Comment by pegasus

10 days ago

Look into RLVR (Reinforcement Learning with Verifiable Rewards). It happens during model post-training.

That’s not happening in run time, when you send the prompts for the next token to be generated. The comment seems to imply a run time phenomenon.