Ekaterina Lobacheva

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

Supriyo Chakraborty

Nima Chitsazan

This work aims to understand how scaling improves language models, specifically in terms of training dynamics. We find that language models … (voir plus)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

2025-06-05

ArXiv (prépublication)

arxiv.org

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

Supriyo Chakraborty

Nima Chitsazan

This work aims to understand how scaling improves language models, specifically in terms of training dynamics. We find that language models … (voir plus)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

2025-01-01

ACL (1) (publié)

doi.org

arxiv.org

How Learning Rates Shape Neural Network Focus: Insights from Example Ranking

Keller Jordan

Aristide Baratin

Nicolas Le Roux

The learning rate is a key hyperparameter that affects both the speed of training and the generalization performance of neural networks. Th… (voir plus)rough a new {\it loss-based example ranking} analysis, we show that networks trained with different learning rates focus their capacity on different parts of the data distribution, leading to solutions with different generalization properties. These findings, which hold across architectures and datasets, provide new insights into how learning rates affect model performance and example-level dynamics in neural networks.

2024-10-10

NeurIPS.cc/2024/Workshop/SciForDL (poster)

Language model scaling laws and zero-sum learning

Supriyo Chakraborty

Nima Chitsazan

This work aims to understand how, in terms of training dynamics, scaling up language model size yields predictable loss improvements. We fin… (voir plus)d that these improvements can be tied back to loss deceleration, an abrupt transition in the rate of loss improvement, characterized by piece-wise linear behavior in log-log space. Notably, improvements from increased model size appear to be a result of (1) improving the loss at which this transition occurs; and (2) improving the rate of loss improvement after this transition. As an explanation for the mechanism underlying this transition (and the effect of model size on loss it mediates), we propose the zero-sum learning (ZSL) hypothesis. In ZSL, per-token gradients become systematically opposed, leading to degenerate training dynamics where the model can't improve loss on one token without harming it on another; bottlenecking the overall rate at which loss can improve. We find compelling evidence of ZSL, as well as unexpected results which shed light on other factors contributing to ZSL.

2024-10-10

NeurIPS.cc/2024/Workshop/SciForDL (poster)

Gradient Dissent in Language Model Training and Saturation

We seek to shed light on language model (LM) saturation from the perspective of learning dynamics. To this end, we define a decomposition o… (voir plus)f the cross-entropy gradient, which forms a shared low-dimensional basis for analyzing the training dynamics of models across scales. Intuitively, this decomposition consists of attractive and repulsive components that increase the logit of the correct class and decrease the logits of incorrect classes, respectively. Our analysis in this subspace reveals a phenomenon we term \textit{gradient dissent}, characterized by gradient components becoming systematically opposed such that loss cannot be improved along one component without being degraded along the other. Notably, we find that complete opposition, which we term \textit{total dissent}, reliably occurs in tandem with the saturation of smaller LMs. Based on these results, we hypothesize that gradient dissent can provide a useful foundation for better understanding and mitigating saturation.

2024-06-16

ICML.cc/2024/Workshop/HiLD (poster)