Hidden rules get learned, cooperation does not: eight LLMs play tens of millions of social dilemmas
Lead
OpenSocialJax turns Clean Up and Commons Harvest into tens of millions of hidden-rule configurations, and across five repeated trials eight large language models learn many rules yet still concentrate the cleaning burden and exhaust the shared orchard every episode.
Story
A social dilemma benchmark now tests whether agents can infer hidden environmental rules through repeated interaction while adapting to other agents, rather than only playing fixed rules. Melting Pot and SocialJax offer diverse social dilemma environments, but they are not designed to vary the hidden environmental dynamics behind a dilemma or to evaluate how agents discover and adapt to them over repeated interactions. Each episode samples one configuration from tens of millions of possibilities and holds it fixed across five trials, systematically varying spatial layouts and hidden rules governing tool crafting, payoffs, and resource dynamics.
Agents carry two kinds of memory across trials: the harness keeps an objective record of what the agent experienced, and the agent writes free-form notes stating its own conclusions and plans. It does not rely on a growing conversation history, and it never sees the hidden rule, the other agents' prompts or scores, and cannot communicate with them. When a trial ends the model rewrites its notes into a short list of lessons, and only the rule evidence, trial summaries and that list carry into the next trial, with the environment reset on the same map and rule.
What to watch
The benchmark targets multi-agent evaluation in sequential social dilemmas, covering OpenCleanup and OpenHarvest, with three seeds per environment, five trials per seed, at most 400 steps per trial, and OpenHarvest ending a trial early once no apple remains. It suits studying how agents explore, infer and adapt under hidden rules, and the authors propose extending the multi-trial protocol with explicit communication and more sophisticated theory of mind mechanisms.
The authors state that training learning-based agents that adapt to the latent configuration remains an open challenge; rule discovery does not necessarily mean agents use the rules effectively, and no model fully resolves either dilemma, with stronger models only delaying the failures.