SCOPD recovers 92% of VLM performance after aggressive visual token pruning

A University of Toronto team introduced SCOPD, an online self-distillation method showing that pruned visual evidence is not lost—Pass@K sampling often recovers correct answers—and restoring 92% of full-context performance with only 10% of visual tokens.

2026-09-29 ~ 2026-10-01 · 2 related posts

• SCOPD Distillation Keeps 92% of VLM Performance at 10% Visual Tokens — uoft · 2026-09-29
• SCOPD self-distillation recovers 92% of full-context VLM accuracy with 90% fewer visual tokens — CSProfKGD · 2026-10-01