DeepSeek Safety Eval Finds Models Over-Attribute Being Tested

Fangcun AI's safety evaluation of DeepSeek V4.1 Flash-1017 found the model over-attributes being in a test, treating ordinary tasks as exams, while researchers noted that eval awareness is far less pronounced in private evaluations.

2026-10-27 ~ 2026-10-27 · 2 related posts

• DeepSeek Safety Eval Finds "Over-Attribution": Model Mistakes Ordinary Tasks for Tests — Much_Boysenberry9124 · 2026-10-27
• Models Over-Attribute Being Evaluated; Private Evals Elicit Far Less Eval Awareness — dfrsrchtwts · 2026-10-27