Go to ACL ARR 2025 February homepageExploring Norm Cognition and Compliance in Multi-Agent Dialogue LLM Systems
Abstract: This research investigates norm cognition and compliance behavior of Large Language Models (LLMs) as agents in dialogue systems. We propose a framework called TBC-TBA, which includes two phases: Think-Before-Chat and Think-Before-Act, designed to enhance collaboration efficiency and decision-making quality in multi-agent systems within complex norm cognition scenarios. Our experiments reveal that: 1) LLM agents demonstrate norm cognition behavior; 2) however, they show significant differences from humans in terms of loss aversion, risk preference, and probability cognitive bias; 3) our proposed methods - Dynamic Norm Cognition Mechanism (DNCM), Norm Consequence Emphasis (NCE), Norm Analysis Reflection (NAE), and Norm Case Demonstration (NCD) - can effectively improve agents' norm compliance, with DNCM, which introduces an identify-infer-internalization rule cognition pattern and a new Dynamic Norm Execution Mechanism framework, showing the most significant effects.
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 analysis
Languages Studied: English
Submission Number: 7422
Loading