Why Post-Norm Transformers Collapse: Attention Amplification and Gradient Repair Failure
Xingjian Wang, Qingyu Han, Xiaodong Luo, Yin Zhang
Abstract
Deep decoder-only Transformers often replace the original Post-Norm architecture with Pre-Norm variants because Post-Norm training is highly sensitive to warmup and learning rate under conventional initialization schemes. Although prior work has identified rank collapse and gradient vanishing as related symptoms, it remains poorly understood how causal attention creates high-similarity representations and why training dynamics fail to repair them. We give a two-stage analysis of Post-Norm rank collapse using token similarity as a scalar state variable. First, at initialization, causal attention acts approximately as a prefix-averaging operator that increases token similarity across depth, while the SwiGLU branch contributes only a smaller damping effect. Second, once training enters a high-similarity regime, growth of pre-normalization residual norms makes the RMSNorm backward factor contractive; under mild conditions, gradients to earlier layers decay geometrically. As a complementary result, we characterize the properties of a collapsed network: its best predictor is frequency distribution with relatively high loss floor, and gradients in collapsed layers vanish at frequency distribution. Experiments on 48-layer decoder-only Transformers trained on C4 dataset match the predicted initialization-time similarity growth and collapse-time gradient contraction, and show that collapsed runs stay near the predicted frequency loss. Together, these results distinguish the forward similarity amplification and backward repair incapacity in Post-Norm collapse, while also characterizing the behavior of collapsed networks.
Create a lesson
Related papers
How Model Growth, Recursion, and Boundary Operators Influence Scaling Exponents
Zixi Chen, Akshay Vegesna, Samip Dahal et al.
Evidence-Grounded Agentic Formulation Development in an Autonomous Laboratory
Michael M. Craig, Riley J. Hickman, Yingshan Ma et al.
Probabilistic Linear Explanations
Frederic Koriche, Jean-Marie Lagniez, Chi Tran
Double descent is the principle of least action
Congzhou M Sha
RLLBC-Lib: An Educational Code Library for Reinforcement Learning and Learning-Based Control
Bernd Frauenknecht, Emma Cramer, Artur Eisele et al.
Higher-order pruning of experts in mixture-of-experts language models
Alex M. Tseng, Prannay Kaul, Luca Zancato et al.