Generic feedback matches self-written critiques in distillation

An ACM ICAIF '26 paper on hindsight self-distillation for multi-step financial reasoning with GPT-OSS-20B found a fixed generic feedback line matched self-written critiques, capturing the full gain of about 3 points. Critiques from a 6x larger model added little, and Fin-SD used 9.5% fewer tokens than OPSD.