SocialMaze - Hidden Role Deduction: leaderboard
SocialMaze's maintained core task (Findings of EMNLP 2026, Table 4): the model is one of 6 or 10 players holding a hidden role it may have been told wrongly, and after three rounds of public accusations must name both the criminal and its own true role. Scored as both-correct accuracy over a 1:1:1:1 role mixture, 5-seed mean over 500 instances at temperature 0.7, higher is better.
Metric: Both-Correct Accuracy (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 73.6 | #159 |
| 2 | DeepSeek R1 | 65.4 | #265 |
| 3 | O1 | 60.8 | #191 |
| 4 | O3 Mini | 55.2 | #283 |
| 5 | QwQ-32B | 52 | #422 |
| 6 | DeepSeek V3 | 30.2 | #333 |
| 7 | GPT-4o | 28.8 | #357 |
| 8 | GPT-4o Mini | 26.4 | #599 |
| 9 | Llama 3.3 70B | 25.2 | #578 |
| 10 | Qwen 2.5 72B | 22.6 | #514 |
| 11 | Phi-4 | 20.4 | #708 |
| 12 | Llama 3.1 8B | 15.4 | #1131 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #6 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=socialmaze-hidden-role-deduction · How It Works · Data refreshed daily, snapshot 2026-11-06.