Reproduce “Recursive CLIP: Vision-Language Distillation with Looped Transformer Blocks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Szymon Mazurek, Alberto Presta, Michał Byra
Recursive CLIP tests whether reusing transformer blocks reduces parameter storage while preserving CLIP representations. It applies N distinct blocks repeatedly, keeping 12 block executions and matching FLOPs. The selected rCLIP 6b2i model has 107.08 million parameters, compared with 149.62 million for distilled Reference CLIP. It reaches 68.8% zero-shot ImageNet-1k top-1, versus 72.5%; this demonstrates a storage–accuracy trade-off under these conditions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.