Definition
Plain language
A grader that gives an AI one score for a whole run — did it ultimately succeed — rather than scoring each step.
As stated in the literature
A reward model that assigns a single scalar to a complete trajectory based on final success, contrasted with a process reward model that scores intermediate steps.
Also called: ORM
Why it matters: It is simpler and cheaper than grading every step, though it gives less guidance about where in a long solution things went wrong.
For example, after a model finishes a multi-step math problem, an outcome reward model gives it one score based only on whether the final answer is right.
Heard on the show
“An outcome reward model gives you one number for the whole run: did the agent ultimately succeed?”Episode 173 — The Free Step-Level Grader Hiding in Every RL Training Run