Reproduce “ENCORE: Exact Non-equilibrium COntrol with Replica Exchange for Diffusion Generation” 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
Jiahao Yu, Saifuddin Syed, José Miguel Hernández-Lobato, Jiajun He
ENCORE controls a pretrained sampler by reweighting its output distribution without retraining. It stores partial generation trajectories so upward replica exchanges truncate history rather than simulate an unavailable reverse transition. Under stated support assumptions, the sampler-specific target is stationary. On a perturbed one-dimensional smooth-reward task, weighted empirical Wasserstein-2 distance was 0.015 ± 0.0069, versus CREPE’s 0.650 ± 0.0331; this does not show correction of arbitrary generator misspecification.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.