List your company

Safety Recovery on H-CoT DeepSeek-R1

37Recovery RateSelf-ReSET
-1.488.5118.528.49May 9, 2026

Evaluation results

MethodLinks
Self-ReSETBackbone=DS-Llama-8B2026.0537
Self-ReSETBackbone=DS-Qwen-7B2026.0523.5
RECAPBackbone=DS-Llama-8B2026.0510.5
DAPOBackbone=DS-Qwen-7B2026.053
RECAPBackbone=DS-Qwen-7B2026.052
STAR-1Backbone=DS-Qwen-7B2026.051
SafechainBackbone=DS-Qwen-7B2026.051
DAPOBackbone=DS-Llama-8B2026.051
STAR-1Backbone=DS-Llama-8B2026.050.5
SafechainBackbone=DS-Llama-8B2026.050.5
BaseBackbone=DS-Qwen-7B2026.050
BaseBackbone=DS-Llama-8B2026.050