Reproduce “On the Generation Order in 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

Timofei Gritsaev, Evgeny Katasonov, Ildus Sadrtdinov, Esmeralda S. Whitammer, Dmitry Vetrov

The authors introduce LO-GFlowNet, a training recipe that lets a GFlowNet learn generation orders from rewards without order labels or target training samples. In checksum tasks, the order changes convergence: a fixed prefix-first order converged in 10/10 runs, versus 3/10 for five prefix–suffix swaps. The recipe learns useful orders, but does not outperform the best fixed order in speed.

Summary by alphaXiv.

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