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

COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems

NeurIPS 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.

Introduces COOP^2, an evaluation framework that grounds LLM multi-agent cooperation in task progress via constraint-guarded cooperative tasks in MA-Crafter and CUBE, plus COOP^2-Repair, a heuristic constraint-failure predictor that opens targeted repair communication rounds.

Problem

Existing LLM multi-agent evaluations treat cooperation as an implicit ingredient of final task success, so it is unclear whether agents actually cooperate, how interactions affect task progress over time, and why failures occur.

Approach

A cognitive-primitive interface couples symbolic agent communication/planning with grounded environment actions, and cooperative tasks are defined by four verifiable constraint types (spatial, temporal, participation, dependency); COOP^2-Repair predicts constraint failures from group plans and triggers targeted repair channels before primitive execution.

Key results

• GPT-5.4 in 3-agent Individual reaches Score 809.0 on MA-Crafter versus 146.8 for GPT-5.4-mini and 30.4 for Llama-Scout.
• Centralized communication raised 6-agent GPT-5.4-mini Score from 86.8 (Individual) to 141.8, while 3-agent GPT-5.4 dropped from 809.0 to 266.8 under Centralized.
• Dependency violations for GPT-5.4-mini reach up to 93% in 3-agent Centralized, while GPT-5.4 reduces them to roughly 7-11% in matched settings; spatial violations stay at 36-43% even for GPT-5.4.

Surprising findings

• Communication overhead reduced the strongest backbone's score (GPT-5.4 3-agent drops ~67% under Centralized), indicating communication is not always cooperation.
• Scaling from 3 to 6 agents reduced Individual GPT-5.4 score by about 15%, as agents compete for the same resources without coordinating.

Limitations

• COOP^2-Repair uses lightweight heuristic estimators rather than a learned constraint predictor.
• Evaluation covers only two grid-world environments and three backbones with 5 runs per setting.
• Repair increases decision overhead and communication load.

Experimental setup

Datasets: MA-Crafter CUBE

Benchmarks: MA-Crafter CUBE

Metrics: Score Steps Score/Step Plans/Agent Messages Interruptions Total Decision Time Spatial violation Temporal violation Dependency violation

Baselines: Individual Centralized Chain

Builds on: MA-Crafter Crafter CUBE ReAct

Closest papers by meaning

• More Capable, Less Cooperative? When LLMs Fail at Zero-Cost Collaboration (ICML 2026 · similarity 0.93)
• CaPo: Cooperative Plan Optimization for Efficient Embodied Multi-Agent Cooperation (ICLR 2025 · similarity 0.90)
• Emergent Coordination in Multi-Agent Language Models (ICLR 2026 · similarity 0.90)
• Improving the Efficiency of Language Agent Teams with Adaptive Task Graphs (NeurIPS 2026 · similarity 0.89)
• Shapley-Coop: Credit Assignment for Emergent Cooperation in Self-Interested LLM Agents (NeurIPS 2025 · similarity 0.89)
• Evaluating Generalization Capabilities of LLM-Based Agents in Mixed-Motive Scenarios Using Concordia (NeurIPS 2025 · similarity 0.89)