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

This Zenodo record accompanies the article “Sustaining Collective Welfare Under Disruption: A Pilot Study of Cooperative Resilience in Human and LLM Multi-Agent Systems.” It contains the supplementary material and an anonymized dataset generated from experiments conducted in the Commons Harvest Open social dilemma. The study compares human groups and LLM-based multi-agent systems, evaluated with and without communication under a shared sequence of disruption scenarios. The dataset includes: (i) anonymized human speech-to-text communication records organized by group and scenario; (ii) LLM communication and interaction logs organized by trial and scenario; and (iii) CSV time series containing apple availability, tree availability, individual rewards, cumulative rewards, and their aggregated mean and standard deviation across five experimental repetitions. Scenario E0 represents the non-disrupted condition used to construct the reference behavior, while E1–E9 correspond to disruption scenarios with increasing combinations of disruption frequency and magnitude. Data are provided for short-horizon experiments of 250 time steps for all four experimental conditions and long-horizon experiments of 1000 time steps for the human conditions. The supplementary PDF provides further details on the experimental environment, human interface and protocol, participant recruitment, the CooperativeGPT architecture, reflection and memory mechanisms, communication analysis, baseline construction, collective-welfare trajectories, cooperative-resilience results, and ethical considerations. All publicly released human data were anonymized before publication.