ESamp: LLMs explore by latent distilling for semantic-novelty sampling
ShanghaiTech University
ESamp is a decoding method that injects semantic (not just lexical) diversity by training a lightweight Distiller at test time to predict deeper-layer hidden states from shallow ones, then using prediction errors as a novelty signal to bias sampling toward less-explored semantic patterns. Reports improved Pass@k on math, science, and code benchmarks with only 1.2-5% inference overhead.
Why it matters
Tackles a long-standing weakness of temperature/top-p sampling — stochastic decoding rarely produces genuinely different reasoning paths. A semantic-novelty signal that breaks the diversity-coherence tradeoff is directly relevant to test-time scaling and self-consistency methods.
Importance: 2/5
Solid method paper, default importance.