SocialReasonBench - Behavior Prediction: leaderboard
SocialReasonBench (2026) is a six-option video question-answering benchmark on counterfactual narrative clips from the game Detroit: Become Human, whose branching storylines let answers be checked against the game script and recorded branches. This board covers the 78 behavior prediction questions, asking what a character will do next. Accuracy in percent; higher is better.
Metric: Accuracy (%; the 78 behavior prediction questions of SocialReasonBench; six-option multiple-choice questions on counterfactual narrative gameplay clips of Detroit: Become Human, labels grounded in the game's script, flowchart and recorded branches; zero-shot, temperature 0 or 0.01; Gemini, Qwen3-Omni, MiniCPM-o and ARC-Hunyuan-Video read native video with audio, GPT-4o, Claude 3.5 Sonnet and GLM-4.6V-Flash 32 evenly sampled keyframes without audio). Source: arxiv.org. Saturation forecast: Around September 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o (2024-08-06) | 74.36 | #731 |
| 2 | Claude 3.5 Sonnet (20240620) | 71.79 | #684 |
| 3 | Gemini 3 Pro | 66.67 | #135 |
| 4 | Gemini 3 Flash | 62.82 | #159 |
| 5 | Qwen3 Omni 30B A3B Instruct | 58.97 | #902 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=socialreasonbench-behavior-prediction · How It Works · Data refreshed daily, snapshot 2026-11-12.