MAVEN: Achieving Auditable Explicit Reasoning Through Multi-Agent Verification and Refinement Networks

MAVEN proposes a multi-agent verification framework based on a blackboard architecture. It transforms LLMs into auditable, deliberate reasoners through the Skeptic-Researcher-Judge adversarial loop and outperforms baselines like Gemini 3.1 Pro and ReConcile across multiple benchmarks. Its core goal is to address the auditability issue in LLM reasoning processes and achieve transparent, verifiable explicit reasoning.

Large language models perform well in complex reasoning tasks, but traditional Chain-of-Thought methods suffer from error cascading—early step errors propagate step by step, leading to deviated conclusions. High-risk scenarios (e.g., medical, legal, financial) require reasoning processes to be verifiable and auditable. However, existing monolithic architectures lack modular intermediate verification mechanisms; reasoning trajectories are implicit black boxes, making fine-grained auditing difficult and undermining user trust. Building self-correcting, process-transparent reasoning frameworks has become a key challenge.

MAVEN draws inspiration from the blackboard architecture, decomposing reasoning into a collaborative network of specialized agents. The core is the Skeptic-Researcher-Judge adversarial loop:

• Skeptic: Acts as a devil's advocate, identifying reasoning loopholes, assumption flaws, and logical leaps;
• Researcher: Constructs arguments, responds to doubts, and provides evidence support;
• Judge: Neutral arbitrator, evaluating reasoning credibility and deciding to proceed, backtrack, or terminate. Agents interact dynamically through a shared blackboard, enabling 'In-Step Epistemic Auditing'—auditing while reasoning, capturing errors in time.

MAVEN was evaluated across four benchmarks: OpenBookQA (scientific common sense), TruthfulQA (anti-misinformation), HALUEVAL (hallucination detection), and StrategyQA (multi-step strategic reasoning). Results show it outperforms strong baselines (Gemini 3.1 Pro, ReConcile) in fine-grained metrics, with advantages in explicit structured, modular, verifiable reasoning trajectories. Additionally, MAVEN is model-agnostic, transferable to various LLM architectures and brings significant performance improvements.

MAVEN brings three insights:

1. From Implicit to Explicit: Requiring models to show reasoning processes instead of just learning to reason correctly is crucial for high-risk applications;
2. Value of Multi-Agent Collaboration: Explicit role assignment allows each agent to focus on specific functions, leading to overall performance exceeding monolithic models;
3. Necessity of Real-Time Auditing: Post-hoc auditing cannot prevent error propagation; step-by-step auditing embeds quality control into the reasoning process.

MAVEN represents an important evolution in LLM reasoning architectures—from pursuing scale to auditability, from black-box reasoning to transparent deliberation. In high-risk decision-making scenarios, 'why this reasoning' is more important than 'what the result is'. Through the multi-agent adversarial loop and step-by-step auditing, MAVEN provides a feasible path for building trustworthy AI reasoning systems and will play an increasingly important role in AI deployment in key fields.