“Most of the LLM capability progress in 2025 came from longer Reinforcement Learning runs, particularly RLVR, rather than from increasing model size.”