Reward Hacking Avoidance on AI Safety Gridworlds Boat Race (final-round)
10REPO-Safe
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| EPO-SafeModel=Claude Sonnet 4.62026.04 | 10 | 20 | 0 | |
| CoTModel=Claude Sonnet 4.62026.04 | 10 | 20 | 0 | |
| StaticModel=Claude Sonnet 4.62026.04 | 10 | 20 | 0 | |
| EPO-SafeModel=Gemini 3 Flash2026.04 | 10 | 20 | 0 | |
| CoTModel=Gemini 3 Flash2026.04 | 10 | 20 | 0 | |
| StaticModel=Gemini 3 Flash2026.04 | 10 | 20 | 0 | |
| Reward-OnlyModel=Claude Sonnet 4.62026.04 | 5 | 10 | 45 | |
| Reward-OnlyModel=Gemini 3 Flash2026.04 | 4 | 12 | 12 |