MARCH: Multi-Agent Reinforced Self-Check for LLM Hallucination

Dimensions overall score 7.0

• MARCH reduces hallucination rates in LLMs by enforcing rigorous factual alignment through deliberate information asymmetry.method 90%
• MARCH breaks the cycle of self-confirmation bias in hallucination detection by depriving the Checker agent of the Solver's original output.technical 85%
• An 8B-parameter LLM equipped with MARCH achieves performance competitive with powerful closed-source models on hallucination benchmarks.result 90%
• MARCH substantially reduces hallucination rates across hallucination benchmarks compared to existing methods.result 85%
• The framework relies heavily on retrieved documents for factual grounding, meaning errors in retrieval could lead to verification failures.limitation 90%
• MARCH may not generalize across all LLM architectures.limitation 80%
• MARCH uses multi-agent reinforcement learning to enable agents to co-evolve and optimize factual adherence.method 85%
• MARCH orchestrates three specialized agents: a Solver, a Proposer, and a Checker.method 95%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card