BehaviorBench (Behavioral Traces) - Belief Choice (DirectGen): leaderboard
BehaviorBench (Yang et al., 2026) tests whether a model can predict what a real pseudonymous user will do, using 2,000 wallets reconstructed from public prediction-market and on-chain records with chronologically held-out targets. The model sees the recent raw history of the wallet (DirectGen). The board is the accuracy in percent of predicting the side (YES or NO) the user finally holds in a market; chance is 50 and higher is better.
Metric: Choice accuracy (%): predicted final revealed side (YES or NO) of the wallet in the target market; history interface DirectGen (the target wallet's most recent visible history, up to 30 events for Belief); 2,000 pseudonymous prediction-market wallets reconstructed from public Polymarket metadata and Polygon on-chain logs, 22,204 chronologically held-out wallet-market test instances; chance 50, context-only baseline 53.13; higher is better. Source: arxiv.org. Saturation forecast: Around March 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 70.75 | #123 |
| 2 | Claude Opus 4 | 70.22 | #322 |
| 3 | Gemma 3 12B (IT) | 70.21 | #1837 |
| 4 | Gemma 3 27B (IT) | 70.08 | #1289 |
| 5 | Claude 3.7 Sonnet | 67.68 | #458 |
| 6 | GPT-5.4 Mini | 66.34 | #389 |
| 7 | GPT-OSS-20B | 66 | #1147 |
| 8 | GPT-5.4 Nano | 65.8 | #635 |
| 9 | Llama 4 Scout Instruct | 64.77 | #1543 |
| 10 | Ministral-3-14B-Instruct-2512 | 59.82 | #1593 |
| 11 | NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | 58.63 | #1767 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=behaviorbench-behavioral-traces-belief-choice-directgen · How It Works · Data refreshed daily, snapshot 2026-11-13.