Correcting Stochastic Update Bias in Preconditioned Language Model Optimizers

• URL: http://arxiv.org/abs/2605.20756v1
• Date: Wed, 20 May 2026 05:54:24 GMT
• Title: Correcting Stochastic Update Bias in Preconditioned Language Model Optimizers
• Abstract summary: We show that introducing and preconditioner are typically estimated from the same minibatch.<n>Even when the preconditioner estimate is unbiased, its inverse or inverse-root is generally biased because inversion is nonlinear.<n>We propose a single-batch bias-correction framework that addresses both effects.
• Score: 6.018167862639113
• License: http://creativecommons.org/licenses/by/4.0/
• 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 $0.15$, $0.07$, and $0.11$ 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.

Related papers

• Entropy Minimization without Model Collapse: Mitigating Prediction Bias in Medical Imaging [2.651701169890824]
  Entropy minimization is the dominant objective for test-time adaptation, yet its failure mode, model collapse, remains poorly understood.<n>We propose Distribution Shift Bias Reduction (DSBR), a bias-correcting objective that specifically targets this failure mode.<n>We find that DSBR consistently stabilizes test-time adaptation, prevents model collapse, and matches or outperforms state-of-the-art methods.
  arXiv  Detail & Related papers  (2026-06-01T14:48:16Z)
• Refining Covariance Matrix Estimation in Stochastic Gradient Descent Through Bias Reduction [9.294518380204154]
  We study online inference and covariance estimation for the gradient descent (SGD) algorithm.<n>Our method employs a bias-reduction technique to achieve a convergence rate of $n(-1)/2 sqrtlog n$, outperforming existing Hessian-free alternatives.
  arXiv  Detail & Related papers  (2026-04-23T01:48:08Z)
• Micro-Diffusion Compression - Binary Tree Tweedie Denoising for Online Probability Estimation [0.0]
  We introduce a micro-diffusion denoising layer for improving probability estimates produced by adaptive statistical models.<n>Midicoth combines five fully online components: an adaptive PPM model, a long-range match model, a trie-based word model, a high-order context model, and the micro-diffusion denoiser applied as the final stage.
  arXiv  Detail & Related papers  (2026-03-09T16:59:24Z)
• Preconditioned Score and Flow Matching [15.378063594292955]
  We show that the covariance $_t$ of $p_t$ governs optimization bias.<n>We propose reversible, label-conditional emphpreconditioning maps that reshape the geometry of $p_t$.<n>We show that preconditioning consistently yields better-trained models by avoiding suboptimal plateaus.
  arXiv  Detail & Related papers  (2026-03-02T19:09:15Z)
• Unregularized Linear Convergence in Zero-Sum Game from Preference Feedback [50.89125374999765]
  We provide the first convergence guarantee for Optimistic Multiplicative Weights Update ($mathtOMWU$) in NLHF.<n>Our analysis identifies a novel marginal convergence behavior, where the probability of rarely played actions grows exponentially from exponentially small values.
  arXiv  Detail & Related papers  (2025-12-31T12:08:29Z)
• On the Optimal Construction of Unbiased Gradient Estimators for Zeroth-Order Optimization [57.179679246370114]
  A potential limitation of existing methods is the bias inherent in most perturbation estimators unless a stepsize is proposed.<n>We propose a novel family of unbiased gradient scaling estimators that eliminate bias while maintaining favorable construction.
  arXiv  Detail & Related papers  (2025-10-22T18:25:43Z)
• From Noisy Traces to Stable Gradients: Bias-Variance Optimized Preference Optimization for Aligning Large Reasoning Models [90.45197506653341]
  Large reasoning models generate intermediate reasoning traces before producing final answers.<n> aligning LRMs with human preferences, a crucial prerequisite for model deployment, remains underexplored.<n>A common workaround optimized a single sampled trajectory, which introduces substantial gradient variance from trace sampling.
  arXiv  Detail & Related papers  (2025-10-06T17:58:01Z)
• The Implicit Bias of Batch Normalization in Linear Models and Two-layer Linear Convolutional Neural Networks [117.93273337740442]
  We show that gradient descent converges to a uniform margin classifier on the training data with an $exp(-Omega(log2 t))$ convergence rate. We also show that batch normalization has an implicit bias towards a patch-wise uniform margin.
  arXiv  Detail & Related papers  (2023-06-20T16:58:00Z)
• Debiased Fine-Tuning for Vision-language Models by Prompt Regularization [56.48290708901531]
  We present a new paradigm for fine-tuning large-scale vision pre-trained models on downstream task, dubbed Prompt Regularization (ProReg)<n>ProReg uses the prediction by prompting the pretrained model to regularize the fine-tuning.<n>We show the consistently strong performance of ProReg compared with conventional fine-tuning, zero-shot prompt, prompt tuning, and other state-of-the-art methods.
  arXiv  Detail & Related papers  (2023-01-29T11:53:55Z)
• Ensembling over Classifiers: a Bias-Variance Perspective [13.006468721874372]
  We build upon the extension to the bias-variance decomposition by Pfau (2013) in order to gain crucial insights into the behavior of ensembles of classifiers. We show that conditional estimates necessarily incur an irreducible error. Empirically, standard ensembling reducesthe bias, leading us to hypothesize that ensembles of classifiers may perform well in part because of this unexpected reduction.
  arXiv  Detail & Related papers  (2022-06-21T17:46:35Z)
• Sample-Efficient Optimisation with Probabilistic Transformer Surrogates [66.98962321504085]
  This paper investigates the feasibility of employing state-of-the-art probabilistic transformers in Bayesian optimisation. We observe two drawbacks stemming from their training procedure and loss definition, hindering their direct deployment as proxies in black-box optimisation. We introduce two components: 1) a BO-tailored training prior supporting non-uniformly distributed points, and 2) a novel approximate posterior regulariser trading-off accuracy and input sensitivity to filter favourable stationary points for improved predictive performance.
  arXiv  Detail & Related papers  (2022-05-27T11:13:17Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.