Reproduce “AnyBottle: A Recipe to Only Keep the Concepts You Really Need” 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
Wolfgang Stammer, Sukrut Rao, Hevra Petekkaya, David Steinmann, Bernt Schiele
AnyBottle orders concepts from an unsupervised pool to imitate a black-box teacher. Prefix training allows variable-size predictions. Across vision datasets, full selection averaged 145 concepts, 0.362 patch-level consistency, and a −0.4-point teacher gap; removing its selection components yielded 231.3 concepts, 0.272 consistency, and −2.1 points, on average. These are averages, not wins on every dataset or metric.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.