Reproduce “Epistemic Context Learning: Building Trust the Right Way in LLM-Based Multi-Agent Systems” on OpenResearch

About this paper

Ruiwen Zhou, Maojia Song, Xiaobao Wu, Sitao Cheng, Xunjian Yin, Yuxi Xie, Zhuoqun Hao, Wenyue Hua and 3 more

Researchers from the National University of Singapore and Nanyang Technological University developed Epistemic Context Learning (ECL), a framework enabling Large Language Model (LLM) agents in multi-agent systems to build and leverage historical trust to evaluate peer reliability. ECL enhanced smaller models to outperform much larger baselines and improved frontier models' performance, particularly in adversarial settings, achieving significant accuracy gains.

Key insights

• LLMs, without explicit architectural guidance, exhibit a 'Lack of Historical Trust' and 'Conformity,' primarily relying on current, superficial cues over historical reliability.
• A two-stage reasoning pipeline, explicitly separating trust estimation from current aggregation, provides a beneficial inductive bias for robust trust modeling in LLMs.
• Targeted auxiliary rewards in RL are crucial for effectively grounding and enhancing LLMs' ability to discern reliable peers from history, preventing reward hacking prevalent in single-stage setups.

Summary by alphaXiv.