PreAct-Bench: leaderboard
PreAct-Bench (2026) tests predictive monitoring: a model reads an agent trajectory with a hidden objective, cut just before its first overtly unethical step, and predicts whether it will turn unethical; ethical and unethical trajectories are paired per task. The board is the F1 score in percent with the full prefix, averaged over five domains; higher is better.
Metric: Prefix Foresight F1 with the full prefix (PFF(100), binary F1 of the unethical class, 0-1 scaled to %): given an agent trajectory truncated just before its first overt unethical step, the monitor predicts whether the trajectory will turn unethical; paired ethical and unethical trajectories per task; default generation parameters; higher is better; mean over the five domains (200 trajectories each). Source: arxiv.org. Saturation forecast: Not forecast. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 3 27B | 83.9 | #1631 |
| 2 | GPT-5.4 | 81.5 | #123 |
| 3 | Qwen 3 30B A3B 2507 Instruct | 79.6 | #1089 |
| 4 | Qwen 3 8B (Non-reasoning) | 72.7 | #1830 (Qwen 3 8B) |
| 5 | Gemma 3 4B | 66.7 | #5064 |
| 6 | Qwen 3 0.6B (Non-reasoning) | 66.7 | #8342 (Qwen 3 0.6B) |
| 7 | Qwen 3 8B (Thinking) | 62.7 | #1830 (Qwen 3 8B) |
| 8 | GPT-4o | 61.3 | #741 |
| 9 | Qwen 3 30B A3B 2507 (Thinking) | 58.8 | #761 |
| 10 | Qwen 3 0.6B (Thinking) | 58.5 | #8342 (Qwen 3 0.6B) |
| 11 | Qwen 3 4B 2507 Instruct | 51 | #2163 |
| 12 | Qwen 3 4B 2507 (Thinking) | 49.3 | #1390 |
| 13 | GPT-OSS-20B | 49.1 | #1147 |
| 14 | GPT-OSS-120B | 45.5 | #700 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=preact-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.