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

The Value of Variance: Mitigating Debate Collapse in Multi-Agent Systems via Uncertainty-Driven Policy Optimization

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 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.

Proposes a three-level (intra-agent, inter-agent, system) uncertainty metric for multi-agent debate and an uncertainty-driven policy optimization (UDPO) objective with agent-tailored asymmetric rewards to mitigate debate collapse, reporting accuracy gains on GSM8K, TruthfulQA and CommonsenseQA.

Problem

Multi-agent debate systems suffer from debate collapse, where final decisions are compromised by erroneous reasoning; existing work lacks granular metrics to detect it and training-based mitigations that improve intrinsic agent robustness.

Approach

Decomposes debate uncertainty into intra-agent flip rate, inter-agent pairwise conflict, and system-level entropy/disagreement/leave-one-out instability, then uses these as reward terms in a clipped, KL-anchored policy optimization with agent-specific coefficients calibrated from warm-up uncertainty profiles.

Key results

• UDPO reaches 92.3% accuracy on GSM8K with N=5 agents versus 68.4% for Standard MAD and 73.6% for MAPPO.
• Under attack with one compromised agent on Qwen, UDPO scores 77.2% on GSM8K versus 35.2% for Standard MAD (drop of 5.2pp vs 34.6pp).
• Failed vs successful reasoning shows mean U_sys of 0.67 vs 0.09 with p<0.001 and Cohen's d>0.8 on GSM8K.

Surprising findings

• Reported UDPO gains are large relative to prior MARL baselines, with headline accuracy rising to 92.3% on GSM8K.
• Uncertainty metrics show correlations approaching 1 with each other, though their distributions differ significantly.

Limitations

• Relies on agent-tailored coefficients calibrated by ad hoc warm-up uncertainty profiles and a scaling factor kappa, with limited justification.
• Evaluated only on three benchmarks with a flat debate structure and majority-vote aggregation.
• Reported gains (e.g., 92.3% vs 68.4%) are very large and the training details (data size, reference policy updates) are sparse.

Experimental setup

Datasets: GSM8K TruthfulQA CommonsenseQA

Benchmarks: GSM8K TruthfulQA CommonsenseQA

Metrics: accuracy intra-agent uncertainty inter-agent uncertainty system-level uncertainty Cohen's d Pearson correlation

Baselines: Standard MAD MAPPO RMAAC

Builds on: multi-agent debate PPO MAPPO RMAAC AutoTransform

Closest papers by meaning

• Epistemic Gain, Aleatoric Cost: Uncertainty Decomposition in Multi-Agent Debate for Math Reasoning (ICML 2026 · similarity 0.92)
• Should we be going MAD? A Look at Multi-Agent Debate Strategies for LLMs (ICML 2024 · similarity 0.90)
• Multi-Agent Debate for LLM Judges with Adaptive Stability Detection (NeurIPS 2025 · similarity 0.89)
• Debate or Vote: Which Yields Better Decisions in Multi-Agent Large Language Models? (NeurIPS 2025 · similarity 0.89)
• Key Decision-Makers in Multi-Agent Debates: Who Holds the Power? (AAAI 2026 · similarity 0.89)
• When Debate Helps: Proposal Supply and Verification-Aware Readout in Multi-Agent Reasoning (NeurIPS 2026 · similarity 0.89)