AffToM - Counterfactual Reasoning: leaderboard
Video-based recursive affective theory-of-mind questions across individual and multiparty social interactions. MCQ accuracy (%). AffToM v1 Table2,1822MCQs plus100open-ended questions from258videos; temperature0 and64uniformframes unless source exception. Native base models and separately trained AffToM-R1 remain distinct. MCQ exact accuracy in%; open answers graded by GPT-5-mini on1–5 Emotion Correctness/Factual Consistency rubrics. Human250-question reference retained in source pins but not emitted as a model. Different LLaVA-OneVision1.5 parameter sizes stay distinct.
Metric: MCQ accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 23 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 (Benchmark Sweep Run: source=AffToM; temperature0; video64uniformframes; native Table2 checkpoint; no subtitle ablation) | 74.01 | |
| 2 | AffToM-R1 (Qwen3.5-4B) (Benchmark Sweep Run: source=AffToM; temperature0; video64uniformframes; native Table2 checkpoint; no subtitle ablation) | 66.67 | |
| 3 | Qwen 2.5 VL 72B Instruct (Benchmark Sweep Run: source=AffToM; temperature0; video64uniformframes; native Table2 checkpoint; no subtitle ablation) | 66.1 | |
| 4 | Qwen 3.5 27B (Benchmark Sweep Run: source=AffToM; temperature0; video64uniformframes; native Table2 checkpoint; no subtitle ablation) | 66.1 | |
| 5 | Seed 2.0 Lite (Benchmark Sweep Run: source=AffToM; temperature0; video64uniformframes; native Table2 checkpoint; no subtitle ablation) | 66.1 | |
| 6 | Qwen 3.5 9B (Benchmark Sweep Run: source=AffToM; temperature0; video64uniformframes; native Table2 checkpoint; no subtitle ablation) | 64.97 | |
| 7 | Qwen 3 VL 32B Instruct (Benchmark Sweep Run: source=AffToM; temperature0; video64uniformframes; native Table2 checkpoint; no subtitle ablation) | 64.41 | |
| 8 | InternVL3.5-38B (Benchmark Sweep Run: source=AffToM; temperature0; video64uniformframes; native Table2 checkpoint; no subtitle ablation) | 59.32 | |
| 9 | Qwen 3.5 4B (Benchmark Sweep Run: source=AffToM; temperature0; video64uniformframes; native Table2 checkpoint; no subtitle ablation) | 58.19 | |
| 10 | LLaVA-OneVision-1.5 8B (Benchmark Sweep Run: source=AffToM; temperature0; video64uniformframes; native Table2 checkpoint; no subtitle ablation) | 57.63 |
Interactive version: theaggregate.ai/benchmark?slug=afftom-counterfactual-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-06.