Reproduce “Truncation Sampling as Language Model Desmoothing” 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
John Hewitt, Christopher D. Manning, Percy Liang
Researchers at Stanford University propose a theoretical framework that reinterprets truncation sampling in language models as 'desmoothing,' aiming to recover the true underlying distribution. Their new "eta-sampling" algorithm, derived from this theory, achieved a 53% human preference rate over top-p sampling for long-document plausibility and demonstrated lower repetition rates in generated text.
Key insights
- Neural language models 'smooth' their output distributions due to KL-divergence training objectives, assigning small probabilities across a vast vocabulary and obscuring the 'true' underlying language distribution.
- The primary goal of truncation sampling should be to 'desmooth' these distributions, effectively recovering the support (set of plausible words) of the true language distribution.
- Existing methods like top-p sampling can aggressively over-truncate in low-entropy contexts, leading to repetition, or under-truncate in high-entropy contexts, allowing implausible continuations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.