Selected work

Representative Papers

HERA: Historical Evidence Routing Adapter for Physical Prediction in Latent World Models

Aug 05, 2026

This work addresses the challenge of effectively leveraging historical visual evidence under occlusion by introducing HERA, a lightweight adapter that enables non-intrusive injection of historical information into a frozen latent predictor for the first time. HERA employs a Register-Routed Patch Memory (RRPM) architecture, integrating a structured memory bank with a register mechanism to selectively route relevant historical evidence into the predictor’s working space, thereby enabling precise modeling of physical regularities. Evaluated on the IntPhys2 Main benchmark, the approach improves the AvgSurprise accuracy of V-JEPA 2-G from 52.57% to 54.35%, achieving 57.69% on the fixed-camera continuity subtask and 63.46% on the invariance subtask.

0 citationsRead paper

Beyond Experience Retrieval: Learning to Generate Utility-Optimized Structured Experience for Frozen LLMs

Jan 30, 2026

This work addresses the tendency of static large language models to repeatedly commit reasoning errors and the limitations of existing retrieval-based experience reuse methods, which suffer from noise, high latency, and reliance solely on similarity matching. To overcome these issues, the authors propose SEAM—a lightweight, executor-specific plug-in module that eschews conventional retrieval mechanisms by internalizing experience into learnable parameters. SEAM generates structured, instance-tailored experience entries via a single forward pass to guide frozen large language models toward improved reasoning. Trained with a utility-driven GRPO algorithm combined with executor rollouts and further refined through supervised fine-tuning on successful trajectory logs, SEAM enables post-deployment performance gains without modifying the main model. Experiments demonstrate significant accuracy improvements across multiple frozen executors on mathematical reasoning benchmarks, with minimal computational overhead, confirming the method’s effectiveness and robustness.

0 citationsRead paper

COMPARE: Clinical Optimization with Modular Planning and Assessment via RAG-Enhanced AI-OCT: Superior Decision Support for Percutaneous Coronary Intervention Compared to ChatGPT-5 and Junior Operators

Dec 11, 2025

Optical coherence tomography (OCT)-guided percutaneous coronary intervention (PCI) heavily relies on operator expertise, and general-purpose AI models lack domain-specific reliability for cardiovascular interventions. Method: We propose CA-GPT—the first large language model specifically designed for OCT-PCI—integrating a modular clinical decision architecture with a medical retrieval-augmented generation (RAG) mechanism to support end-to-end tasks: semantic parsing of coronary anatomy, pre-procedural planning (lumen diameter/length measurement), and post-procedural assessment (stent apposition, expansion, etc.). Results: In real-world clinical validation, CA-GPT achieved inter-rater agreement rates of 90.3% (diameter) and 80.6% (length) with expert cardiologists—significantly outperforming ChatGPT-4o and junior physicians (p<0.01). For post-procedural evaluation, it attained perfect median consistency (5/5; IQR: 4.75–5) and demonstrated robust performance in complex cases. This study provides the first clinical validation of domain-specialized LLMs for OCT-PCI, establishing their feasibility and superiority over generalist models and novice practitioners.

0 citationsRead paper

Sigma: The Key for Vision-Language-Action Models toward Telepathic Alignment

Nov 30, 2025

Humanoid robot cognitive systems lack a dynamically updatable intermediate “mental space” bridging semantic understanding and continuous action control. Method: This paper proposes Sigma—a novel Vision-Language-Action-Alignment (VLAA) architecture—built upon the pi05_base foundation model. Sigma achieves cross-temporal “intention alignment” by deeply integrating hierarchical semantic understanding with associative modeling, while preserving textual semantic fidelity and empathetic norms. It employs LoRA fine-tuning, inference-time adapter optimization, and offline closed-loop replay, trained on a customized svla_so101_pickplace dataset. Results: Sigma demonstrates consistent MSE reduction across vector-, segment-, and full-trajectory control levels. It establishes, for the first time, a quantifiable and reproducible dynamic mental space—enabling real-time, semantics-driven behavior generation without retraining the base model.

0 citationsRead paper

SURGE: On the Potential of Large Language Models as General-Purpose Surrogate Code Executors

Feb 16, 2025

This work investigates whether large language models (LLMs) can serve as universal “proxy code executors”—accurately predicting program outputs and behaviors without actual execution. Method: We introduce SURGE, a comprehensive benchmark comprising eight challenging domains: multi-language programming, competitive algorithm solving, repository-scale analysis, high-cost scientific computation, time-complexity-intensive tasks, erroneous code diagnosis, compiler- or environment-dependent programs, and formal mathematical proof verification. We conduct the first systematic evaluation of LLMs’ proxy execution feasibility across cross-lingual, high-cost, environment-sensitive, and formal verification dimensions, propose a novel error attribution taxonomy, and perform scaling studies across open- and closed-source models of varying sizes, integrating program behavior modeling and execution trace prediction. Contribution/Results: Results show that LLMs exhibit nascent proxy execution capability on certain tasks, but generalization remains limited; model scale and training data volume yield nonlinear performance gains.

