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

2026-03-25 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

The Multi-Agent Reinforced Self-Check for Hallucination (MARCH) framework addresses the critical problem of hallucination in large language models (LLMs), particularly within Retrieval-Augmented Generation (RAG) systems. Existing LLM-as-a-judge methods for hallucination detection often suffer from confirmation bias, as the verifier can reproduce errors from the initial generation. MARCH mitigates this by employing deliberate information asymmetry across three specialized agents: a Solver, a Proposer, and a Checker. The Solver creates an initial RAG response, which the Proposer then breaks down into atomic, verifiable claims. The Checker validates these claims against retrieved evidence independently, without access to the Solver's original output, thereby breaking the self-confirmation bias cycle. This pipeline is trained using multi-agent reinforcement learning (MARL), allowing agents to co-evolve for optimized factual adherence. Experiments show MARCH significantly reduces hallucination rates, with an 8B-parameter LLM achieving competitive performance against powerful closed-source models.

Key takeaway

For AI Architects and Research Scientists developing RAG systems, MARCH offers a robust method to combat LLM hallucination. By adopting its multi-agent, information-asymmetric verification pipeline, you can significantly improve factual accuracy and reliability. Consider integrating this reinforcement learning approach to enable LLMs to self-improve their factual adherence, potentially making 8B-parameter models competitive with larger, closed-source alternatives.

Key insights

MARCH uses multi-agent reinforcement learning and information asymmetry to reduce LLM hallucination in RAG systems.

Principles

• Information asymmetry breaks confirmation bias.
• Decompose responses into atomic propositions.
• Co-evolution optimizes factual adherence.

Method

MARCH employs a Solver for initial RAG response, a Proposer to decompose it into atomic claims, and a Checker to validate claims against evidence in isolation, trained via MARL.

In practice

• Implement multi-agent systems for verification.
• Isolate verifiers from original outputs.
• Apply MARL for factual self-improvement.

Topics

• LLM Hallucination
• Retrieval-Augmented Generation
• Multi-Agent Reinforcement Learning
• Factual Alignment
• Information Asymmetry

Code references

• Qwen-Applications/MARCH

Best for: Research Scientist, AI Architect, AI Engineer, AI Researcher, AI Scientist, Machine Learning Engineer

Related on AIssential

• MODE-RAG: Manifold Outlier Diagnosis and Energy-based Retrieval-Augmented Generation Evaluation — MODE-RAG dynamically gates M-RAG interventions using Variational Free Energy and attention states to reduce cross-modal hallucinations and fabrications.
• Google researchers introduce 'faithful uncertainty,' allowing LLMs to offer best guesses instead of hallucinations — Faithful uncertainty aligns LLM linguistic expression with internal confidence to mitigate confident errors and improve agentic AI reliability.
• [D] Self-Promotion Thread — The AI/ML community actively develops diverse tools, from cost-efficient LLM interaction to advanced model evaluation and production monitoring.
• Extrinsic Hallucinations in LLMs — LLM hallucinations arise from pre-training data issues and fine-tuning new knowledge, necessitating robust detection and mitigation strategies.
• Hallucination as Trajectory Commitment: Causal Evidence for Asymmetric Attractor Dynamics in Transformer Generation — Hallucination in LLMs is an early, asymmetric attractor phenomenon, easy to trigger but hard to correct.
• The Reasoning Trap: How Enhancing LLM Reasoning Amplifies Tool Hallucination — Enhanced LLM reasoning, especially via RL, causally increases tool hallucination, revealing a fundamental reliability-capability trade-off.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.