Sustaining Collective Welfare Under Disruption: A Pilot Study of Cooperative Resilience in Human and LLM Multi-Agent Systems

Version note. This entry corresponds to the revised camera-ready manuscript accepted at FAIEMA 2026. It is distinct from the earlier arXiv preprint, “Evaluating Cooperative Resilience in Multiagent Systems: A Comparison Between Humans and LLMs,” which remains available as the original preprint. The accepted manuscript is a substantially revised and polished version of the work and will be linked here once it becomes available in the official Springer proceedings.

Abstract. The ability to sustain collective welfare under disruption is key for deploying agents in real-world multi-agent environments. However, most evaluations focus on stable scenarios; as a result, agent behavior under adverse conditions remains poorly understood. In this work, we present a pilot comparative study of cooperative resilience—the ability of a system to act before, during, and after a disruption to preserve collective welfare—between human groups and LLM-based multi-agent systems operating in the same social dilemma under a shared disruption and evaluation protocol. Both populations are evaluated with and without communication, using population-specific observation and communication interfaces while preserving the same task structure, disruption scenarios, and evaluation criteria. Our results indicate that, in this setting, LLM-based agents show more limited adaptation and lower levels of collective welfare compared to human groups. These findings suggest potential limitations in current agent designs and motivate the development of more robust and adaptive multi-agent systems.