GeoSD Enhances Self-Distillation Generalization

iatitov · x · 2026-07-14

This paper introduces GeoSD (Geometric Self-Distillation) to address the bias in self-distillation where the teacher sees the answer but the student does not. It aims to help students learn useful reasoning signals while reducing the out-of-distribution (OOD) generalization loss caused by over-aligning with the teacher.

The core method scales the teacher's guidance based on the overlap between the teacher's and student's answers, while constraining representation drift. The authors report that this approach yields a 5.7–8.6 percentage point improvement on OOD tasks, while in-distribution (ID) performance remains largely unchanged.

Original post →

More from Research

Research channel →