Correcting Stochastic Update Bias in Preconditioned Language Model Optimizers
The paper discusses a new framework for correcting biases in preconditioned language model optimizers. It identifies two main biases that arise from the stochastic update rules used in these optimizers. The proposed bias-correction methods show improvements in pretraining loss and overall performance.
- ▪Preconditioned optimizers are essential for training language models but often suffer from stochastic update biases.
- ▪The authors propose a bias-correction framework that addresses gradient-preconditioner coupling and nonlinear inversion biases.
- ▪The framework has been shown to reduce pretraining loss in various models, demonstrating its effectiveness.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Machine Learning arXiv:2605.20756 (cs) [Submitted on 20 May 2026] Title:Correcting Stochastic Update Bias in Preconditioned Language Model Optimizers Authors:Nikhil Nayak, Julia White, Urchade Zaratiana, Kelton Zhang, Henrijs Princis, Dhruv Atreja, Henry Fawcett, Matthew Thomas, George Hurn-Maloney, Ash Lewis View a PDF of the paper titled Correcting Stochastic Update Bias in Preconditioned Language Model Optimizers, by Nikhil Nayak and 9 other authors View PDF HTML (experimental) Abstract:Preconditioned optimizers are central to language model training, but their stochastic update rules are usually treated as direct approximations to population preconditioned descent. We show that this view misses two finite-sample biases.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
