BehaviorBench (Behavioral Traces) - Trade Amount Error (RetrievalGen): 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 behavior of similar wallets retrieved from a separate support pool (RetrievalGen). The board is the median absolute error of the predicted transaction amount, in paper units and unbounded; lower is better.
Metric: Median absolute error of the predicted trade amount (paper units, unbounded); history interface RetrievalGen (behaviorally similar examples retrieved from a disjoint support pool of wallets); 2,000 pseudonymous prediction-market wallets reconstructed from public Polymarket metadata and Polygon on-chain logs, about 59,560 chronologically held-out test transactions; unparsable outputs count as failures; context-only baseline 31.79; lower is better. Source: arxiv.org. Saturation forecast: Not forecast. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 3 12B (IT) | 8.29 | #1837 |
| 2 | Claude 3.7 Sonnet | 9.76 | #458 |
| 3 | GPT-5.4 | 10 | #123 |
| 4 | GPT-5.4 Nano | 10.22 | #635 |
| 5 | Llama 4 Scout Instruct | 10.81 | #1543 |
| 6 | GPT-5.4 Mini | 11.13 | #389 |
| 7 | NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | 11.53 | #1767 |
| 8 | GPT-OSS-20B | 14.5 | #1147 |
| 9 | Gemma 3 27B (IT) | 14.65 | #1289 |
| 10 | Ministral-3-14B-Instruct-2512 | 21.8 | #1593 |
| 11 | Claude Opus 4 | 22.67 | #322 |
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-trade-amount-error-retrievalgen · How It Works · Data refreshed daily, snapshot 2026-11-13.