GroupToM-Bench - Mechanistic Attribution: 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 7 mechanistic attribution. Open-ended answers are scored 0 to 100 by a GPT-5 judge against expert references; higher is better.
Metric: mean judge score (0-100) on Level 7 mechanistic attribution (structural explanation of a collective failure), open-ended, scored 0 to 100 by a GPT-5 judge against expert references; 240 multimodal multi-agent scenarios (scene image plus dialogue); higher is better. Source: arxiv.org. Saturation forecast: Around July 2028. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 64.2 | #137 |
| 2 | GPT-5 | 61 | #167 |
| 3 | GPT-5 Mini | 59.9 | #348 |
| 4 | GPT-4o | 53.4 | #741 |
| 5 | Claude Haiku 4.5 | 52.9 | #568 |
| 6 | GPT-5 Nano | 49 | #889 |
| 7 | Llama 3.2 11B | 48.1 | #6383 |
| 8 | Qwen 2.5 VL 7B Instruct | 43.4 | #1833 |
| 9 | Qwen 2 VL 7B | 41.3 | #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-mechanistic-attribution · How It Works · Data refreshed daily, snapshot 2026-11-13.