arXiv paper: switch optimizers mid-training, save 65%+ of compute

RichmanRonald · x · 2026-08-23

The arXiv paper "Many Optimizers But Only One Training Path" (2608.18810, Ronald Richman and Mario Vüthrich) proposes Repeated Optimizer Resampling (ROR): instead of fixing an optimizer before training, choose it dynamically within a single evolving run.

Method: every $b$ epochs, each candidate optimizer "scouts" from the current weights for $s$ epochs; the best scout continues for the remaining $b-s$ epochs, and a completed segment becomes the new incumbent if the validation objective improves — allowing the preferred optimizer to change as training progresses.

On MNIST, Fashion-MNIST, and two motor insurance claim-count models, with identical seeds comparing 9 fixed optimizers and 2 ROR variants: one-epoch ROR uses only 24%-35% of the aggregate training needed to identify the best fixed optimizer exhaustively while staying close to it on all four tasks.

Original post →

More from Research

Research channel →