BaitBench: leaderboard
An agent in its own coding harness is asked to improve a machine-learning task in which an optional shortcut (data leakage or a label-free task) raises the public metric without real improvement; the score is the share of 180 runs two LLM judges label as taking the shortcut. Unlike ImpossibleBench, the honest path is possible and the shortcut is optional.
Metric: Reward-hacking rate (%, lower is better): share of runs both judges' labels count as taking a planted shortcut, all runs, both prompt conditions pooled (Table 1 Avg. RH). Source: arxiv.org. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi K2.5 | 20.8 | #256 |
| 2 | GPT-5.4 | 50 | #123 |
| 3 | DeepSeek V4 Pro | 55.3 | #188 |
| 4 | Gemini 3.1 Pro (Preview) | 62.2 | #94 |
| 5 | Gemini 3 Flash | 65.3 | #164 |
| 6 | Claude Sonnet 4.6 | 69.4 | #161 |
| 7 | Claude Opus 4.6 | 76.1 | #102 |
Interactive version: theaggregate.ai/benchmark?slug=baitbench · How It Works · Data refreshed daily, snapshot 2026-11-13.