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

Stochastic Self-Organization in Multi-Agent Systems

ICLR 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 SelfOrg, a judge-free multi-agent framework that builds a per-round DAG from response-embedding cosine similarity and a Shapley-approximating contribution score, and reports gains on math, science, and knowledge benchmarks with Qwen, LLaMA, Falcon, and Mistral backbones.

Problem

Multi-agent LLM systems rely on fixed topologies, pretrained graph generators, edge optimization, or external judges, which add overhead and fail to adapt to stochastic agent responses, especially with weak backends.

Approach

Agents answer independently; each response is embedded, contributions are approximated as cosine similarity to the average embedding (a Shapley approximation), and a per-round DAG routes messages from high-contribution to lower-contribution agents, with the top response initiating the next round.

Key results

• On Qwen-2.5-1.5B-Instruct, SelfOrg averages 45.05% across seven benchmarks versus 41.24% for single-agent and 33-37% for multi-agent baselines, with best AVG-R of 1.00.
• On LLaMA-3.3-70B-Instruct, SelfOrg reaches 70.19% average accuracy (AVG-R 1.25) versus 68.46% for CoT.
• On Qwen-2.5-72B-Instruct, SelfOrg reaches 68.95% average accuracy (AVG-R 1.38) versus 67.82% for single agent.

Surprising findings

• At 72B on AQUA-RAT, SelfOrg slightly decreases accuracy from 81.10 to 80.71 relative to single agent, suggesting diminishing returns for strong models.
• G-Designer performs worse than the single-agent baseline on the weak 1.5B backbone and was omitted for larger models.

Limitations

• Theoretical guarantees rely on an assumption of clustered correct responses and a loose approximation bound with unnormalized constants.
• Gains diminish for strong backbones (e.g., 72B on AQUA-RAT).
• Relies on embedding similarity, which may not capture semantic correctness when answers differ in surface form.

Experimental setup

Datasets: MATH GSM8K GSM-Hard AQUA-RAT AIME-2024 GPQA MMLU MMLU-Pro

Benchmarks: MATH GSM8K AQUA-RAT GSM-Hard MMLU MMLU-Pro AIME-2024 GPQA

Metrics: accuracy average rank (AVG-R) token consumption

Baselines: Single CoT DyLAN MacNet G-Designer AgentVerse AutoGen

Builds on: Shapley value chain-of-thought MASLab G-Designer DyLAN

Closest papers by meaning

• Emergent Coordination in Multi-Agent Language Models (ICLR 2026 · similarity 0.90)
• Graph-of-Agents: A Graph-based Framework for Multi-Agent LLM Collaboration (ICLR 2026 · similarity 0.89)
• MAS$^2$: Self-Generative, Self-Configuring, Self-Rectifying Multi-Agent Systems (ICLR 2026 · similarity 0.89)
• Systematic Failures in Collective Reasoning under Distributed Information in Multi-Agent LLMs (ICML 2026 · similarity 0.89)
• Benefits and Limitations of Communication in Multi-Agent Reasoning (ICLR 2026 · similarity 0.88)
• AgentNet: Decentralized Evolutionary Coordination for LLM-based Multi-Agent Systems (NeurIPS 2025 · similarity 0.88)