Research map / Multi-agent systems and games / LLM agents
Key Decision-Makers in Multi-Agent Debates: Who Holds the Power?
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.
Shows that role allocation order matters in multi-agent debate (placing correct-viewpoint agents last boosts accuracy) and proposes MADC, a path-consistency-based allocation that approximates this without ground truth.
Problem
Multi-agent debate (MAD) performance is sensitive to how roles/viewpoints are ordered across agent positions, a factor that prior MAD frameworks leave fixed and unexamined; the truth is unknown in practice, so the ideal allocation cannot be directly used.
Approach
Identifies role allocation as a scaling dimension for MAD: later-speaking agents exert more influence, so placing correct viewpoints last ("Truth Last") improves accuracy. MADC estimates the truth without labels by scoring each agent's path consistency with others and ordering the highest-consistency agent last, requiring no prompt or context changes.
Key results
- Truth Last allocation improves MAD accuracy by up to 22.8 points (Qwen2.5-3B-Instruct, Logical Deduction and Geometric Shapes).
- MADC improves over MAD on Qwen2.5-3B logic by +8.8 points and DeepSeek-R1-Distill-Qwen-7B Geometric Shapes by +9.6 points.
- With 20 agents, roughly half the agent count relative to the 40-agent base can flip a majority outcome in the constructed setting.
Surprising findings
- Alternating allocation of correct and incorrect agents yields unstable outcomes that do not converge with more agents.
- Truth First yields lower entropy (more consensus) yet lower accuracy than other strategies.
Limitations
- Truth Last is an oracle that requires known ground truth; MADC only partially recovers its gains (often a few points).
- Gains of MADC are modest on several models (e.g., GPT-4o-mini +0.0 to +2.0 on some tasks, MATH500 often unchanged).
- Results are reported on a small set of tasks and models with limited repetitions in some settings.
Experimental setup
Datasets: MATH500 BIG-Bench Hard Logical Deduction BIG-Bench Hard Geometric Shapes
Benchmarks: MATH500 BBH Logical Deduction BBH Geometric Shapes
Metrics: accuracy log-likelihood entropy
Baselines: Single CoT Self-Consistency MAD MAD+Truth
Builds on: Multi-Agent Debate (Du et al.) Chain-of-Thought Self-Consistency
Closest papers by meaning
- Should we be going MAD? A Look at Multi-Agent Debate Strategies for LLMs (ICML 2024 · similarity 0.91)
- Debate or Vote: Which Yields Better Decisions in Multi-Agent Large Language Models? (NeurIPS 2025 · similarity 0.90)
- Multi-Agent Debate with Memory Masking (ICLR 2026 · similarity 0.90)
- Efficient Multi-Agent Reasoning via Confidence-Guided Adaptive Debate (ICML 2026 · similarity 0.90)
- When Debate Helps: Proposal Supply and Verification-Aware Readout in Multi-Agent Reasoning (NeurIPS 2026 · similarity 0.89)
- The Value of Variance: Mitigating Debate Collapse in Multi-Agent Systems via Uncertainty-Driven Policy Optimization (ICML 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.