Reproduce “Active Flow Expansion for Out-of-Distribution Discovery: from Theory to Molecules” 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

Riccardo De Santi, Bruce Lee, Cristian Perez Jensen, Kimon Protopapas, Sophia Tang, Cheng-Hao Liu, Pranam Chatterjee, Yisong Yue and 1 more

ActFlow expands a model’s generable set of valid designs by using verifier feedback to sample uncertain frontier regions while staying near its current distribution, then updates the model with verified samples. On QM9, FlowMol Gaussian reaches 88.40 valid molecular clusters, versus 45.40 for REC-F and 42.80 for REC-NF, with 95.90% validity; this demonstrates broader measured coverage, not real-world discoveries.

Key insights

  • Scientific discovery necessitates "generable set expansion" (covering the valid design space) rather than merely matching observed data distributions.
  • Active exploration, guided by verifier uncertainty and regularized by the current generative model, efficiently discovers sparse, frontier valid regions beyond initial data.
  • A "signed" model update, incorporating both positive (valid) and negative (invalid) feedback, effectively reallocates model mass toward newly verified valid regions and away from invalid ones.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.