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

Multi-Agent Teams Hold Experts Back

ICML 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 2 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.

Shows that self-organizing LLM teams fail to achieve strong synergy (matching their best member), even when told who the expert is, attributing this to integrative compromise that worsens with team size.

Problem

Whether self-organizing multi-agent LLM teams with unconstrained deliberation can match or exceed their strongest individual member (strong synergy), and what causes shortfalls: identifying or leveraging expertise.

Approach

Adapts human teamwork tasks and frontier ML benchmarks to measure a relative synergy gap versus the best member, decomposes it into identification vs. leveraging gaps via Reveal Expert ablations, and uses conversational coding to show integrative compromise drives expertise dilution.

Key results

• On ML benchmarks, teams trail the At Least One Correct bound with relative synergy gaps from 6.3% (MMLU Pro) to 41.1% (HLE Text-Only), even under Reveal Expert prompts.
• Integrative compromise correlates positively with the synergy gap (NASA r=0.55, p<0.001; SBP r=0.69, p<0.001), while epistemic deference correlates negatively.
• A team of four full-information experts matches a single expert (e.g., NASA 16.22 vs 15.03 error), while concentrating expertise in one agent yields 25.35, isolating expertise asymmetry as the cause.

Surprising findings

• Teams fail to match the expert even when explicitly told who the expert is, using prompts optimized with GEPA.
• Consensus-seeking that causes expertise dilution also provides robustness to adversarial team members.

Limitations

• Attribution of consensus-seeking to alignment training is correlational; no base-vs-aligned model comparison.
• ML benchmark evaluation uses 100-problem subsamples per benchmark due to inference cost.
• Only a small subset of tasks and collaboration scenarios is evaluated.

Experimental setup

Datasets: MMLU Pro GPQA Diamond SimpleQA Humanity's Last Exam (text-only) MATH-500 NASA Moon Survival Lost at Sea Student Body President

Benchmarks: MMLU Pro GPQA Diamond SimpleQA HLE Text-Only MATH-500 NASA Moon Survival Lost at Sea Student Body President

Metrics: accuracy L1 ranking error relative synergy gap At Least One Correct upper bound Pearson correlation

Baselines: Debate Opt-Out CoT + majority vote Best Individual

Builds on: multi-agent debate GEPA Mixture-of-Agents GPTSwarm

Closest papers by meaning

• More Capable, Less Cooperative? When LLMs Fail at Zero-Cost Collaboration (ICML 2026 · similarity 0.89)
• TeamTR: Trust-Region Fine-Tuning for Multi-Agent LLM Coordination (ICML 2026 · similarity 0.89)
• AgentNet: Decentralized Evolutionary Coordination for LLM-based Multi-Agent Systems (NeurIPS 2025 · similarity 0.88)
• Emergent Coordination in Multi-Agent Language Models (ICLR 2026 · similarity 0.88)
• AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors (ICLR 2024 · similarity 0.87)
• Model Swarms: Collaborative Search to Adapt LLM Experts via Swarm Intelligence (ICML 2025 · similarity 0.87)