New Study Proposes Sink-Guided Key Editing to Address Attention Sink Tokens in Vision-Language Models
A new arXiv paper examines how a small number of sink tokens in vision-language models attract a disproportionate share of attention and proposes a method for editing attention keys guided by these sink tokens.
Read the full story at arXiv cs.CV ↗
Timeline · 1 report
- 2026-11-19 05:00 · arXiv cs.CV
Beyond the Sink Token: Sink-Guided Key Editing in Vision-Language Models