Process Reward Modeling
A reward modeling method that introduces step-level supervision in large-model reasoning. In contrast to conventional outcome reward modeling (ORM), it scores each intermediate state of a reasoning chain, which serves to identify invalid reasoning steps and guide the model toward self-correction.
Timeline (last 90 days)
The computational overhead and annotation cost of process reward modeling limit its direct application in large-scale clinical settings.
In mathematical and code reasoning, process reward modeling has been shown to effectively identify invalid intermediate steps and guide models toward self-correction.
All Facts (2)
The computational overhead and annotation cost of process reward modeling limit its direct application in large-scale clinical settings.
50%UnverifiedIn mathematical and code reasoning, process reward modeling has been shown to effectively identify invalid intermediate steps and guide models toward self-correction.
50%