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

CoMAS: Co-Evolving Multi-Agent Systems via Interaction Rewards

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 1 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 CoMAS, a framework where LLM agents co-evolve via RL using intrinsic rewards from their own solve/evaluate/score discussions (LLM-as-judge), evaluated on Qwen2.5-3B-Instruct agents across math, coding, science benchmarks.

Problem

Self-evolution of LLM-based agents typically relies on external verifiers/reward models or single-model intrinsic rewards, diverging from how humans improve through collective discussion. The paper asks whether agents can self-evolve purely from inter-agent interaction without external rewards.

Approach

Agents alternate solution proposal, critical evaluation, and scoring in a decentralized discussion; a scoring agent (LLM-as-judge) rates each solution-evaluation pair, and the score yields complementary zero-sum rewards for solver and evaluator, which train each agent with REINFORCE++.

Key results

• CoMAS on Qwen2.5-3B under Vanilla setup reaches 85.40 on GSM8K and 70.73 on HumanEval, best among methods on those benchmarks.
• Under AutoGen setup, CoMAS improves GSM8K from 52.60 to 72.40 (+19.80), while TTRL drops to 41.00.
• Removing scoring leads to reward hacking with rewards rising toward 1.0; removing evaluation causes rewards to decline.

Surprising findings

• Removing the scoring step causes rewards to converge to the maximum, indicating reward hacking by unanimous support of solutions.
• TTRL collapses under the AutoGen setup, dropping GSM8K by 11.60 points below untrained agents.

Limitations

• Evaluated only on 3B and 7B Qwen-family and one Llama model with a single training dataset of 2000 samples.
• Gains over untrained agents are sometimes modest (e.g., under 1 point on several benchmarks), and some baselines are close.
• Reward precision and recall remain moderate (precision about 50%, recall about 29-38%) per the reward-effectiveness analysis.

Experimental setup

Datasets: MATH KodCode WebInstruct-verified GSM8K MATH-500 HumanEval MBPP SciBench GPQA MMLU

Benchmarks: GSM8K MATH-500 HumanEval MBPP SciBench GPQA MMLU

Metrics: accuracy

Baselines: Untrained SRLM MAPoRL TTRL

Builds on: REINFORCE++ MARTI MAPoRL TTRL Self-rewarding language models

Closest papers by meaning

• MemEvolve: Meta-Evolution of Agent Memory Systems (ICML 2026 · similarity 0.89)
• MAGE: Multi-Agent Self-Evolution with Co-Evolutionary Knowledge Graphs (NeurIPS 2026 · similarity 0.89)
• Self-evolving LLM agents with in-distribution Optimization (ICML 2026 · similarity 0.89)
• Emergent Coordination in Multi-Agent Language Models (ICLR 2026 · similarity 0.89)
• Adaptive Collaboration with Humans: Metacognitive Policy Optimization for Multi-Agent LLMs with Continual Learning (ICLR 2026 · similarity 0.88)
• COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.88)