The authors systematically replaced the 12 independently parameterized transformer blocks of a ViT-B vision encoder with N distinct blocks each applied M times (N×M = 12), creating recursive CLIP variants.
Reusing transformer blocks repeatedly preserves CLIP's vision-language capabilities
Recursive CLIP models achieve parameter efficiency while maintaining strong zero-shot and retrieval performance through depth-wise block reuse.
Big Tech
Szymon Mazurek · Alberto Presta · Michał Byra
Samsung AI Center Warsaw · AGH University of Krakow · ACC Cyfronet AGH · Polish Academy of Sciences
Research Digest··3 min read
The authors introduce recursive transformer backbones for CLIP where a small set of blocks is applied multiple times, reducing distinct parameters while keeping computational depth constant.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.