PreSTO preprint reports faster sampling for language-model reasoning

Researchers report that PreSTO speeds up a sampling method used to improve language-model reasoning while preserving its underlying acceptance rules. Across seven datasets and five base models, the preprint reports median speedups of 1.71× for PowerMH, 1.91× for EntropyCut and 2.33× for MultiTryMH. PreSTO batches possible future proposals, then follows the sequence selected by the original sampler. For the first two methods, it used a median of 3.9× fewer model calls, with each call taking 2.1× longer.