The authors developed OpenSocialJax, built on Clean Up and Commons Harvest from Melting Pot and SocialJax.
LLM agents discover hidden rules in social dilemmas but fail to sustain cooperation
A new benchmark reveals that even the strongest models cannot jointly adapt to unfamiliar environmental dynamics and other agents' behavior.
Big Tech
Zihao Guo · Meng Fang · Joel Z. Leibo · Yali Du
King's College London · University of Liverpool · The Alan Turing Institute · Google DeepMind
Research Digest··3 min read
The authors introduce OpenSocialJax, a benchmark extending Clean Up and Commons Harvest with procedurally generated layouts and hidden rules from tens of millions of configurations.
Why this paper
From Google DeepMind and 3 others
In one line
LLM agents in OpenSocialJax discover hidden environment rules but fail to sustain cooperation, depleting shared resources in social dilemmas.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§