Reproduce “Plug-and-Play Guidance for Discrete Diffusion Models via Gradient-Informed Logit Correction” 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
Hongkun Dou, Zike Chen, Fengji Li, Hongjue Li, Yue Deng
GILC guides pretrained discrete diffusion models toward reward goals by correcting clean-token predictions during sampling, without retraining the generator. Its differentiable and black-box variants estimate reward from proxy samples and adjust logits, omitting the denoiser Jacobian. On HepG2 DNA design, GILC-DB reached predicted activity 7.04 ± 0.26 and ATAC-Acc 95.2 ± 2.1%, above fine-tuned DRAKES; evaluations were computational, not laboratory tests.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.