MAS-ProVe: Process Verification in MAS

• The paper demonstrates that process verification in LLM-based MAS yields inconsistent performance gains and high variance across diverse reasoning domains.
• It introduces a comprehensive experimental framework assessing verification paradigms, granularity levels, and context management strategies.
• Empirical results reveal the need for specialized, context-aware verifiers to mitigate noise and out-of-distribution challenges in multi-agent coordination.

MAS-ProVe is a systematic empirical study assessing the effectiveness of process verification in Multi-Agent Systems (MAS) composed of LLMs. The study introduces a comprehensive experimental framework to evaluate verification paradigms, verifiers, and context management strategies within various MAS frameworks across multiple reasoning domains. While process verification—where automatic verifiers evaluate and select among intermediate reasoning steps—has been proposed as a promising approach for improving MAS coordination, MAS-ProVe establishes that its efficacy is neither consistent nor universally beneficial, revealing substantial variance and several critical challenges (Venkataramani et al., 3 Feb 2026).

1. Formalization of Multi-Agent Systems and Process Verification

A Multi-Agent System (MAS) consists of a set $A = \{A_1, A_2, ..., A_n\}$ of LLM-based agents, each maintaining its own local context $C_i$ and producing partial outputs iteratively. Problem-solving proceeds by alternating agent calls (where each agent $A_i$ generates text conditioned on $C_i$ and possibly via external tools) and coordination steps (aggregating or routing outputs), culminating in a final answer. The system’s reasoning can be modeled by a trajectory:

$\tau(q) = [s_1, s_2, \ldots, s_T]$

where each $s_t$ is an agent output or a coordination result for user query $q$.

Process verification introduces an automated verifier $V:H\to\mathbb{R}$—with $H$ the set of all partial trajectories—that evaluates intermediate steps. For $N$ parallel candidate continuations $C_i$0 of a current state, process verification selects:

$C_i$1

and execution resumes from the selected $C_i$2. This mechanism can be executed at each reasoning step or iteration, facilitating potential error correction, path selection, or intra-trajectory optimization.

2. Verification Paradigms and Granularity

MAS-ProVe examines three primary verification paradigms:

• Generative Judge (LLM-as-a-Judge): $C_i$3 scores or ranks trajectories by prompting general-purpose or specialized LLMs, via $C_i$4.
• Reward Model (RM): $C_i$5 delivers a scalar reward for only terminal outputs.
• Process Reward Model (PRM): $C_i$6 evaluates every intermediate prefix of a reasoning trajectory.

Process verification can intervene at two granularities:

• Agent-Level Verification: After each agent’s output at time $C_i$7, among $C_i$8 candidates, $C_i$9 is selected and propagated.
• Iteration-Level Verification: After one full cycle through all agents, $A_i$0 parallel runs of the entire iteration yield $A_i$1 for trajectory selection.

No single paradigm or granularity universally dominates; empirical results show strong dependence on MAS architecture and problem domain (Venkataramani et al., 3 Feb 2026).

3. Evaluated Verifiers and Context Management

MAS-ProVe benchmarks five representative, off-the-shelf verifiers:

• General-purpose LLM judges: GPT-4o-Mini, GPT-5-Mini
• Task-specialized fine-tuned judge: FARE-20B
• Reward model: Skywork-Reward-V2-Llama-3.1-8B (RM)
• Process reward model: Qwen2.5-Math-PRM-7B (PRM)

For managing verifier input, four context strategies are examined:

1. Step-Only: Only the latest agent output.
2. Raw (Full History): Full sequence of prior and current step outputs.
3. Summary + Step: Summary of full history (via Summ(·), an auxiliary LLM-generated summary) along with the current output.
4. Refined Summary: Refined summary including both summarized history and current output.

A context-length-performance trade-off is observed. Step-Only approaches can match or outperform Raw history configurations, likely due to noise overload in long trajectories. Summarization regains performance when full context becomes excessively lengthy (Venkataramani et al., 3 Feb 2026).

4. Experimental Framework and Benchmarks

MAS-ProVe covers six MAS frameworks across two MAS paradigms:

MAS Category	MAS Frameworks
Fixed-Topology MAS	LLM-Debate, Agentic Workflow (AFlow), ADAS
Adaptive MAS	DyLAN, MaAS, MAS-Zero

Benchmarks include:

• AIME24 and AIME25: Middle-school level math problem sets (30 problems each)
• GAIA Information-Extraction: 103 real-world assistant tasks with tool usage

Primary metrics are:

• Success Rate (Accuracy): Fraction of test problems solved correctly
• Variance: Standard deviation across three random runs
• Performance Gain: Accuracy gain relative to no-verification baseline
• Evaluator Stability: Outcome variance across $A_i$2 parallel evaluation trials
• Intra-MAS Resurrection: Frequency with which baseline failures ($A_i$3 accuracy) are converted to at least $A_i$4 success under verification

5. Empirical Results and Comparative Insights

Process verification does not consistently yield performance gains, frequently displaying high variance—standard deviations up to 15 percentage points were observed. There is strong entanglement between MAS design and the effectiveness or instability of each verification paradigm or granularity. For example:

MAS	LLM-as-a-Judge	RM	PRM
Debate	81.11% ± 1.92	71.67% ± 1.44	68.10% ± 4.50
AFlow	76.39% ± 6.36	76.22% ± 6.59	74.17% ± 3.01

In 24 of 36 configurations, LLM-as-a-Judge outperforms reward-based verifiers. Task-specialized judges (FARE-20B) surpass general-purpose LLM judges by margins of 1–5 percentage points. The gap between judge-based and reward-based paradigms is significant; however, the difference between specialized and general-purpose judges is relatively modest, suggesting that verification may be intrinsically easier than generating correct solutions (Venkataramani et al., 3 Feb 2026).

A context-length-performance trade-off is highlighted in experiments under agent-level GPT-5-Mini judging:

	Step-Only	Raw	Summary	Summary+Step
DyLAN	86.67% ± 3.34	85.56% ± 1.93	82.22% ± 5.09	86.67% ± 3.34
MaAS	81.94% ± 2.41	80.55% ± 8.67	84.72% ± 4.82	84.72% ± 6.37

Step-Only and summarized contexts are often competitive with full-history, mitigating issues of context noise.

6. Limitations, Open Challenges, and Future Directions

MAS-ProVe finds that reliable process verification remains unresolved. Verification of partial MAS trajectories is an out-of-distribution task for current verifier models; context noise, missing dependencies, and agent output heterogeneity impede scoring reliability. High variance and non-monotonicity are persistent across settings.

Task-specialized judges outperform general LLMs by a modest edge, but the magnitude of this improvement is smaller than the solving gap between single-agent LLMs and verification-augmented MAS, indicating that verification is not the primary bottleneck under current architectures.

Proposed future research avenues include:

• Joint training of verifiers with MAS on actual agent trajectories to address out-of-distribution errors.
• Learning context-selection policies to optimize which trajectory features are provided to the verifier.
• Co-optimization of MAS and verifier architecture, potentially via policy-gradient techniques using verifier feedback signals.
• Development of multi-objective verifiers, extending beyond correctness to include consistency and factuality (Venkataramani et al., 3 Feb 2026).

A plausible implication is that simply transplanting verification strategies from single-agent or final-output settings is inadequate for current-generation LLM-based MAS, and that deeper integration, context selection, and verifier specialization are required for robust gains.