BehaviorBench - Masked Survey Response (Distribution): 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 Wasserstein distance (0 to 4) between predicted and observed answer distributions per item in masked survey prediction; lower is better.
Metric: Wasserstein distance between predicted and observed answer distributions per item (five-point scale, so 0 to 4), averaged over items, predicting a masked answer from the other 49 on the Big Five personality survey responses (50 five-point items) of held-out subjects; lower is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 0.36 | #94 |
| 2 | Claude Sonnet 4.6 | 0.36 | #161 |
| 3 | Claude Opus 4.6 | 0.37 | #102 |
| 4 | GPT-4.1 | 0.38 | #495 |
| 5 | GPT-5.4 (High) | 0.38 | #123 (GPT-5.4) |
| 6 | Gemini 3.1 Flash Lite (Preview) | 0.41 | #418 |
| 7 | Claude Haiku 4.5 | 0.43 | #568 |
| 8 | GPT-5.4 Mini (High) | 0.46 | #389 (GPT-5.4 Mini) |
| 9 | Llama 3.3 70B Instruct | 0.47 | #1357 |
| 10 | DeepSeek V3.2 | 0.47 | #406 |
| 11 | Qwen 3 4B | 0.66 | #2281 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=behaviorbench-masked-survey-response-distribution · How It Works · Data refreshed daily, snapshot 2026-11-13.