Avalon-ToM-Bench - Perspective-Taking: leaderboard
Avalon-ToM-Bench (arXiv 2608.09638v1, Table 1): theory of mind in The Resistance: Avalon, where models judge perspective-constrained true/false statements about players' knowledge and motives from game state and public history. Accuracy on Perspective-Taking (epistemic inference) items, the hardest ability in the paper; balanced true/false, chance 50%. Higher is better.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Around February 2027. 27 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 82.35 | #105 |
| 2 | Gemini 2.5 Pro | 79.41 | #159 |
| 3 | Gemini 2.5 Flash | 77.27 | #255 |
| 4 | Qwen 3 4B 2507 (Thinking) | 72.73 | #527 |
| 5 | GPT-5 Mini | 69.12 | #193 |
| 6 | Mistral 7B Instruct (v0.3) | 69.12 | #1255 |
| 7 | Llama 3.2 3B Instruct | 68.89 | #1327 |
| 8 | Qwen 3.5 9B (Thinking) | 68.63 | #391 (Qwen 3.5 9B) |
| 9 | Qwen 3 8B (Thinking) | 68.42 | #678 (Qwen 3 8B) |
| 10 | Gemma 3 4B (IT) | 67.65 | #997 |
| 11 | GPT-OSS-120B | 63.24 | #346 |
| 12 | GPT-4o Mini | 63.24 | #598 |
| 13 | Llama 3 8B Instruct | 63.24 | #1116 |
| 14 | Phi-4 | 63.24 | #707 |
| 15 | Gemma 3 27B (IT) | 61.76 | #528 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=avalon-tom-bench-perspective-taking · How It Works · Data refreshed daily, snapshot 2026-11-05.