Self-Refine Learning in LLM Mult-Agent Systems for Norm Cognition and Compliance

ACL ARR 2025 May Submission7133 Authors

Abstract: With the rise of large language models (LLMs) as social simulation agents, understanding how they recognize and follow social norms has become increasingly important. We propose a novel TBC-TBA self-refine learning multi-agent framework integrated with self-refine learning to investigate LLM agents' norm cognition, behavioral alignment with human expectations, and compliance enhancement, thus improving LLM collaboration and decision-making in complex norm scenarios. Our experiments reveal while LLMs can recognize and apply norms partially, they also exhibit reward hacking (RH) that lead to norm violations. Further analysis of alignment with human behavior shows that LLMs are strongly consistent with human moral judgments, but differ in their perception of risk and probability. Our proposed methods including Dynamic Norm Learning Mechanism (DNLM), Deep MaxPain (DMP), Norm Analysis Chain-of-Thought (NA-CoT), and Few-shot Norm Learning (FNL), have been shown to effectively improve the norm compliance of LLMs, with DNLM achieving the most significant impact through its identify-infer-internalization pattern in a novel norm cognition model. The code will be released on GitHub.

Paper Type: Long

Research Area: Dialogue and Interactive Systems

Research Area Keywords: human-in-the-loop,commonsense reasoning,embodied agents,evaluation and metrics

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Data resources, Data analysis

Languages Studied: English

Submission Number: 7133