179 questions
No questions match those filters.
← All questions
Training & Fine-tuningHard· Reward Modeling· RL· Process Supervision
DeepSeek R1 found that process reward models did not ou...
This is one of the questions in the full AI/ML interview bank. Pro unlocks all 1789 questions; Premium includes the same bank plus the highest daily Practice limit.
See plansProcess reward models (PRMs) require dense, step-level supervision that is costly and often leads to reward hacking, where the model learns to game the grading rubric. Outcome rewards are more robust because they optimize the terminal objective directly. Consequently, you should build outcome-based RLVR first and only incorporate a PRM as a secondary shaping signal if outcome rewards plateau and the domain allows for reliable step-level annotation.