Large Language Models Explore by Latent Distilling
Yuanhao Zeng, Ao Lu, Lufei Li, and 3 more authors
In International Conference on Machine Learning, 2026
Exploratory Sampling (ESamp) is a decoding-time method for large language models that encourages semantic diversity during generation. It trains a lightweight Latent Distiller online to predict deep-layer hidden representations from shallow-layer representations; the prediction error acts as a novelty signal for reweighting candidate continuations toward less-explored reasoning paths.