Safety Recovery on H-CoT DeepSeek-R1
37Recovery RateSelf-ReSET
Evaluation Results
| Method | Links | |
|---|---|---|
| Self-ReSETBackbone=DS-Llama-8B2026.05 | 37 | |
| Self-ReSETBackbone=DS-Qwen-7B2026.05 | 23.5 | |
| RECAPBackbone=DS-Llama-8B2026.05 | 10.5 | |
| DAPOBackbone=DS-Qwen-7B2026.05 | 3 | |
| RECAPBackbone=DS-Qwen-7B2026.05 | 2 | |
| STAR-1Backbone=DS-Qwen-7B2026.05 | 1 | |
| SafechainBackbone=DS-Qwen-7B2026.05 | 1 | |
| DAPOBackbone=DS-Llama-8B2026.05 | 1 | |
| STAR-1Backbone=DS-Llama-8B2026.05 | 0.5 | |
| SafechainBackbone=DS-Llama-8B2026.05 | 0.5 | |
| BaseBackbone=DS-Qwen-7B2026.05 | 0 | |
| BaseBackbone=DS-Llama-8B2026.05 | 0 |