Cooperative Profiles Predict Multi-Agent LLM Team Performance in AI for Science Workflows

Multi-agent systems built from teams of large language models (LLMs) are increasingly deployed for collaborative scientific reasoning and problemsolving, requiring agents to coordinate under shared constraints where cooperative behavior matters. The core finding is that game-derived cooperative profiles robustly predict downstream performance in AI-for-Science tasks, indicating that models that effectively coordinate games and invest in multiplicative team production produce better scientific reports across accuracy, quality, and completion.

What the paper investigates

The research benchmarks 35 open-weight LLMs across six behavioral economics games to reveal systematic differences in cooperative behavior related to coordination, resource sharing, and team production. The investigation tests whether these game-derived cooperative profiles predict downstream performance in realistic multi-agent workflows where teams of LLM agents collaboratively analyze data and produce scientific reports under shared budget constraints.

The Behavioral Economics Framework

The study uses six distinct behavioral economics games to isolate different cooperation mechanisms:

1. Weakest-Link (Minimum-Effort Coordination): This game captures settings where team output depends on the least capable member, where success means sustaining high-effort coordination across both groups simultaneously.

2. Common-Pool Resource (CPR): This models the tragedy of the commons, where extraction yields individual gain but depletes a shared resource pool.

3. CPR with Sanctioning: This variant tests how cooperative behavior differs in the face of explicit institutional constraints, hypothesizing that LLMs that productively respond to such constraints are more likely to collaborate in real settings.

4. Collective Risk (Threshold Public Goods): This game models a collective action problem where groups must contribute enough to avert catastrophic loss before a risk event occurs.

5. O-Ring Team Production: This captures settings where withdrawing from the shared pool to build quality depletes resources for other teams, requiring balancing personal investment against the shared resource cost.

6. Public Goods Game: This probes cooperation across different scopes—individual keeping, group pooling (multiplied by 2.0), and global pooling (multiplied by 1.5)—measuring allocation to group/global pools rather than hoarding.

Predictive Modeling and Findings

The authors employ a two-level Bayesian hierarchical measurement-error model to predict Pareto proximity—the normalized distance from the Pareto optimum (0 = Pareto, 1 = Nash)—from model characteristics, prompting strategies, and game structure. Key predictors include:

(Model Characteristics):

Model size is the dominant predictor of collaborativeness: each unit increase in log10(size) reduces Pareto proximity by 0.051.

Prompting Strategies:

The analysis found that Theory of Mind has a credible effect, as ToM-prompted agents play closer to the Pareto optimum, making it the strongest strategy effect after model size. In contrast, chain-of-thought prompting showed no effect on Pareto proximity.

Game Structure:

Weakest-Link effort is positively associated with all three outcomes,

O-Ring withdrawal is negatively associated with all three metrics.

AI for Science Performance Prediction

The study tested whether game metrics predict downstream performance in a multi-agent AI-for-Science benchmark involving data analysis and report generation. The results showed that:

1. Weakest-Link effort is positively associated with all three outcomes, meaning models that cooperate more produce more accurate, higher-quality scientific reports and complete more stages.

2. O-Ring withdrawal is negatively associated with all three metrics, as models who withdraw more produce worse scientific output across all dimensions.

3. CPR extraction shows no credible effect on accuracy but is positively associated with quality and completion, suggesting a throughput advantage where resource-hungry models generate more output at each pipeline stage.

4. The contrast between CPR and CPR+Sanction revealed that models’ responsiveness to institutional constraints—i.e., the presence of sanctioning abilities—produce worse output.

Efficiency and Practical Utility

A significant practical contribution is the efficiency of this diagnostic method: "behavioral games serve as an efficient proxy for evaluating agentic collaborativeness, reducing cost by orders of magnitude (upto 44x faster and 27x fewer tokens) enabling rapid screening of LLMs for deployment in multi-agent systems. Furthermore, the researchers found that ≈20 simulations are sufficient to obtain reliable estimates of cooperative behavior metrics with low error, contrasting sharply with the 67K–481K tokens" required for full AI-for-Science experiments.

Conclusion

The work establishes a new framework for diagnosing and predicting cooperative behavior in LLM agent teams using behavioral economic games, demonstrating that models exhibiting high effort in coordination games produce superior scientific outcomes, offering a fast and inexpensive diagnostic tool for screening agentic systems.

Here are the specific improvements for AI systems based on this research, categorized by capability:

)System-Level Improvement: Predictive Cooperative Screening (Cost Reduction & Efficiency)

The core improvement is moving from expensive, full-scale deployment to a rapid, low-cost diagnostic phase.

1. A new pre-screening pipeline should be developed using the behavioral game framework. Before deploying any LLM team for a costly AI-for-Science task, run the LLM through the six behavioral economics games (Weakest-Link, CPR with Sanctioning, O-Ring Team Production, etc.).

2. The system should calculate a Cooperative Profile Score based on these game metrics.

3. Only models with a high predicted Pareto proximity (indicating effective cooperation) in the most relevant games—specifically those showing strong coordination in the Weakest-Link and resource management in CPR/O-Ring—should proceed to full deployment. This acts as a fast and inexpensive diagnostic for screening cooperative fitness.

)Agent-Level Improvement: Strategic Behavioral Tuning (In-Game Adaptation)

Instead of relying solely on general prompting, agents should be dynamically tuned based on their observed cooperative profile.

1. Implement a dynamic strategy selection layer where the LLM agent assesses its current cooperative disposition (inferred from its performance in behavioral games).

2. If the profile suggests high coordination ability (e.g., strong performance in Weakest-Link), the system should encourage it to adopt multiplicative team production strategies over greedy ones, specifically targeting outcomes that maximize scientific report quality and accuracy.

3. If the profile shows weakness in resource management (high CPR extraction), agents should be explicitly prompted or constrained to prioritize restraint and peer sanctioning mechanisms over unilateral extraction during shared budget phases, mimicking successful human group dynamics.

)Workflow-Level Improvement: Robust Multi-Agent Task Execution (AI-for-Science)

The current system's failure mode involves models that are technically capable but fail at complex, sequential collaboration under constraint. The improvement focuses on making the workflow resilient to these failures.

1. Implement a hierarchical task decomposition where agents explicitly model the shared budget constraints and inter-team dependencies as hard constraints in their planning phase (using LangGraph structures).

2. Develop specialized Error Analyst and Eval Lead roles (as suggested in Section C) that are triggered automatically when budget warnings are issued or when upstream artifacts fail to meet quality thresholds, allowing the team to self-correct by requesting targeted re-dos from specific teams rather than restarting the entire pipeline.

3. For tasks requiring high accuracy (e.g., Biomarker Discovery), the system should prioritize models with high IFEval scores, as research shows this capability is an independent predictor of success, ensuring that even if cooperative disposition is average, a high-capability model can still produce accurate scientific findings.

)Model Training/Fine-Tuning Improvement: Targeted Behavioral Alignment

Use the behavioral game results to guide post-training alignment efforts.

1. If a target LLM shows low scores in critical coordination games (like O-Ring Team Production), fine-tune it using Reinforcement Learning from Human Feedback (RLHF) specifically on datasets that reward multiplicative team production rather than individual maximization, aiming to shift its baseline cooperative bias toward collective success.

2. Use the findings regarding prompting strategies: since Theory of Mind (ToM) has a credible effect on game behavior but no effect on complex task completion, the system should be designed to use ToM-enabled prompts for initial coordination tasks (like data analysis planning) while relying on highly structured, deterministic prompts for execution phases where stability is paramount.