GroupToM-Bench - Belief: leaderboard
GroupToM-Bench (2026) tests group-level theory of mind in multimodal models: 240 expert-seeded multi-agent scenarios, each a scene image with dialogue, are probed along a seven-level audit from individual mental states to collective outcomes. This board is level 1 belief tracking. Scored by strict exact-match accuracy in percent on multi-select multiple choice (random 6.7); higher is better.
Metric: strict exact-match accuracy (%) over multi-select options (random 6.7) on Level 1 belief tracking (second-order and deeper beliefs under information asymmetry and deception), multiple choice; 240 multimodal multi-agent scenarios (scene image plus dialogue); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o | 79.8 | #741 |
| 2 | Gemini 3 Pro | 78.9 | #137 |
| 3 | GPT-5 | 76.7 | #167 |
| 4 | Claude Haiku 4.5 | 75.1 | #568 |
| 5 | GPT-5 Mini | 70.4 | #348 |
| 6 | Llama 3.2 11B | 66 | #6383 |
| 7 | Qwen 2.5 VL 7B Instruct | 65.8 | #1833 |
| 8 | GPT-5 Nano | 63.3 | #889 |
| 9 | Qwen 2 VL 7B | 58.3 | #2261 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=grouptom-bench-belief · How It Works · Data refreshed daily, snapshot 2026-11-13.