Large Language Models Explore by Latent Distilling

Yuanhao Zeng, Ao Lu, Lufei Li, Zheng Zhang, Yexin Li, Kan Ren · 2026

Read Large Language Models Explore by Latent Distilling by Yuanhao Zeng, Ao Lu, Lufei Li, Zheng Zhang, Yexin Li, Kan Ren (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-24927.txt.

Description

Zeng and five colleagues propose Exploratory Sampling, a decoding approach that seeks varied meanings while making text. They train a small model at test time to predict deep states from shallow states. Guess error serves as a novelty sign to reweight token choices, biasing decoding toward fresh meanings. ESamp runs with less than 5 percent overhead and boosts model scores, breaking the variety-clarity trade-off in creative writing.

Catalog record

CREATORYuanhao Zeng, Ao Lu, Lufei Li, Zheng Zhang, Yexin Li, Kan Ren
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE77900
LICENSEhttp://creativecommons.org/licenses/by-sa/4.0/
RIGHTSCC BY SA
SOURCEarXiv

Source: https://arxiv.org/abs/2604.24927

Browse the full library · About the Tower of Babel