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

Systematic Failures in Collective Reasoning under Distributed Information in Multi-Agent LLMs

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

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.

Introduces HiddenBench, a 65-task Hidden Profile benchmark, and evaluates 15 frontier LLMs to show multi-agent groups fail to surface unshared information, with a lightweight structured protocol partially fixing it.

Problem

Multi-agent LLM systems are expected to pool distributed information for better decisions, but no controlled benchmark isolates collective reasoning under distributed information from individual reasoning ability.

Approach

Adapts the social-psychology Hidden Profile paradigm into a verifiable benchmark (with an automated GPT-4.1 generation and validation pipeline) where each task is solvable with full information but not from shared or individual private information, then diagnoses failures via ablations and a structured Exchange-then-Decide protocol.

Key results

• Multi-agent LLMs reach 30.1% accuracy under Hidden Profile versus 80.7% for single agents with full information (as stated in abstract).
• Gemini-2.5-Pro achieves 0.671 post-discussion Hidden Profile accuracy, the highest among 15 models, while GPT-5-Nano improves only -0.004.
• Structured Exchange-then-Decide protocol raises GPT-4.1 from 0.037 to 0.800 on 18 tasks; Reveal-All raises GPT-4.1 to 0.926.

Surprising findings

• Adding more agents worsens collective improvement: communication gain drops from +0.348 at N=3 to +0.006 at N=7.
• Simple dissent hurts Gemini-2.5-Flash, whose accuracy falls from 0.889 with no adversaries to 0.306 with one adversarial agent.

Limitations

• Tasks are validated only with GPT-4.1 agents in the generation pipeline, which may bias task selection toward that model.
• Human study is small (24 groups of four) and the structured protocol is only evaluated on 18 tasks with three models.

Experimental setup

Datasets: HiddenBench (65 Hidden Profile tasks: 3 manual, 5 adapted from prior literature, 57 generated by GPT-4.1)

Benchmarks: HiddenBench

Metrics: post-discussion accuracy (average rule) majority accuracy pre-discussion accuracy improvement over pre-discussion

Baselines: Full Profile single-agent accuracy Reveal-All Secretary Share All Information prompting Gemini-2.5-Pro

Builds on: Hidden Profile paradigm zero-shot chain-of-thought

Closest papers by meaning

• Emergent Coordination in Multi-Agent Language Models (ICLR 2026 · similarity 0.91)
• Autonomous Agents for Collaborative Task under Information Asymmetry (NeurIPS 2024 · similarity 0.90)
• Benefits and Limitations of Communication in Multi-Agent Reasoning (ICLR 2026 · similarity 0.90)
• AgentCDM: Enhancing Multi-Agent Collaborative Decision-Making via ACH-Inspired Structured Reasoning (AAAI 2026 · similarity 0.90)
• More Capable, Less Cooperative? When LLMs Fail at Zero-Cost Collaboration (ICML 2026 · similarity 0.90)
• Unlocking the Power of Multi-Agent LLM for Reasoning: From Lazy Agents to Deliberation (ICLR 2026 · similarity 0.90)