Starting from coarse topic descriptions, the model generates questions and associated answers or reasoning information.
Models improve by training on questions and answers they generate
Sick and colleagues report gains in language and visual reasoning using self-generated supervision rather than human-curated training questions or reference answers.
Big Tech
Leon Sick · Lukas Hoyer · Yannick Strümpler · Christian Reisswig · Timo Ropinski · Federico Tombari
Google · Ulm University
Research Digest··2 min read
Sick and colleagues introduce On-Policy Self-Training, a method in which a model generates its own training problems and learns from a teacher version of itself given extra context.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.