MAVEN: Multi-Agent Framework for LLM Reasoning with Epistemic Auditing

MAVEN (Multi-Agent Verification-Elaboration Network with In-Step Epistemic Auditing) is a framework inspired by blackboard architecture that enables large language models (LLMs) to engage in intentional reasoning by distinctly separating roles. It implements an adversarial loop involving a Skeptic, Researcher, and Judge, mimicking expert discussions by dissociating logical justification from factual accuracy. Tests conducted on OpenBookQA, TruthfulQA, HALUEVAL, and StrategyQA benchmarks indicate that MAVEN significantly improves reasoning quality across four detailed metrics, consistently exceeding the performance of latent reasoning approaches. This framework tackles the issue of missing intermediate verification in unified reasoning chains, which can lead to unchecked early mistakes, undermining epistemic trust in critical applications. MAVEN aims to improve model interpretability and detailed auditing.