Michael Y. Li

QuasiMoTTo: Quasi-Monte Carlo Test-Time Scaling

Michael Y. Li, Anthony Zhan, Kanishk Gandhi, Noah D. Goodman, Emily B. Fox

Scaling inference compute, by generating many parallel attempts per problem, is a costly but reliable lever for improving language model capabilities. By default these attempts are generated independently, wasting inference compute on redundant solutions. This waste seems unavoid…Read more