General Social Agents: 883K novel games beat game theory at predicting humans
soumitrashukla9 · x · 2026-10-02
In arXiv paper 'General Social Agents' (Manning & Horton), AI agents built from theory-grounded natural language instructions, human data from a few 'seed' games, and pretraining knowledge generalize across novel social settings. Across a population of 883,320 novel games, preregistered experiments show the agents predict initial human play on 1,500 sampled games better than cognitive hierarchy models, game-theoretic equilibria, and out-of-the-box agents — and on separate new games they outpredict even the most plausibly relevant published human data. Horton frames this as a call for open research on building and validating digital twins.
More from Research
- One policy for millions of robots: team builds generalist motor control policy trained on 200+ real robot models — breadli428 · 2026-10-02
- Google's Cogentic uses multi-agent orchestration to produce novel results on five open math problems — KyeGomezB · 2026-10-02
- Pioneer Labs unveils sPL.001, the first engineered microbe designed to build with Martian soil — jiqizhixin · 2026-10-02
- Survey: 76% of Americans say torturing sentient AI is wrong, 63% want AGI banned — jacyanthis · 2026-10-02
- New paper explores training risk aversion into AI to make misaligned models negotiable — sethlazar · 2026-10-02
- Loop Scaling Laws: First Scaling Law Jointly Modeling Recurrence and MoE Sparsity, Promising ~2x Parameter Savings — anirudhg9119 · 2026-10-02