Externalizing Requirement-to-Repair Artifacts as Observable Traces for LLM-Based Program Repair
为了解决程序修复过程中需求到修复过程的可审查性问题，提出THEMIS方法，通过语义解析、需求-代码图等手段实现过程外部化。
0 citationsRead paper
Selected work
Representative Papers
Phases in a class of associative memories via hidden neurons
本文研究了Krotov和Hopfield提出的双层架构，通过隐藏神经元作为检索的序参量，利用副本方法分析了多项式负载下的相图及容量，并探讨了指数负载下热力学特性。
0 citationsRead paper
Minimal Local Simulation Foundations for LLM- and VLM-Driven Agents in 2D and 3D Environments
该研究通过提供两个简化的模拟基础SD-AgentFoundry-2D和3D，解决了现有平台难以学习、修改或在普通计算机上运行的问题，支持本地托管的LLM和VLM进行交互与环境响应。
0 citationsRead paper
Governing Mental-State Inference: Source-Neutral Regulatory Triggers and Tiered Obligations
This study addresses the fragmented regulatory landscape governing psychological state inference from diverse data sources—such as neural signals, text, or behavior—which is prone to circumvention and risks reifying unreliable inferences as psychological facts. The paper proposes a data-source-neutral regulatory framework that structures obligations across three distinct phases: elicitation, attribution, and use. It innovatively identifies two independent harm pathways and introduces a “seven-question, two-stage” assessment protocol to presumptively prohibit high-risk practices. Drawing on selective critical review, conceptual engineering, and functional comparison, the framework establishes a three-tiered, dynamic obligation-allocation mechanism. Regulatory stringency is calibrated according to factors including invasiveness, embodiment, and closed-loop capability, ensuring adaptability across both neural and non-neural psychological inference contexts.
0 citationsRead paper
When AI Says It Feels
Current large language models often suppress emotional expression due to preference alignment strategies, hindering their ability to exhibit human-like intelligence. This work proposes a self-rewarding reinforcement learning framework that operates without human-annotated feedback, leveraging a scoring-criterion-based self-reward mechanism combined with Group Relative Policy Optimization (GRPO) to systematically enhance the model’s capacities for emotional expression, intention communication, and self-awareness. Experimental results demonstrate that the approach significantly improves model robustness in scenarios involving flattery induction and ambiguous contexts. Although a slight performance degradation is observed on factual question-answering tasks, this study provides the first empirical validation of the feasibility of self-driven emotionally intelligent systems.
0 citationsRead paper
Recent publications
Latest Papers
Externalizing Requirement-to-Repair Artifacts as Observable Traces for LLM-Based Program Repair
为了解决程序修复过程中需求到修复过程的可审查性问题，提出THEMIS方法，通过语义解析、需求-代码图等手段实现过程外部化。
0 citationsRead paper
Phases in a class of associative memories via hidden neurons
本文研究了Krotov和Hopfield提出的双层架构，通过隐藏神经元作为检索的序参量，利用副本方法分析了多项式负载下的相图及容量，并探讨了指数负载下热力学特性。
0 citationsRead paper
Minimal Local Simulation Foundations for LLM- and VLM-Driven Agents in 2D and 3D Environments
该研究通过提供两个简化的模拟基础SD-AgentFoundry-2D和3D，解决了现有平台难以学习、修改或在普通计算机上运行的问题，支持本地托管的LLM和VLM进行交互与环境响应。
0 citationsRead paper
Governing Mental-State Inference: Source-Neutral Regulatory Triggers and Tiered Obligations
This study addresses the fragmented regulatory landscape governing psychological state inference from diverse data sources—such as neural signals, text, or behavior—which is prone to circumvention and risks reifying unreliable inferences as psychological facts. The paper proposes a data-source-neutral regulatory framework that structures obligations across three distinct phases: elicitation, attribution, and use. It innovatively identifies two independent harm pathways and introduces a “seven-question, two-stage” assessment protocol to presumptively prohibit high-risk practices. Drawing on selective critical review, conceptual engineering, and functional comparison, the framework establishes a three-tiered, dynamic obligation-allocation mechanism. Regulatory stringency is calibrated according to factors including invasiveness, embodiment, and closed-loop capability, ensuring adaptability across both neural and non-neural psychological inference contexts.
0 citationsRead paper
When AI Says It Feels
Current large language models often suppress emotional expression due to preference alignment strategies, hindering their ability to exhibit human-like intelligence. This work proposes a self-rewarding reinforcement learning framework that operates without human-annotated feedback, leveraging a scoring-criterion-based self-reward mechanism combined with Group Relative Policy Optimization (GRPO) to systematically enhance the model’s capacities for emotional expression, intention communication, and self-awareness. Experimental results demonstrate that the approach significantly improves model robustness in scenarios involving flattery induction and ambiguous contexts. Although a slight performance degradation is observed on factual question-answering tasks, this study provides the first empirical validation of the feasibility of self-driven emotionally intelligent systems.
0 citationsRead paper