Output Embedding Centering for Stable LLM Pretraining

• URL: http://arxiv.org/abs/2601.02031v1
• Date: Mon, 05 Jan 2026 11:44:05 GMT
• Title: Output Embedding Centering for Stable LLM Pretraining
• Abstract summary: A specific instability that often occurs for large learning rates at the end of training is output logit divergence.<n>We propose output embedding centering (OEC) as a new mitigation strategy.<n>OEC can be implemented in two different ways, as a deterministic operation called -centering, or a regularization method called -loss.
• Score: 3.0718163016168067
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Pretraining of large language models is not only expensive but also prone to certain training instabilities. A specific instability that often occurs for large learning rates at the end of training is output logit divergence. The most widely used mitigation strategy, z-loss, merely addresses the symptoms rather than the underlying cause of the problem. In this paper, we analyze the instability from the perspective of the output embeddings' geometry and identify its cause. Based on this, we propose output embedding centering (OEC) as a new mitigation strategy, and prove that it suppresses output logit divergence. OEC can be implemented in two different ways, as a deterministic operation called μ-centering, or a regularization method called μ-loss. Our experiments show that both variants outperform z-loss in terms of training stability and learning rate sensitivity. In particular, they ensure that training converges even for large learning rates when z-loss fails. Furthermore, we find that μ-loss is significantly less sensitive to regularization hyperparameter tuning than z-loss.

Related papers

• Demystifying Variance in Circuit Discovery of LLMs [51.97489679448601]
  We develop a new circuit discovery method that substantially lessens resampling variance.<n>We show that rephrasing variance arises because prompts with different templates tend to activate different circuits in the model.<n>We argue that sparsity, which has been claimed to form more compact and interpretable task circuits, fails to solve this problem.
  arXiv  Detail & Related papers  (2026-06-15T16:25:07Z)
• Beyond Precision: Training-Inference Mismatch is an Optimization Problem and Simple LR Scheduling Fixes It [24.70923739848818]
  We show that gradient noise and training-inference mismatch escalate in tandem as training progresses.<n>We find that the mismatch can be effectively suppressed by shrinking the update size.<n>We propose a simple yet effective solution: a specialized Learning Rate scheduler.
  arXiv  Detail & Related papers  (2026-02-02T09:00:53Z)
• Memorize Early, Then Query: Inlier-Memorization-Guided Active Outlier Detection [5.34972604411447]
  Inlier-memorization (IM) effect provides a promising signal for distinguishing outliers.<n>Existing unsupervised approaches that rely solely on the IM effect still struggle when inliers and outliers are not well-separated.<n>We propose IMBoost, a novel framework that explicitly reinforces the IM effect to improve outlier detection.
  arXiv  Detail & Related papers  (2026-01-16T04:55:46Z)
• Exploring Structural Degradation in Dense Representations for Self-supervised Learning [84.52554180480037]
  We observe a counterintuitive phenomenon in self-supervised learning (SSL): longer training may impair the performance of dense prediction tasks.<n>We refer to this phenomenon as Self-supervised Dense Degradation (SDD) and demonstrate its consistent presence across sixteen state-of-the-art SSL methods.<n>We introduce a Dense representation Structure Estimator (DSE) composed of a class-relevance measure and an effective dimensionality measure.
  arXiv  Detail & Related papers  (2025-10-20T08:40:16Z)
• Efficient local linearity regularization to overcome catastrophic overfitting [59.463867084204566]
  Catastrophic overfitting (CO) in single-step adversarial training results in abrupt drops in the adversarial test accuracy (even down to 0%) We introduce a regularization term, called ELLE, to mitigate CO effectively and efficiently in classical AT evaluations.
  arXiv  Detail & Related papers  (2024-01-21T22:55:26Z)
• Learning Prompt-Enhanced Context Features for Weakly-Supervised Video Anomaly Detection [37.99031842449251]
  Video anomaly detection under weak supervision presents significant challenges. We present a weakly supervised anomaly detection framework that focuses on efficient context modeling and enhanced semantic discriminability. Our approach significantly improves the detection accuracy of certain anomaly sub-classes, underscoring its practical value and efficacy.
  arXiv  Detail & Related papers  (2023-06-26T06:45:16Z)
• Leveraging Heteroscedastic Uncertainty in Learning Complex Spectral Mapping for Single-channel Speech Enhancement [20.823177372464414]
  Most speech enhancement (SE) models learn a point estimate, and do not make use of uncertainty estimation in the learning process. We show that modeling heteroscedastic uncertainty by minimizing a multivariate Gaussian negative log-likelihood (NLL) improves SE performance at no extra cost.
  arXiv  Detail & Related papers  (2022-11-16T02:29:05Z)
• MaxMatch: Semi-Supervised Learning with Worst-Case Consistency [149.03760479533855]
  We propose a worst-case consistency regularization technique for semi-supervised learning (SSL) We present a generalization bound for SSL consisting of the empirical loss terms observed on labeled and unlabeled training data separately. Motivated by this bound, we derive an SSL objective that minimizes the largest inconsistency between an original unlabeled sample and its multiple augmented variants.
  arXiv  Detail & Related papers  (2022-09-26T12:04:49Z)
• Scale-Equivalent Distillation for Semi-Supervised Object Detection [57.59525453301374]
  Recent Semi-Supervised Object Detection (SS-OD) methods are mainly based on self-training, generating hard pseudo-labels by a teacher model on unlabeled data as supervisory signals. We analyze the challenges these methods meet with the empirical experiment results. We introduce a novel approach, Scale-Equivalent Distillation (SED), which is a simple yet effective end-to-end knowledge distillation framework robust to large object size variance and class imbalance.
  arXiv  Detail & Related papers  (2022-03-23T07:33:37Z)
• A Simple but Tough-to-Beat Data Augmentation Approach for Natural Language Understanding and Generation [53.8171136907856]
  We introduce a set of simple yet effective data augmentation strategies dubbed cutoff. cutoff relies on sampling consistency and thus adds little computational overhead. cutoff consistently outperforms adversarial training and achieves state-of-the-art results on the IWSLT2014 German-English dataset.
  arXiv  Detail & Related papers  (2020-09-29T07:08:35Z)
• Unbiased Risk Estimators Can Mislead: A Case Study of Learning with Complementary Labels [92.98756432746482]
  We study a weakly supervised problem called learning with complementary labels. We show that the quality of gradient estimation matters more in risk minimization. We propose a novel surrogate complementary loss(SCL) framework that trades zero bias with reduced variance.
  arXiv  Detail & Related papers  (2020-07-05T04:19:37Z)

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.