Reproduce “Empowering Masked Diffusion Models to Self-Correct with Leave-One-Out Transformers” 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
Sofian Zalouk, Vincent Counathe, Paul Jünger, Daniel Cao, Daniel Zlotnick, Kilian Q. Weinberger, Christopher De Sa
LOOT predicts each token from surrounding tokens without using its own input at that position. This supports training on all positions and proposing replacements for filled tokens. On unconditional OpenWebText generation, Blocked Gibbs correction reaches MAUVE 0.408 at 256 forward evaluations, versus ReMDM’s 0.348 at 2,048, with matched sample entropy. This comparison uses 3.5× fewer inference FLOPs than ReMDM, not 8× fewer compute.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.