BehaviorBench - Scientific Workflow Prediction: leaderboard
BehaviorBench (University of Michigan, 2026) evaluates foundation models on behavioral science tasks at the individual and population levels, using economic-game records, Big Five personality surveys and research literature. This board is the BLEURT score of generated research-workflow aspects (key idea, method, outcome, impact and title) against 2025 economics and behavioral science papers; higher is better.
Metric: BLEURT score of generated research-workflow aspects (key idea, method, outcome, projected impact and title, each from the preceding aspects) against 2025 American Economic Review and Nature Human Behaviour papers; higher is better. Source: arxiv.org. Saturation forecast: Around June 2027. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 0.48 | #102 |
| 2 | Claude Sonnet 4.6 | 0.47 | #161 |
| 3 | Gemini 3.1 Pro (Preview) | 0.47 | #94 |
| 4 | GPT-5.4 (High) | 0.46 | #123 (GPT-5.4) |
| 5 | GPT-4.1 | 0.46 | #495 |
| 6 | GPT-5.4 Mini (High) | 0.45 | #389 (GPT-5.4 Mini) |
| 7 | Qwen 3 4B | 0.45 | #2281 |
| 8 | Gemini 3.1 Flash Lite (Preview) | 0.43 | #418 |
| 9 | Llama 3.3 70B Instruct | 0.43 | #1357 |
| 10 | DeepSeek V3.2 | 0.43 | #406 |
| 11 | Claude Haiku 4.5 | 0.43 | #568 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=behaviorbench-scientific-workflow-prediction · How It Works · Data refreshed daily, snapshot 2026-11-13.