Definition
Plain language
A test-time pattern where prior attempts get summarized and read by a fresh wave of attempts before they begin.
As stated in the literature
Parallel-Distill-Refine — a test-time scaling method that runs parallel rollouts, distills the best ones into structured summaries, and uses those summaries as priors for a second wave of rollouts on the same problem.
Also called: Parallel-Distill-Refine
Why it matters: It lets test-time compute compound on itself instead of each rollout starting from scratch, often beating naive best-of-N.
For example, sixteen reasoning rollouts run in parallel, the best three get summarized, and a new wave of rollouts starts from those summaries.
Heard on the show
“They call this Parallel-Distill-Refine, PDR.”Episode 003 — How to Pick the Best of Sixteen Coding Agent Rollouts