Exploratory Sampling Boosts Semantic Diversity in LLMs

Researchers have introduced Exploratory Sampling (ESamp), a decoding technique designed for large language models that actively promotes semantic diversity in text generation. Traditional stochastic sampling results in minimal lexical variation, which restricts exploration. ESamp employs a lightweight Distiller during testing to forecast deep-layer hidden representations based on shallow-layer data, effectively modeling transitions in representation depth. This Distiller adjusts according to the current generation context, utilizing prediction error as a signal for novelty to modify token probabilities. This method is based on the insight that neural networks tend to make more accurate predictions on familiar inputs while struggling with novel ones, ultimately aiming to enhance test-time scaling for richer semantic exploration.