Reproduce “Tilted Schrödinger Bridge Matching” 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
Sergei Kholkin, Evgeny Burnaev, Alexander Korotin
TSBM adapts a pretrained Schrödinger bridge to steer its target distribution with an exponential reward tilt while retaining the source constraint. It uses the pretrained bridge as a reference and learns an incremental controller plus a correction for input–output dependence. On CelebA, it improves reward-faithfulness trade-offs; PickScore reaches 18.2345 versus 17.6109 for the baseline, while exact source-marginal preservation remains unproven.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.