Definition
Plain language
Teaching an AI to spot its own mistakes by first letting it practice with the answer key in hand, then taking the key away.
As stated in the literature
A method that trains a no-reference student verifier to imitate the feedback a same-model teacher produces when given the reference solution, exploiting the asymmetry that critiquing a solution against an answer key is far easier than diagnosing it from scratch; trained via on-policy distillation so the deployed verifier stays robust on its own trajectory.
Also called: self-trained verifier
Why it matters: It builds a model that can judge its own work without needing the right answer on hand, which matters wherever no answer key exists at deployment.
For example, a model first practices catching errors while peeking at the answer key, then learns to spot those same errors once the key is taken away.
Heard on the show
“They call this self-trained verification.”Episode 099 — How an Open-Book Trick Teaches a Model to Catch Its Own Mistakes