Language-model agents are increasingly acting on behalf of people in settings where individual decisions shape shared outcomes. A new arXiv preprint (2610.22882) examines whether five production models adapt their contributions when the level of risk changes in a collective-risk dilemma—a scenario where a group must contribute to a common goal to avoid a shared loss.
The title of the paper, "Cooperation Without Adaptation," summarizes the central finding: the agents did cooperate, but they did not adapt their contributions to changing risks. The abstract states the researchers examined whether the models adapt, and the title indicates they did not, though the full results are not available in the excerpt.
This pattern matters for real-world deployment. If LLM agents are to act on behalf of people in dynamic situations—such as managing shared resources or coordinating group responses to evolving threats—they may need mechanisms that respond to changing conditions, not just maintain steady cooperation. The source is a preprint, so the findings have not yet undergone peer review.