Glossary · Term

discriminative utility

← all terms

Definition

Plain language

How useful a grading rubric is, measured by whether a weaker grader using it picks the right winner.

As stated in the literature

A definition of rubric quality as the increase in a frozen low-capacity judge's preference-classification accuracy when conditioned on the rubric.

Why it matters: It gives a concrete, measurable definition of when a rubric is actually doing work, instead of just looking thorough.

For example, a weaker model picks the human-preferred answer 60% of the time on its own, but 85% of the time when shown the rubric, so the rubric scores a 25-point utility gain.

Heard on the show

“The paper calls this discriminative utility, and it's the entire definition of rubric quality.”
Episode 019 — When the Best Reward Model Trains the Worst Policy: Inside EvoLM

Mentioned in 1 episode

  1. 019
    When the Best Reward Model Trains the Worst Policy: Inside EvoLM

Related terms