outcome

Definition

The final result of a process, as opposed to the steps that produced it. The distinction matters in reasoning models: outcome supervision rewards a rollout only for its final answer, which is cheap and often automatable, while process supervision scores each intermediate step. Outcome-only signal is easy to collect but reinforces lucky answers reached through faulty reasoning — the main argument for process reward models.

Explainers using this term

• Paper Walkthrough: EarlyEval — Making Agent Evaluation Cheaper by Stopping EarlyEarlyEval: Cheaper Agent Evaluation via Early Outcome Prediction