English

Correcting Stochastic Update Bias in Preconditioned Language Model Optimizers

Machine Learning 2026-05-21 v1 Artificial Intelligence Optimization and Control Machine Learning

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. First, the gradient and preconditioner are typically estimated from the same minibatch, introducing gradient--preconditioner coupling bias. Second, even when the preconditioner estimate is unbiased, its inverse or inverse-root is generally biased because inversion is nonlinear. We propose a single-batch bias-correction framework that addresses both effects: cross-fitted preconditioning estimates the numerator and preconditioner from independent microbatch groups, while variance-corrected inversion uses microbatch variability to subtract the leading delta-method bias term. The framework applies to diagonal moment, diagonal curvature, and matrix preconditioning methods, instantiated in AdamW, Sophia, and Shampoo. Bias correction reduces held-out pretraining loss on Qwen2.5-0.5B by 0.150.15, 0.070.07, and 0.110.11 nats, respectively; the effects on mixed-quality pretraining and downstream instruction tuning are consistently neutral-to-positive. Together, these results establish bias correction as a practical mechanism for reducing finite-sample update bias and improving the performance of preconditioned optimizers.

Keywords

Cite

@article{arxiv.2605.20756,
  title  = {Correcting Stochastic Update Bias in Preconditioned Language Model Optimizers},
  author = {Nikhil Nayak and Julia White and Urchade Zaratiana and Kelton Zhang and Henrijs Princis and Dhruv Atreja and Henry Fawcett and Matthew Thomas and George Hurn-Maloney and Ash Lewis},
  journal= {arXiv preprint arXiv:2605.20756},
  year   = {2026}
}

Comments

32 pages, 3 figures, 13 tables