Study: Similarity signals can induce cooperation among LLM agents

conitzer · x · 2026-08-18

A study honored at the ICML AI4GOOD Workshop investigates how similarity signals affect cooperation between LLM agents. Findings show that models vary drastically in navigating these signals, with modern models showing consistent behavior across different scenarios. Surprisingly, the dataset used for similarity signals had minimal impact, and LLMs systematically self-identify as highly similar to others. A behavioral game-theoretic model was also developed to explain the reasoning rationale.

Original post →

More from AGI Musings

AGI Musings channel →