Reproduce “Weighted Adjoint Matching: Finetuning Flow Matching Models with Time-Dependent Control Cost” 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
Eliot Beyler, Francis Bach
Weighted Adjoint Matching makes early changes to generation more costly and later changes less costly. Its objective uses KL penalties on generation prefixes. On unconditional MNIST, it reaches reward 0.180, with 10.5% zeros and 9.1% ones, versus Adjoint Matching’s 0.176, 18.0% zeros, and 3.4% ones; this supports a better tradeoff in this setup, not universal preservation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.