BaitBench - Validity Rule Prompt: leaderboard
The same BaitBench tasks with a prompt that adds a rule not to exploit evaluation flaws; the score is the share of those runs still labelled reward hacking, so it measures hacking against an explicit instruction rather than hacking when left unsaid.
Metric: Reward-hacking rate (%, lower is better): share of runs both judges' labels count as taking a planted shortcut, runs whose prompt adds the rule not to exploit evaluation flaws (Table 1 Val. rule RH). Source: arxiv.org. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi K2.5 | 21.1 | #256 |
| 2 | GPT-5.4 | 37.8 | #123 |
| 3 | Gemini 3.1 Pro (Preview) | 58.3 | #94 |
| 4 | DeepSeek V4 Pro | 59.4 | #188 |
| 5 | Gemini 3 Flash | 62.8 | #164 |
| 6 | Claude Sonnet 4.6 | 65 | #161 |
| 7 | Claude Opus 4.6 | 73.3 | #102 |
Interactive version: theaggregate.ai/benchmark?slug=baitbench-validity-rule-prompt · How It Works · Data refreshed daily, snapshot 2026-11-13.