MARCH: Three-Agent Reinforcement Learning Pipeline Makes 8B Model Match Closed-Source LLMs on Hallucination Detection

arXiv·high signalXBlueskyLinkedInCopy link

Li et al. introduce MARCH, a multi-agent reinforced self-check framework where a Solver generates RAG responses, a Proposer decomposes them into atomic claims, and a Checker validates propositions in isolation — deliberately deprived of the Solver's output to prevent confirmation bias. Using multi-agent RL for co-evolution, an 8B parameter model with MARCH achieves hallucination detection competitive with powerful closed-source models. The information asymmetry design between agents is the key architectural insight.