Reproduce “Active Flow 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
Yashvir S. Grewal, Daniel M. Steinberg, Thang D. Bui, Cheng Soon Ong, Edwin V. Bonilla
Active Flow Matching (AFM) integrates implicit discrete flow models with variational active generation frameworks by reformulating optimization objectives to utilize conditional endpoint distributions. This method enables the efficient steering of generative models toward high-fitness designs, demonstrating competitive performance across several online protein and small molecule design tasks under tight experimental budgets.
Key insights
- By shifting the optimization objective from intractable marginal likelihoods to tractable conditional endpoint distributions, AFM effectively unifies implicit generative models with rigorous active generation frameworks.
- The Forward-KL AFM variant provides a consistency guarantee (Theorem 3.1), ensuring that optimizing the objective effectively steers the generative model towards sampling from the desired high-fitness distribution.
- A carefully constructed proposal distribution, leveraging self-normalised importance sampling, is crucial for efficient gradient estimation and effective navigation of high-dimensional design spaces with implicit models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.