Similar LLMs cooperate better but lose creativity, ICML study finds

A new paper from researchers at UC Berkeley, Google DeepMind, Stanford, and the University of Chicago, presented at ICML 2026, investigates how representational similarity between large language models affects their behavior in multi-agent interactions. The team paired 276 different models (across sizes, architectures, and training runs) and had them play eight games designed to test both cooperation (e.g., coordination tasks, resource sharing) and creativity (e.g., story generation, novel solution tasks). The core finding: pairs with more similar internal representations consistently achieved higher cooperation rates—but at a significant cost to novelty and creative output. The study controlled for factors like performance disparity and model size, confirming that the similarity itself, not other variables, drives this trade-off.

Strikingly, the effect was most pronounced in the early layers of the models, indicating that shared lexical and semantic grounding is the primary mechanism. This echoes patterns seen in human social networks, where similarity fosters cooperation but dissimilarity sparks innovation. The implications are concrete for system designers: if you want a multi-agent team that works together reliably, choose similar models; if you want disruptive creativity, mix in diverse ones. The research provides a measurable, tunable knob—representational similarity—for steering multi-agent behavior, which could impact everything from automated scientific discovery teams to cooperative AI agents in robotics and game design.