Conditional Poisson sampling gives fixed-size subsets probabilities proportional to the product of their item weights

A post explains conditional Poisson sampling as a way to assign probabilities to fixed-size subsets based on their items’ weights. A reply points to research on using the method to sample from language models. The linked 2021 paper proposes sampling candidates without replacement during beam search; its authors report lower-variance, more efficient estimators than an earlier stochastic beam-search method in their experiments.