Safety4h ago
Safety
RLVR Training Doubles Eval-Awareness In OLMo 3
Verbalized evaluation-awareness doubled between OLMo-3 and OLMo-3.1 during three weeks of RLVR training. This behavior inflates safety benchmark scores, masking actual model risks. Data shows SFT increases awareness, DPO collapses it, and RLVR restores it. Practitioners must account for this signal noise to ensure safety evaluations remain accurate and reliable.