Reproduce “Stochastic Beams and Where to Find Them: The Gumbel-Top-k Trick for Sampling Sequences Without Replacement” 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
Wouter Kool, Herke van Hoof, Max Welling
The authors introduce Stochastic Beam Search, adding coordinated Gumbel noise to sequence probabilities and expanding the highest-scoring prefixes. For locally normalized, factorized sequence models, it produces exact samples without replacement without enumerating all sequences. In tested English-to-French translation, it has higher mean and maximum BLEU than Diverse Beam Search at matched diversity.
Key insights
- The Gumbel-Max trick can be generalized to draw multiple ordered samples without replacement (Gumbel-Top-k trick) from a categorical distribution.
- This generalized trick can be efficiently applied to factorized sequence distributions, providing a principled method for stochastic sequence generation without explicit enumeration.
- Stochastic Beam Search establishes a clear probabilistic connection between random sampling and deterministic beam search, combining advantages of both for sequence generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.