Reproduce “Towards Detecting LLMs Hallucination via Markov Chain-based Multi-agent Debate Framework” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiaoxi Sun, Jinpeng Li, Yan Zhong, Dongyan Zhao, Rui Yan
A Markov Chain-based multi-agent debate framework detects hallucinations in Large Language Model outputs without requiring model retraining, simulating human-like discussions among specialized LLM agents to rigorously verify claims against retrieved evidence. This approach led to improved hallucination detection accuracy across Knowledge Base Question Answering, Summarization, and Dialogue tasks.
Key insights
- Modeling the multi-agent debate as a Markov chain allows for dynamic adjustment of debate modes (Trust-initiated vs. Skeptic-initiated) based on the outcomes of previous rounds, enabling a more adaptive verification process.
- Assigning specialized roles (Trust, Skeptic, Leader) to LLM agents promotes diverse perspectives and a thorough, human-like debate, where claims are challenged and synthesized.
- Challenging claims initially deemed factual by transitioning to a Skeptic-initiated debate mode proves particularly effective for optimal hallucination detection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.