Definition
Plain language
Getting some points for solving part of a problem instead of all-or-nothing.
As stated in the literature
Graded per-subtask scoring that returns a structured signal about which input regimes a solution handles, enabling targeted iterative refinement rather than blind retry.
Also called: partial-credit
Why it matters: It converts a dead-end failure into a usable clue, letting the next attempt target the specific case that went wrong.
For example, a program that handles lists of up to a thousand items but times out on a million earns points for the small case and reveals exactly where it broke.
Heard on the show
“It falls apart on long ones, where there are many partial-credit situations and many ways to be partially right.”Episode 017 — When the Agent Grades Its Own Homework: A Brutal New Benchmark for AI Workers