LLM agents struggle with cooperation in new study

New research from Kazuya Horibe, Kenji Itao, and Wataru Toyokawa reveals critical limitations in how large language model (LLM) agents cooperate. Published on arXiv (2608.04507), the study examines whether LLM agents can develop evolutionarily stable cooperation strategies when facing free-riders (agents that exploit generosity without reciprocating). The team tested four major LLM backends—including proprietary and open-source models—and found their robustness varied by over an order of magnitude.

The study introduced a 'donation game' where agents observed behavioral traces and donated on a continuous scale. While models showed varying sensitivity to opponent endowment (ability to distinguish cooperative vs. uncooperative agents), none successfully adopted the more robust 'Leading-Eight L1 norm' strategy. Instead, all defaulted to simpler 'Image Scoring' mechanisms, which proved vulnerable to defector exploitation. The key vulnerability? Defector exclusion stringency—how strictly models punish uncooperative agents—directly predicted resistance to free-rider invasion.