordering proficiency, quantifying the performance gap between them. Furthermore, the analysis demonstrates that next-action prediction relies predominantly on priors rather than real-time visual evidence. Collectively, this research establishes a novel paradigm for evaluating specific model capabilities in embodied AI contexts.

0 citationsRead paper

OCA: ODE-Driven Cross-Attention for Image-to-Point-Cloud Registration

Sep 28, 2026

This study addresses the attention ambiguity problem inherent in cross-attention mechanisms for image–point cloud registration by proposing an ODE-driven cross-attention module. To our knowledge, this work is the first to introduce ordinary differential equations into cross-attention, modeling ideal feature interaction dynamics to jointly optimize the attention matrix and feature representations. This formulation significantly enhances the discriminability of 2D–3D feature correspondences and can be seamlessly integrated into existing frameworks. Extensive evaluations across four benchmark datasets demonstrate that the proposed method improves registration recall by 5%, 9%, and 15% under standard, fine-tuned, and zero-shot settings, respectively, thereby validating both its effectiveness and generalization capability.

0 citationsRead paper

Recent publications

Latest Papers

X-OPM: Explainable Automatic Digital On-Chip Power Modeling for Enhanced Robustness

Oct 06, 2026

This study addresses the lack of physical interpretability and insufficient cross-workload generalization in existing on-chip power consumption models by proposing a robust feature engineering framework grounded in VLSI design principles. Methodologically, it introduces a human-in-the-loop workflow to balance modeling accuracy against overhead, integrating tree-based models to capture nonlinear interactions with linear models for prediction, while leveraging EDA tools for layout verification. Experimental results demonstrate that the proposed model achieves an R² exceeding 0.93 with an area overhead below 0.1%. It outperforms state-of-the-art methods such as APOLLO in predictive accuracy and maintains stable generalization performance under previously unseen workloads.

0 citationsRead paper

Rewarding Novel Deductions: Solver-guided Process Rewards for Logical Reasoning

Sep 28, 2026

This study addresses the lack of process supervision in the logical reasoning of large language models and the tendency of smaller models to produce inconsistent and redundant steps by proposing the SPRING framework. This method introduces the first formal definition of a "novel reasoning step" and employs an SMT solver to verify the logical validity of intermediate deductions. Based on this verification, it constructs a process-level reward signal that balances consistency with information gain, guiding the model via reinforcement learning to generate high-quality reasoning chains. Experimental results demonstrate that SPRING significantly outperforms existing baselines on benchmarks such as ZebraLogic, achieving accuracy improvements of up to 49.71 percentage points and reaching 93.14% accuracy on the Knights and Knaves task.

0 citationsRead paper

Safe Greenhouse Climate Control Using Lagrangian-Constrained PPO with Kolmogorov-Arnold Networks

Sep 28, 2026

This study addresses the limitation of fixed penalties in traditional reinforcement learning for greenhouse control, which struggle to constrain long-term climate violations. We formulate greenhouse regulation as a Constrained Markov Decision Process (CMDP) and propose a Lagrangian safe reinforcement learning framework based on RCPO-PPO. The core innovation lies in pioneering the integration of Kolmogorov-Arnold Networks (KANs) in place of standard MLPs to enhance nonlinear representation, combined with sinusoidal temporal features to capture diurnal cycles, thereby achieving decoupled optimization of economic returns and climate safety. Experimental results demonstrate that, compared to baseline methods, the proposed approach reduces cumulative climate violations by 18.65% while increasing lettuce profit by 2.91%, effectively balancing production efficiency with long-term risk mitigation.

0 citationsRead paper

RoboChrono: A Real Robot Benchmark for Streaming Task Understanding

Sep 28, 2026

This study addresses the persistent challenge of correlating visual observations with interaction histories and task progress in robotic manipulation. To this end, it constructs a streaming task understanding benchmark grounded in real-world robot and human demonstrations, encompassing seven task categories such as recognition, alignment, and temporal localization. Moving beyond aggregate scoring, the work introduces a diagnostic evaluation perspective to conduct zero-shot assessments and input ablation studies across eighteen vision-language models. The findings reveal that strong visual matching capabilities do not inherently translate to temporal ordering proficiency, quantifying the performance gap between them. Furthermore, the analysis demonstrates that next-action prediction relies predominantly on priors rather than real-time visual evidence. Collectively, this research establishes a novel paradigm for evaluating specific model capabilities in embodied AI contexts.

0 citationsRead paper

OCA: ODE-Driven Cross-Attention for Image-to-Point-Cloud Registration

Sep 28, 2026

This study addresses the attention ambiguity problem inherent in cross-attention mechanisms for image–point cloud registration by proposing an ODE-driven cross-attention module. To our knowledge, this work is the first to introduce ordinary differential equations into cross-attention, modeling ideal feature interaction dynamics to jointly optimize the attention matrix and feature representations. This formulation significantly enhances the discriminability of 2D–3D feature correspondences and can be seamlessly integrated into existing frameworks. Extensive evaluations across four benchmark datasets demonstrate that the proposed method improves registration recall by 5%, 9%, and 15% under standard, fine-tuned, and zero-shot settings, respectively, thereby validating both its effectiveness and generalization capability.

0 citationsRead paper