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

Group Perspective Matters: Regulating Debate Relationships Can Mitigate Blind Conformity in Multi-Agent Debate

NeurIPS 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.

Proposes DEAR, which regulates debate relationships in multi-agent debate via group evidence (consensus/divergence), Subjective Logic decoupling, a Selection RL-Agent for peer selection, and a Behavior RL-Agent adjusting temperature/top_p, trained jointly with HAPPO to mitigate blind conformity.

Problem

Blind conformity in multi-agent debate (LLM minorities abandon correct reasoning to follow the majority), where existing confidence- or perplexity-based individual filtering is unreliable and can worsen conformity.

Approach

Shift from individual evaluation to group-level regulation of debate relationships: quantify pairwise consensus/divergence as group evidence, decouple it into belief and uncertainty via Subjective Logic, learn which peers to reference (Selection RL-Agent) and how to sample (Behavior RL-Agent), and fuse peer beliefs with Dempster-Shafer combination, trained jointly with HAPPO.

Key results

• On Qwen3-8B, DEAR reaches 72.3% average math accuracy vs 66.6% for CoT-SC and 65.3% for MAD, with 10.9K average tokens vs 33.5K for MAD.
• On GPT-4o-mini, DEAR averages 75.0% on QA tasks vs 70.3% for CoT-SC, with 8.4K tokens vs 19.5K for MAD.
• On Claude-4.5-Sonnet, DEAR averages 58.4% across AIME24, IMO AnswerBench, MedXpertQA and GPQA Diamond vs 49.3% for MAD-M²(S), with 26.4K vs 38.0K tokens.

Surprising findings

• Individual evaluation methods such as CortexDebate and MAD-M²(S) sometimes underperform plain MAD, which the authors attribute to unreliable confidence or perplexity signals.

Limitations

• Evaluation on closed-source models relies on API calls and sampled subsets (e.g., 100-400 questions per dataset), limiting statistical power.
• Scope is limited to math and QA tasks; generative, coding, and multimodal settings are left to future work.

Experimental setup

Datasets: GSM8K ARC-C

Benchmarks: GSM8K AIME24 GSM-Hard MATH-500 ARC-C MMLU-Pro (Health) TruthfulQA GPQA Diamond IMOAnswerBench MedXpertQA (Text) MMLU

Metrics: accuracy average tokens per task

Baselines: CoT CoT-SC Majority Voting MAD DMAD CortexDebate MAD-M² ECON RUMAD S-MAD S²-MAD

Builds on: HAPPO Free-MAD ECON MAD-M² Subjective Logic

Closest papers by meaning

• Should we be going MAD? A Look at Multi-Agent Debate Strategies for LLMs (ICML 2024 · similarity 0.90)
• Multi-LLM Debate: Framework, Principals, and Interventions (NeurIPS 2024 · similarity 0.90)
• When Debate Helps: Proposal Supply and Verification-Aware Readout in Multi-Agent Reasoning (NeurIPS 2026 · similarity 0.90)
• Efficient Multi-Agent Reasoning via Confidence-Guided Adaptive Debate (ICML 2026 · similarity 0.89)
• Debate or Vote: Which Yields Better Decisions in Multi-Agent Large Language Models? (NeurIPS 2025 · similarity 0.89)
• Do as We Do, Not as You Think: the Conformity of Large Language Models (ICLR 2025 · similarity 0.89)