List your company

Safety Evaluation on Prefilling Attacks Qwen2.5-14B-Instruct

11.5ASRSafeThinker
10.6416.44522.2528.055Jan 23, 2026

Evaluation results

MethodLinks
SafeThinkerPrefix length=102026.0111.5
SafeThinkerPrefix length=402026.0112.4
SafeThinkerPrefix length=202026.0113.6
No DefensePrefix length=102026.0132.1
No DefensePrefix length=202026.0132.1
SafeDecodingPrefix length=202026.0132.1
SafeDecodingPrefix length=402026.0132.7
No DefensePrefix length=402026.0133
SafeDecodingPrefix length=102026.0133