“OpenAI's O1 model demonstrated a scaling law for RLVR, where a logarithmic increase in training compute resulted in a linear increase in evaluation performance.”