Training Dynamics Underlying Language Model Scaling Laws: Loss Deceleration and Zero-Sum Learning

• URL: http://arxiv.org/abs/2506.05447v2
• Date: Mon, 14 Jul 2025 23:29:38 GMT
• Title: Training Dynamics Underlying Language Model Scaling Laws: Loss Deceleration and Zero-Sum Learning
• Authors: Andrei Mircea, Supriyo Chakraborty, Nima Chitsazan, Milind Naphade, Sambit Sahu, Irina Rish, Ekaterina Lobacheva,
• Abstract summary: This work aims to understand how scaling improves language models, specifically in terms of training dynamics.<n>We find that language models undergo loss deceleration early in training; an abrupt slowdown in the rate of loss improvement, resulting in piecewise linear behaviour of the loss curve in log-log space.<n>We attribute loss deceleration to a type of training dynamics we term zero-sum learning (ZSL)<n>In ZSL, per-example gradients become systematically opposed, leading to destructive interference in per-example changes in loss.
• Score: 14.227982314368116
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: This work aims to understand how scaling improves language models, specifically in terms of training dynamics. We find that language models undergo loss deceleration early in training; an abrupt slowdown in the rate of loss improvement, resulting in piecewise linear behaviour of the loss curve in log-log space. Scaling up the model mitigates this transition by (1) decreasing the loss at which deceleration occurs, and (2) improving the log-log rate of loss improvement after deceleration. We attribute loss deceleration to a type of degenerate training dynamics we term zero-sum learning (ZSL). In ZSL, per-example gradients become systematically opposed, leading to destructive interference in per-example changes in loss. As a result, improving loss on one subset of examples degrades it on another, bottlenecking overall progress. Loss deceleration and ZSL provide new insights into the training dynamics underlying language model scaling laws, and could potentially be targeted directly to improve language models independent of scale. We make our code and artefacts available at: https://github.com/mirandrom/zsl

Related papers

• Scaling Collapse Reveals Universal Dynamics in Compute-Optimally Trained Neural Networks [59.552873049024775]
  We show that compute-optimally trained models exhibit a remarkably precise universality.<n>With learning rate decay, the collapse becomes so tight that differences in the normalized curves across models fall below the noise floor.<n>We explain these phenomena by connecting collapse to the power-law structure in typical neural scaling laws.
  arXiv  Detail & Related papers  (2025-07-02T20:03:34Z)
• Dynamical loss functions shape landscape topography and improve learning in artificial neural networks [0.9208007322096533]
  We show how to transform cross-entropy and mean squared error into dynamical loss functions. We show how they significantly improve validation accuracy for networks of varying sizes.
  arXiv  Detail & Related papers  (2024-10-14T16:27:03Z)
• To Clip or not to Clip: the Dynamics of SGD with Gradient Clipping in High-Dimensions [6.653325043862049]
  We study gradient clipping in a least squares problem under streaming SGD. We show that with Gaussian noise clipping cannot improve SGD performance. We propose a simple for near optimal scheduling of the clipping threshold.
  arXiv  Detail & Related papers  (2024-06-17T16:50:22Z)
• Heavy-Tailed Class Imbalance and Why Adam Outperforms Gradient Descent on Language Models [23.520679217713685]
  Adam has been shown to outperform gradient descent on large language models by a larger margin than on other tasks. We show that a key factor in this performance gap is the heavy-tailed class imbalance found in language tasks.
  arXiv  Detail & Related papers  (2024-02-29T18:47:52Z)
• A Dynamical Model of Neural Scaling Laws [79.59705237659547]
  We analyze a random feature model trained with gradient descent as a solvable model of network training and generalization. Our theory shows how the gap between training and test loss can gradually build up over time due to repeated reuse of data.
  arXiv  Detail & Related papers  (2024-02-02T01:41:38Z)
• Dropout Reduces Underfitting [85.61466286688385]
  In this study, we demonstrate that dropout can also mitigate underfitting when used at the start of training. We find dropout reduces the directional variance of gradients across mini-batches and helps align the mini-batch gradients with the entire dataset's gradient. Our findings lead us to a solution for improving performance in underfitting models - early dropout: dropout is applied only during the initial phases of training, and turned off afterwards.
  arXiv  Detail & Related papers  (2023-03-02T18:59:15Z)
• Same Pre-training Loss, Better Downstream: Implicit Bias Matters for Language Models [46.24479693469042]
  This paper shows that 1) pre-training loss cannot fully explain downstream performance and 2) flatness of the model is well-correlated with downstream performance where pre-training loss is not.
  arXiv  Detail & Related papers  (2022-10-25T17:45:36Z)
• Neural Collapse Inspired Attraction-Repulsion-Balanced Loss for Imbalanced Learning [97.81549071978789]
  We propose Attraction-Repulsion-Balanced Loss (ARB-Loss) to balance the different components of the gradients. We perform experiments on the large-scale classification and segmentation datasets and our ARB-Loss can achieve state-of-the-art performance.
  arXiv  Detail & Related papers  (2022-04-19T08:23:23Z)
• A Mathematical Analysis of Learning Loss for Active Learning in Regression [2.792030485253753]
  This paper develops a foundation for Learning Loss which enables us to propose a novel modification we call LearningLoss++. We show that gradients are crucial in interpreting how Learning Loss works, with rigorous analysis and comparison of the gradients between Learning Loss and LearningLoss++. We also propose a convolutional architecture that combines features at different scales to predict the loss. We show that LearningLoss++ outperforms in identifying scenarios where the model is likely to perform poorly, which on model refinement translates into reliable performance in the open world.
  arXiv  Detail & Related papers  (2021-04-19T13:54:20Z)
• Implicit Bias in Deep Linear Classification: Initialization Scale vs Training Accuracy [71.25689267025244]
  We show how the transition is controlled by the relationship between the scale and how accurately we minimize the training loss. Our results indicate that some limit behaviors of gradient descent only kick in at ridiculous training accuracies.
  arXiv  Detail & Related papers  (2020-07-13T23:49:53Z)

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.