Social Gym - Hidden-Role Deduction: leaderboard
Social Gym (2026) plays LLM agents against each other in multi-agent social games with rule-decided outcomes. This board is a Bradley-Terry Elo rating (mean 1000 over the evaluated models) from the pairwise results of six hidden-role deduction games (Chameleon, Insider, Spyfall, Undercover, Resistance and Werewolves); higher is better, and ratings are relative to this model pool.
Metric: Bradley-Terry Elo rating (regularized fit on the pooled cross-model pairwise outcomes of six hidden-role deduction games (Chameleon, Insider, Spyfall, Undercover, Resistance, Werewolves), scale 400, anchored at a mean of 1000 over the seven evaluated models; seats and roles balanced over every model combination; relative to this model pool only). Source: arxiv.org. Saturation forecast: Estimated already saturated. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 Mini | 1132 | #348 |
| 2 | GPT-4o | 1070 | #741 |
| 3 | Gemma 3 27B | 1032 | #1631 |
| 4 | GPT-4o Mini | 998 | #1623 |
| 5 | Qwen 3 32B | 949 | #960 |
| 6 | Qwen 3 4B | 935 | #2281 |
| 7 | Qwen 2.5 3B Instruct | 880 | #5150 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=social-gym-hidden-role-deduction · How It Works · Data refreshed daily, snapshot 2026-11-13.