0 citationsRead paper

Recent publications

Latest Papers

HERA: Historical Evidence Routing Adapter for Physical Prediction in Latent World Models

Aug 05, 2026

This work addresses the challenge of effectively leveraging historical visual evidence under occlusion by introducing HERA, a lightweight adapter that enables non-intrusive injection of historical information into a frozen latent predictor for the first time. HERA employs a Register-Routed Patch Memory (RRPM) architecture, integrating a structured memory bank with a register mechanism to selectively route relevant historical evidence into the predictor’s working space, thereby enabling precise modeling of physical regularities. Evaluated on the IntPhys2 Main benchmark, the approach improves the AvgSurprise accuracy of V-JEPA 2-G from 52.57% to 54.35%, achieving 57.69% on the fixed-camera continuity subtask and 63.46% on the invariance subtask.

0 citationsRead paper

Beyond Experience Retrieval: Learning to Generate Utility-Optimized Structured Experience for Frozen LLMs

Jan 30, 2026

This work addresses the tendency of static large language models to repeatedly commit reasoning errors and the limitations of existing retrieval-based experience reuse methods, which suffer from noise, high latency, and reliance solely on similarity matching. To overcome these issues, the authors propose SEAM—a lightweight, executor-specific plug-in module that eschews conventional retrieval mechanisms by internalizing experience into learnable parameters. SEAM generates structured, instance-tailored experience entries via a single forward pass to guide frozen large language models toward improved reasoning. Trained with a utility-driven GRPO algorithm combined with executor rollouts and further refined through supervised fine-tuning on successful trajectory logs, SEAM enables post-deployment performance gains without modifying the main model. Experiments demonstrate significant accuracy improvements across multiple frozen executors on mathematical reasoning benchmarks, with minimal computational overhead, confirming the method’s effectiveness and robustness.

0 citationsRead paper

COMPARE: Clinical Optimization with Modular Planning and Assessment via RAG-Enhanced AI-OCT: Superior Decision Support for Percutaneous Coronary Intervention Compared to ChatGPT-5 and Junior Operators

Dec 11, 2025

Optical coherence tomography (OCT)-guided percutaneous coronary intervention (PCI) heavily relies on operator expertise, and general-purpose AI models lack domain-specific reliability for cardiovascular interventions. Method: We propose CA-GPT—the first large language model specifically designed for OCT-PCI—integrating a modular clinical decision architecture with a medical retrieval-augmented generation (RAG) mechanism to support end-to-end tasks: semantic parsing of coronary anatomy, pre-procedural planning (lumen diameter/length measurement), and post-procedural assessment (stent apposition, expansion, etc.). Results: In real-world clinical validation, CA-GPT achieved inter-rater agreement rates of 90.3% (diameter) and 80.6% (length) with expert cardiologists—significantly outperforming ChatGPT-4o and junior physicians (p<0.01). For post-procedural evaluation, it attained perfect median consistency (5/5; IQR: 4.75–5) and demonstrated robust performance in complex cases. This study provides the first clinical validation of domain-specialized LLMs for OCT-PCI, establishing their feasibility and superiority over generalist models and novice practitioners.

0 citationsRead paper

Sigma: The Key for Vision-Language-Action Models toward Telepathic Alignment

Nov 30, 2025

Humanoid robot cognitive systems lack a dynamically updatable intermediate “mental space” bridging semantic understanding and continuous action control. Method: This paper proposes Sigma—a novel Vision-Language-Action-Alignment (VLAA) architecture—built upon the pi05_base foundation model. Sigma achieves cross-temporal “intention alignment” by deeply integrating hierarchical semantic understanding with associative modeling, while preserving textual semantic fidelity and empathetic norms. It employs LoRA fine-tuning, inference-time adapter optimization, and offline closed-loop replay, trained on a customized svla_so101_pickplace dataset. Results: Sigma demonstrates consistent MSE reduction across vector-, segment-, and full-trajectory control levels. It establishes, for the first time, a quantifiable and reproducible dynamic mental space—enabling real-time, semantics-driven behavior generation without retraining the base model.

0 citationsRead paper

SURGE: On the Potential of Large Language Models as General-Purpose Surrogate Code Executors

Feb 16, 2025

This work investigates whether large language models (LLMs) can serve as universal “proxy code executors”—accurately predicting program outputs and behaviors without actual execution. Method: We introduce SURGE, a comprehensive benchmark comprising eight challenging domains: multi-language programming, competitive algorithm solving, repository-scale analysis, high-cost scientific computation, time-complexity-intensive tasks, erroneous code diagnosis, compiler- or environment-dependent programs, and formal mathematical proof verification. We conduct the first systematic evaluation of LLMs’ proxy execution feasibility across cross-lingual, high-cost, environment-sensitive, and formal verification dimensions, propose a novel error attribution taxonomy, and perform scaling studies across open- and closed-source models of varying sizes, integrating program behavior modeling and execution trace prediction. Contribution/Results: Results show that LLMs exhibit nascent proxy execution capability on certain tasks, but generalization remains limited; model scale and training data volume yield nonlinear performance gains.

0 citationsRead paper