Decoding (how a language model picks the next word) isnβt a bag of tricks; itβs a clean optimisation problem over probabilities.
This paper builds a Google-for-theorems: a semantic search engine that finds exact theorems, lemmas, and propositions instead of just entire papers.