CSD: Common-Support Distillation for Cross-Tokenizer Knowledge Transfer
arXiv:2610.28895v1 Announce Type: new Abstract: Cross-tokenizer logits distillation enables knowledge transfer across model families, allowing students to learn from stronger or more specialized teachers. However, differences in vocabulary and segmentation make teacher and student next-token distri…
Read the full story at arXiv cs.CL ↗
Timeline · 1 report
- 2026-10-22 04:00 · arXiv cs.CL
CSD: Common-Support Distillation for Cross-Tokenizer Knowledge Transfer