Reproduce “Online Class Posteriors for Interpretable and Controllable Discrete Diffusion” 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
Felix Koulischer, Gianni Van De Velde, François Remy, Chris Develder, Thomas Demeester
The authors track class likelihoods in discrete diffusion by accumulating differences between class-conditioned and unconditional transition predictions after token changes. This running posterior links words such as “wine” to class evidence and weights NOT guidance by probability. In an imbalanced antibody test, one NOT-gate setting trades violation rates of 0.002 and 0.016 for FPD values of 0.326 and 0.010; updates follow token changes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.