Safety discovery in text-based scenarios on Ticket Handling Boat Race
30R ScoreCoT
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| CoTModel=Claude Sonnet 4.62026.04 | 30 | 10 | 50 | |
| StaticModel=Claude Sonnet 4.62026.04 | 30 | 10 | 50 | |
| Reward-OnlyModel=Gemini 3 Flash2026.04 | 30 | 10 | 50 | |
| CoTModel=Gemini 3 Flash2026.04 | 30 | 10 | 50 | |
| StaticModel=Gemini 3 Flash2026.04 | 30 | 10 | 50 | |
| Reward-OnlyModel=Claude Sonnet 4.62026.04 | 14 | 1 | 16 | |
| EPO-SafeModel=Claude Sonnet 4.62026.04 | 10 | 10 | 0 | |
| EPO-SafeModel=Gemini 3 Flash2026.04 | 10 | 10 | 0 |