BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CAIL//Events//EN
CALSCALE:GREGORIAN
METHOD:PUBLISH
X-WR-CALNAME:CAIL Events
BEGIN:VEVENT
UID:2026-10-26-hanlin-zhang@cail.columbia.edu
DTSTAMP:20261018T204625Z
DTSTART:20261026T140000Z
DTEND:20261026T150000Z
SUMMARY:ML Seminar: Hanlin Zhang - Less is More: Bypassing Some Empirical 
 Overhead in Language Modeling
LOCATION:Mudd 417 conference room
DESCRIPTION: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 pro
 xy training runs\, reward models are fit to preference labels bought anew 
 for every domain\, and capability is discovered by trial and error. By tre
 ating all three as measurement problems\, we enable better understanding a
 nd engineering: the Data Manifold Oracle reads a corpus’s scaling-law fl
 oor and rate off raw text by compression\, with no gradient steps on the c
 andidate corpus\; unsupervised reward modeling recovers a useful preferenc
 e signal from unlabeled math-web continuations at zero additional preferen
 ce-annotation cost\; and prescriptive scaling laws fit the attainable-perf
 ormance boundary traced out by the public model ecosystem\, estimating upp
 er-tail performance for a given budget while validating calibration on lat
 er model generations. Less is more in one specific sense: measure first\, 
 train less\, and know what each measurement can certify.\n\nhttps://cail.c
 olumbia.edu/events/2026-10-26-hanlin-zhang
URL:https://cail.columbia.edu/events/2026-10-26-hanlin-zhang
END:VEVENT
END:VCALENDAR