Research map / Multi-agent systems and games / LLM agents

MAS-ProVe: Understanding the Process Verification of Multi-Agent Systems

ICML 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Systematic empirical study of process verification (LLM-as-a-Judge, reward models, PRMs) for multi-agent systems across verification granularity, context management, and solvability, finding that process verification does not consistently help and is high-variance.

Problem

Whether process-level automatic verification actually improves LLM-based multi-agent systems, given high variance in MAS reasoning trajectories and the out-of-distribution nature of partial MAS trajectories for existing verifiers.

Approach

A plug-and-play middleware protocol (MAS-ProVe) that samples N=3 candidate continuations at agent-call or iteration granularity and uses a verifier to greedily select the best, letting any off-the-shelf MAS be tested with multiple verifier types and context-management strategies.

Key results

• Judge (GPT-5-mini) verification on Debate lifts AIME24 from 74.13 to 81.11 at agent level, but RM and PRM often degrade performance, e.g. Debate RM at iteration level reaches 81.11 while Agent-level PRM drops to 68.10.
• Generative judges beat reward-based verifiers in 24 of 36 configurations across agent and iteration levels.
• Context strategy study: Summarized context matches or beats raw history while using far fewer tokens; for MaAS raw history costs about 3x more tokens with lower accuracy.

Surprising findings

• Process verification with reward models or PRMs often degrades MAS performance relative to no verification, and gains are noisy across settings.
• Verification with minimal step-only context performs comparably to or better than full-context verification.

Limitations

• Process verification rarely recovers problems that the MAS generator cannot solve at all.
• Results are high-variance and depend on the MAS architecture, verifier, and context format.
• Experiments use a limited set of benchmarks with small test sets (e.g., 30 AIME questions) and a single primary backbone.

Experimental setup

Datasets: AIME24 AIME25 GAIA

Benchmarks: AIME24 AIME25 GAIA HumanEval GPQA

Metrics: accuracy pass rate standard deviation over three runs Pass@3 Pass@128

Baselines: Debate AFlow ADAS MaAS DyLAN MAS-Zero Skywork-Reward-V2-Llama-3.1-8B Qwen2.5-Math-PRM-7B GPT-4o-mini FARE-20B

Builds on: LLM-Debate AFlow ADAS MaAS MAS-Zero

Closest papers by meaning

• MAS-Orchestra: Understanding and Improving Multi-Agent Reasoning Through Holistic Orchestration and Controlled Benchmarks (ICML 2026 · similarity 0.90)
• DoVer: Intervention-Driven Auto Debugging for LLM Multi-Agent Systems (ICLR 2026 · similarity 0.89)
• Why Do Multi-Agent LLM Systems Fail? (NeurIPS 2025 · similarity 0.88)
• interwhen: A Generalizable Framework for Steering Reasoning Models with Test-time Verification (NeurIPS 2026 · similarity 0.88)
• Benefits and Limitations of Communication in Multi-Agent Reasoning (ICLR 2026 · similarity 0.88)
• The Illusion of Multi-Agent Advantage (NeurIPS 2026 · similarity 0.88)