Reproduce “Spherical Flows for Sampling Categorical Data” 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
Jannis Chemseddine, Gregor Kornhardt, Gabriele Steidl
The authors propose a sequence-generation method that places token embeddings on a sphere and corrupts them with von Mises–Fisher noise. A learned token-posterior model supplies both transport velocities and scores for predictor–corrector sampling, which updates sequence positions together. Trained on TinyGSM, the self-conditioned model reaches 43.5% accuracy on GSM8K with predictor–corrector sampling at NFE 1024, below the 63.3% greedy autoregressive reference.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.