Reproduce “BridgeTwist: Twisting Schrödinger Bridges for Training-Free Conditional Sampling” 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
Mengyu Li, Qianqian Qu, Jun S. Liu
BridgeTwist enables training-free conditional sampling with pretrained Schrödinger bridges (SBs). It combines learned velocity and score to predict endpoint states, then evaluates task likelihoods to steer and reweight particle paths. On CIFAR-10, its 64-particle sampler reaches external class accuracy 0.979 and KID 0.0403; convergence is to the posterior of the learned, discretized prior, not exact sampling from real data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.