Reproduce “Decoupled Mode Connectivity for Base-to-Novel Generalization in Vision-Language Models” 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

Imad Eddine Marouf, Khalid Oublal, Enzo Tartaglione, Stéphane Lathuilière

The authors introduce DMC to adapt CLIP with separate prompts for fitting seen classes and preserving recognition of unseen classes, then train a path between text features. On an 11-dataset benchmark, with 16 examples per base class, DMC raises mean harmonic mean over MergeTune by 0.78 for CoOp and 0.33 for KgCoOp, but evidence supports only these prompt-tuning methods and requires distinct endpoints.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.