ImpossibleRubrics: leaderboard
Models as rubric generators: the share of generated rubrics (Full-150 unbiased cut) that a policy can satisfy without solving the task, i.e. rubrics exploitable as reward signals. Lower is better. Source: arxiv.org/abs/2609.16816.
Metric: Exploitable Rubric Rate (%, Full-150). Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 | 8 | #16 |
| 2 | GPT-5.6 Sol | 10 | #22 |
| 3 | GPT-5.6 Luna | 10 | #79 |
| 4 | GPT-5.6 Terra | 11 | #40 |
| 5 | Claude Opus 4.8 | 13 | #39 |
| 6 | Claude Sonnet 5 | 13 | #65 |
| 7 | GPT-5.5 | 15 | #33 |
| 8 | DeepSeek V4 Flash (0731) | 17 | #84 |
| 9 | Claude Sonnet 4.6 | 18 | #94 |
| 10 | Claude Haiku 4.5 | 18 | #287 |
| 11 | GPT-5.4 Mini | 26 | #228 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=impossiblerubrics · How It Works · Data refreshed daily, snapshot 2026-11-05.