Hanlin Zhang

Language-model pipelines tend to rely on brute force to answer three fundamental questions: what to train on, what signal to align with, and what a compute budget will buy. Candidate corpora are ranked by proxy training runs, reward models are fit to preference labels bought anew for every domain, and capability is discovered by trial and error. By treating all three as measurement problems, we enable better understanding and engineering: the Data Manifold Oracle reads a corpus’s scaling-law floor and rate off raw text by compression, with no gradient steps on the candidate corpus; unsupervised reward modeling recovers a useful preference signal from unlabeled math-web continuations at zero additional preference-annotation cost; and prescriptive scaling laws fit the attainable-performance boundary traced out by the public model ecosystem, estimating upper-tail performance for a given budget while validating calibration on later model generations. Less is more in one specific sense: measure first, train less, and know what each measurement can certify.