RATIONALYST

Process-supervision for reasoning via pretraining: mines 79K implicit rationales from unlabeled web text and the Pile, then trains a model that supervises another LLM's chain-of-thought step by step at inference. Improves accuracy by an average ~3.9% across seven reasoning benchmarks (math, commonsense, scientific, logical), including GSM8K and MATH. JHU (Jiang, Khashabi, Van Durme et al.).