Reproduce “Local Search GFlowNets” 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

Minsu Kim, Taeyoung Yun, Emmanuel Bengio, Dinghuai Zhang, Yoshua Bengio, Sungsoo Ahn, Jinkyoo Park

LS-GFN augments GFlowNet training with policy-guided revisions that undo part of a molecule or biological sequence and rebuild it. The method retains from-scratch candidate generation, and local search is not used at inference. On QM9, TB + LS-GFN finds 793 ± 4 modes versus 699 ± 14 for TB across three seeds; gains are not universal.

Key insights

  • A synergistic combination of GFlowNets' global, inter-mode exploration and local search's intra-mode exploitation significantly improves sampling efficiency and quality.
  • Integrating local search allows GFlowNets to focus more on exploiting high-reward regions, addressing the "over-exploration" problem while maintaining diversity.
  • The iterative process enables a positive feedback loop where improved GFlowNet policies enhance local search effectiveness, and vice-versa.

Summary by alphaXiv.

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