Paper claims RL for reasoning only changes 1-3% of tokens, and they replicate the gains without RL at ~1000x less compute