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

The key methodological bet here is that short, cheap behavioral games can substitute for expensive end-to-end task evaluations when assembling LLM teams — essentially treating cooperation as a measurable trait rather than an emergent accident of deployment.

This sits in direct conversation with CoopEval (covered April 16), which tested LLM agents in social dilemmas like prisoner's dilemma and found that recent models default to defection. Where CoopEval diagnosed the problem and tested game-theoretic repair mechanisms, this paper takes a different angle: rather than fixing uncooperative models, it asks whether cooperative tendency can be measured upfront to predict team composition outcomes. The two papers together sketch a nascent research program around LLM cooperation as a first-class engineering concern, not just a behavioral curiosity. That framing also connects loosely to the LLM judge reliability work from April 16, which showed that aggregate metrics can mask per-instance inconsistency — a warning worth keeping in mind when interpreting cooperative profile scores at the model level.

The predictive validity claim needs stress-testing on tasks outside the paper's specific AI-for-Science scope. If independent groups replicate the correlation between cooperative profiles and team performance on heterogeneous agentic benchmarks within the next six months, the screening approach becomes practically useful; if it only holds for narrow scientific workflows, it is a domain-specific finding, not a general team-assembly tool.