Revisiting Anisotropy in Language Transformers: The Geometry of Learning Dynamics

• URL: http://arxiv.org/abs/2604.08764v1
• Date: Thu, 09 Apr 2026 21:02:20 GMT
• Title: Revisiting Anisotropy in Language Transformers: The Geometry of Learning Dynamics
• Abstract summary: Transformer architectures have dominated Natural Language Processing (NLP)<n>Recent research has highlighted an inherent anisotropy phenomenon in these models.<n>We use concept-based mechanistic interpretability during training, rather than only post hoc, to fit activation-derived low-rank tangent proxies.<n>We find that these activation-derived directions capture both unusually large gradient energy and a substantially larger share of gradient anisotropy than matched-rank normal controls.
• Score: 5.987889368693271
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Since their introduction, Transformer architectures have dominated Natural Language Processing (NLP). However, recent research has highlighted an inherent anisotropy phenomenon in these models, presenting a significant challenge to their geometric interpretation. Previous theoretical studies on this phenomenon are rarely grounded in the underlying representation geometry. In this paper, we extend them by deriving geometric arguments for how frequency-biased sampling attenuates curvature visibility and why training preferentially amplify tangent directions. Empirically, we then use concept-based mechanistic interpretability during training, rather than only post hoc, to fit activation-derived low-rank tangent proxies and test them against ordinary backpropagated true gradients. Across encoder-style and decoder-style language models, we find that these activation-derived directions capture both unusually large gradient energy and a substantially larger share of gradient anisotropy than matched-rank normal controls, providing strong empirical support for a tangent-aligned account of anisotropy.

Related papers

• The Inductive Bias of Convolutional Neural Networks: Locality and Weight Sharing Reshape Implicit Regularization [57.37943479039033]
  We study how architectural inductive bias reshapes the implicit regularization induced by the edge-of-stability phenomenon in gradient descent.<n>We show that locality and weight sharing fundamentally change this picture.
  arXiv  Detail & Related papers  (2026-03-05T04:50:51Z)
• Truth as a Trajectory: What Internal Representations Reveal About Large Language Model Reasoning [19.48395840455837]
  We introduce Truth as a Trajectory (TaT) for Large Language Models.<n>TaT shifts analysis from static activations to layer-wise geometric displacement.<n>We show that TaT effectively mitigates reliance on static lexical confounds, outperforming conventional probing.
  arXiv  Detail & Related papers  (2026-03-01T23:52:10Z)
• Scale-Consistent State-Space Dynamics via Fractal of Stationary Transformations [9.983526161001997]
  Recent deep learning models increasingly rely on depth without structural guarantees on the validity of intermediate representations.<n>We address this limitation by formulating a structural requirement for state-space model's scale-consistent latent dynamics.<n>We empirically verify the predicted scale-consistent behavior, showing that adaptive efficiency emerges from the aligned latent geometry.
  arXiv  Detail & Related papers  (2026-01-27T12:44:20Z)
• Tracing the Representation Geometry of Language Models from Pretraining to Post-training [22.18942718274405]
  We take a spectral approach to investigate the geometry of learned representations across pretraining and post-training.<n>We uncover a consistent non-monotonic sequence of three geometric phases during autoregressive pretraining.<n>Post-training further transforms geometry: SFT and DPO drive "entropy-seeking" dynamics to integrate specific instructional or preferential data.
  arXiv  Detail & Related papers  (2025-09-27T00:46:29Z)
• Generative Model Inversion Through the Lens of the Manifold Hypothesis [98.37040155914595]
  Model inversion attacks (MIAs) aim to reconstruct class-representative samples from trained models.<n>Recent generative MIAs utilize generative adversarial networks to learn image priors that guide the inversion process.
  arXiv  Detail & Related papers  (2025-09-24T14:39:25Z)
• A U-turn on Double Descent: Rethinking Parameter Counting in Statistical Learning [68.76846801719095]
  We show that double descent appears exactly when and where it occurs, and that its location is not inherently tied to the threshold p=n. This provides a resolution to tensions between double descent and statistical intuition.
  arXiv  Detail & Related papers  (2023-10-29T12:05:39Z)
• Towards Training Without Depth Limits: Batch Normalization Without Gradient Explosion [83.90492831583997]
  We show that a batch-normalized network can keep the optimal signal propagation properties, but avoid exploding gradients in depth. We use a Multi-Layer Perceptron (MLP) with linear activations and batch-normalization that provably has bounded depth. We also design an activation shaping scheme that empirically achieves the same properties for certain non-linear activations.
  arXiv  Detail & Related papers  (2023-10-03T12:35:02Z)
• Curve Your Attention: Mixed-Curvature Transformers for Graph Representation Learning [77.1421343649344]
  We propose a generalization of Transformers towards operating entirely on the product of constant curvature spaces. We also provide a kernelized approach to non-Euclidean attention, which enables our model to run in time and memory cost linear to the number of nodes and edges.
  arXiv  Detail & Related papers  (2023-09-08T02:44:37Z)
• Gradient-Based Feature Learning under Structured Data [57.76552698981579]
  In the anisotropic setting, the commonly used spherical gradient dynamics may fail to recover the true direction. We show that appropriate weight normalization that is reminiscent of batch normalization can alleviate this issue. In particular, under the spiked model with a suitably large spike, the sample complexity of gradient-based training can be made independent of the information exponent.
  arXiv  Detail & Related papers  (2023-09-07T16:55:50Z)
• On the Bias Against Inductive Biases [34.10348216388905]
  Self-supervised feature learning for visual tasks has seen state-of-the-art success using these extremely deep, isotropic networks. In this work, we analyze the effect of inductive biases on small to moderately-sized isotropic networks used for unsupervised visual feature learning.
  arXiv  Detail & Related papers  (2021-05-28T19:41:48Z)
• Coherent Gradients: An Approach to Understanding Generalization in Gradient Descent-based Optimization [15.2292571922932]
  We propose an approach to answering this question based on a hypothesis about the dynamics of gradient descent. We show that changes to the network parameters during training are biased towards those that (locally) simultaneously benefit many examples.
  arXiv  Detail & Related papers  (2020-02-25T03:59:31Z)

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.