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

TL;DR

MAS-ProVe systematically evaluates process-level verification in multi-agent systems built on LLMs across verification types, granularities, and context strategies. The findings show high variance and limited universal gains, with generative LLM judges typically outperforming reward-based verifiers, and with performance depending on MAS architecture and context design. The work highlights solvability ceilings and stability enhancements rather than broad recovery of fundamentally unsolvable tasks. It also provides a modular, plug-and-play framework to enable reproducible evaluation and future research in robust MAS verification.

Abstract

Multi-Agent Systems (MAS) built on Large Language Models (LLMs) often exhibit high variance in their reasoning trajectories. Process verification, which evaluates intermediate steps in trajectories, has shown promise in general reasoning settings, and has been suggested as a potential tool for guiding coordination of MAS; however, its actual effectiveness in MAS remains unclear. To fill this gap, we present MAS-ProVe, a systematic empirical study of process verification for multi-agent systems (MAS). Our study spans three verification paradigms (LLM-as-a-Judge, reward models, and process reward models), evaluated across two levels of verification granularity (agent-level and iteration-level). We further examine five representative verifiers and four context management strategies, and conduct experiments over six diverse MAS frameworks on multiple reasoning benchmarks. We find that process-level verification does not consistently improve performance and frequently exhibits high variance, highlighting the difficulty of reliably evaluating partial multi-agent trajectories. Among the methods studied, LLM-as-a-Judge generally outperforms reward-based approaches, with trained judges surpassing general-purpose LLMs. We further observe a small performance gap between LLMs acting as judges and as single agents, and identify a context-length-performance trade-off in verification. Overall, our results suggest that effective and robust process verification for MAS remains an open challenge, requiring further advances beyond current paradigms. Code is available at https://github.com/Wang-ML-Lab/MAS-ProVe.

Figures (12)

• Figure 1: Overview of Multi-Agent System Process Verification Framework (MAS-ProVe). The execution of any multi-agent system (MAS) can be abstracted into a unified workflow, where stages are executed sequentially and atomic LLM calls within each stage may run in parallel. MAS-ProVe systematically evaluates the integration of MAS with process verification by performing parallel search at two complementary granularities: agent-level and iteration-level, while remaining agnostic to the choice of process verifier. Prior to verification, customizable context-management strategies can be applied to partial multi-agent trajectories, enabling flexible control over the information passed to the verifier and the selection of the best candidate to return to the workflow.
• Figure 2: Accuracy vs. Token Cost (Log Scale) across MAS architectures. Strategies closer to the top-left corner are to be preferred. Summarized contexts consistently outperform Raw History while consuming significantly fewer tokens.
• Figure 3: AIME 2024 performance across MAS architectures using different process verification methods: Judge (GPT-5-mini), RM (Skywork-Reward-V2-Llama-3.1-8B), PRM (Qwen2.5-Math-PRM-7B), and Best Configuration. Error bars show standard deviation over three runs. Stars indicate Pass@3 performance.
• Figure 4: Average Pass rate on Mathematical Reasoning tasks by difficulty stratum. Baseline (red), Process Eval (green), and Best (blue) configurations compared across Hard, Medium, and Easy questions for six MAS architectures.
• Figure 5: Stability distribution aggregated across all MAS architectures on mathematical reasoning tasks. Stacked bars show trial outcome proportions by difficulty: Stable Failure (red, 0/3), High Variance (yellow, 1-2/3), and Stable Success (green, 3/3) for Baseline, Judge, and Best configurations.