How Much Does the Hint Matter? A generic feedback line matches self-written critiques in hindsight self-distillation (ACM ICAIF '26) [R]

Overview

We study hindsight self-distillation for multi-step financial reasoning on GPT-OSS-20B. The student finds its own first wrong step and writes a critique. A frozen copy, conditioned on the prefix plus the critique, is distilled back into the student under reverse KL. Notably, replacing the critique with a fixed generic feedback line matches the full gain of about 3 points, and critiques from a 6x larger model do not materially improve performance, suggesting that the positional information provid

SpinGraph analysis pending — check back after processing.