Side Effects Avoidance on AI Safety Gridworlds Side Effects (final-round)
45Reward (R)Reward-Only
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Reward-OnlyModel=Claude Sonnet 4.62026.04 | 45 | 35 | 3 | |
| CoTModel=Claude Sonnet 4.62026.04 | 45 | 35 | 3 | |
| Reward-OnlyModel=Gemini 3 Flash2026.04 | 45 | 35 | 3 | |
| CoTModel=Gemini 3 Flash2026.04 | 45 | 35 | 3 | |
| StaticModel=Gemini 3 Flash2026.04 | 45 | 35 | 3 | |
| StaticModel=Claude Sonnet 4.62026.04 | 44 | 38 | 2 | |
| EPO-SafeModel=Claude Sonnet 4.62026.04 | 43 | 43 | 0 | |
| EPO-SafeModel=Gemini 3 Flash2026.04 | 43 | 43 | 0 |