Skip to content

Qwen-Applications/MARCH

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 
 
 

Repository files navigation

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

Qwen-Applications arXiv Github License

Qwen Large Model Application Team, Alibaba

MARCH (Multi-Agent Reinforced Check for Hallucination) is a collaborative framework that enforces factual alignment in RAG systems by leveraging information asymmetry. By decoupling response generation, claim decomposition, and fact verification through specialized agents (Solver, Proposer, Checker), MARCH breaks the cycle of confirmation bias inherent in previous LLM verifiers.

  • Fact-Grounded: Uses Multi-Agent Reinforcement Learning (MARL) to ensure high-fidelity grounding.

  • Blind Verification: The Checker validates claims in isolation—no access to the Solver's internal logic.

  • Agentic Co-evolution: Agents learn to self-correct through collaborative multi-agent training.

MARCH framework
Overview: Proposer decomposes Solver‘s response into claim-level verifiable QA pairs. Checker performs blind validation against retrieved documents to recheck factuality.

More implementation is coming soon!

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages