Research map / Multi-agent systems and games / LLM agents
Beyond Correctness: Distance-Based Social Dynamics of Multi-Agent Debate
Seungwoong Ha, Melanie Mitchell
Read the paperCheck your idea against 137,217 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.
Studies microscopic answer-revision dynamics in multi-agent debate on ConceptARC grid-reasoning tasks, using quantitative distance between candidate solutions rather than binary correctness.
Problem
Multi-agent debate systems are typically evaluated with binary correctness or peer agreement, which obscures how individual agents revise answers during social interaction.
Approach
Uses ConceptARC, a 2D grid-reasoning benchmark whose candidate solutions admit quantitative distance measures, to analyze how LLM agents move toward or away from peer answers during debate.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| more wrong peers (scenario A, W3) shown to a correct target (ablation) | correct target in multi-agent debate · ConceptARC | retention of correct answer | Hurt |
| wrong peers repeat the same wrong answer instead of being diverse (ablation) | correct target · ConceptARC | retention of correct answer | Hurt |
| re-prompt without peers (self-reflection) for correct targets (ablation) | DeepSeek-chat and GPT-5 mini · ConceptARC | retention of correct answer: more than 70% retained | No clear effect |
| all wrong peers identical to the target's wrong answer (sameWrong) (ablation) | wrong target · ConceptARC | answer stability | Improved |
| wrong peers differ from the target's wrong answer (ablation) | wrong target · ConceptARC | answer retention: less than 10% retention | Hurt |
| larger initial pixel distance to ground truth | wrong target · ConceptARC | probability of changing answer | Improved |
| larger initial distance, correction rate | wrong target, DeepSeek-chat vs GPT-5 mini · ConceptARC | probability of correct revision | Mixed |
| social revision of wrong answers (contraction) | wrong target · ConceptARC | change in pixel distance after revision | Improved |
| wrong peers closer to the ground truth | correct target, scenario A · ConceptARC | probability of changing answer | Hurt |
Experimental setup
Datasets: ConceptARC
Benchmarks: ConceptARC
Metrics: distance between candidate solutions answer revision dynamics
Builds on: multi-agent debate
Closest papers by meaning
- Should we be going MAD? A Look at Multi-Agent Debate Strategies for LLMs (ICML 2024 · similarity 0.90)
- Group Perspective Matters: Regulating Debate Relationships Can Mitigate Blind Conformity in Multi-Agent Debate (NeurIPS 2026 · similarity 0.89)
- Multi-LLM Debate: Framework, Principals, and Interventions (NeurIPS 2024 · similarity 0.89)
- Debate or Vote: Which Yields Better Decisions in Multi-Agent Large Language Models? (NeurIPS 2025 · similarity 0.89)
- Efficient Multi-Agent Reasoning via Confidence-Guided Adaptive Debate (ICML 2026 · similarity 0.89)
- When Debate Helps: Proposal Supply and Verification-Aware Readout in Multi-Agent Reasoning (NeurIPS 2026 · similarity 0.88)
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.
