Understanding Post-Training Structural Changes in Large Language Models

• URL: http://arxiv.org/abs/2509.17866v1
• Date: Mon, 22 Sep 2025 15:03:36 GMT
• Title: Understanding Post-Training Structural Changes in Large Language Models
• Authors: Xinyu He, Xianghui Cao,
• Abstract summary: Post-training fundamentally alters the behavior of large language models (LLMs)<n>This work focuses on two widely adopted post-training methods: instruction tuning and long-chain-of-thought (Long-CoT) distillation.
• Score: 3.054513120350576
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Post-training fundamentally alters the behavior of large language models (LLMs), yet its impact on the internal parameter space remains poorly understood. In this work, we conduct a systematic singular value decomposition (SVD) analysis of principal linear layers in pretrained LLMs, focusing on two widely adopted post-training methods: instruction tuning and long-chain-of-thought (Long-CoT) distillation. Our analysis reveals two consistent and unexpected structural changes:(1) a near-uniform geometric scaling of singular values across layers, which theoretically modulates attention scores; and (2) highly consistent orthogonal transformations are applied to the left and right singular vectors of each matrix. Disrupting this orthogonal consistency leads to catastrophic performance degradation. Based on these findings, we propose a simple yet effective framework that interprets post-training as a reparameterization of fixed subspaces in the pretrained parameter space. Further experiments reveal that singular value scaling behaves as a secondary effect, analogous to a temperature adjustment, whereas the core functional transformation lies in the coordinated rotation of singular vectors. These results challenge the prevailing view of the parameter space in large models as a black box, uncovering the first clear regularities in how parameters evolve during training, and providing a new perspective for deeper investigation into model parameter changes.

Related papers

• Low-Dimensional Execution Manifolds in Transformer Learning Dynamics: Evidence from Modular Arithmetic Tasks [0.0]
  We investigate the structure of learning dynamics in transformer models through carefully controlled arithmetic tasks.<n>Our results suggest a unifying geometric framework for understanding transformer learning.
  arXiv  Detail & Related papers  (2026-02-11T03:57:46Z)
• An Equivariance Toolbox for Learning Dynamics [13.651450618432094]
  We develop a general equivariance toolbox that yields coupled first- and second-order constraints on learning dynamics.<n>At the first order, our framework unifies conservation laws and implicit-bias relations as special cases of a single identity.<n>At the second order, it provides structural predictions about curvature.
  arXiv  Detail & Related papers  (2025-12-24T23:42:07Z)
• From Coefficients to Directions: Rethinking Model Merging with Directional Alignment [66.99062575537555]
  We introduce a unified geometric framework, emphMerging with Directional Alignment (method), which aligns directional structures consistently in both the parameter and feature spaces.<n>Our analysis shows that directional alignment improves structural coherence, and extensive experiments across benchmarks, model scales, and task configurations further validate the effectiveness of our approach.
  arXiv  Detail & Related papers  (2025-11-29T08:40:58Z)
• VIKING: Deep variational inference with stochastic projections [48.946143517489496]
  Variational mean field approximations tend to struggle with contemporary overparametrized deep neural networks.<n>We propose a simple variational family that considers two independent linear subspaces of the parameter space.<n>This allows us to build a fully-correlated approximate posterior reflecting the overparametrization.
  arXiv  Detail & Related papers  (2025-10-27T15:38:35Z)
• From Condensation to Rank Collapse: A Two-Stage Analysis of Transformer Training Dynamics [3.247992990696076]
  We use the gradient flow analytical framework to systematically investigate linearized Transformer training dynamics.<n>Our theoretical analysis dissects the dynamics of attention modules into two distinct stages.
  arXiv  Detail & Related papers  (2025-10-08T12:37:53Z)
• Time-Scale Coupling Between States and Parameters in Recurrent Neural Networks [3.924071936547547]
  Gated neural networks (RNNs) implicitly induce adaptive learning-rate behavior.<n>Effect arises from the coupling between state-space time scales--parametrized by the gates--and parameter-space dynamics.<n> Empirical simulations corroborate these claims.
  arXiv  Detail & Related papers  (2025-08-16T18:19:34Z)
• Generalized Linear Mode Connectivity for Transformers [87.32299363530996]
  A striking phenomenon is linear mode connectivity (LMC), where independently trained models can be connected by low- or zero-loss paths.<n>Prior work has predominantly focused on neuron re-ordering through permutations, but such approaches are limited in scope.<n>We introduce a unified framework that captures four symmetry classes: permutations, semi-permutations, transformations, and general invertible maps.<n>This generalization enables, for the first time, the discovery of low- and zero-barrier linear paths between independently trained Vision Transformers and GPT-2 models.
  arXiv  Detail & Related papers  (2025-06-28T01:46:36Z)
• PAID: Pairwise Angular-Invariant Decomposition for Continual Test-Time Adaptation [70.98107766265636]
  This paper takes the geometric attributes of pre-trained weights as a starting point, systematically analyzing three key components: magnitude, absolute angle, and pairwise angular structure.<n>We find that the pairwise angular structure remains stable across diverse corrupted domains and encodes domain-invariant semantic information, suggesting it should be preserved during adaptation.
  arXiv  Detail & Related papers  (2025-06-03T05:18:15Z)
• Weight Spectra Induced Efficient Model Adaptation [54.8615621415845]
  Fine-tuning large-scale foundation models incurs prohibitive computational costs.<n>We show that fine-tuning predominantly amplifies the top singular values while leaving the remainder largely intact.<n>We propose a novel method that leverages learnable rescaling of top singular directions.
  arXiv  Detail & Related papers  (2025-05-29T05:03:29Z)
• Generalized Tensor-based Parameter-Efficient Fine-Tuning via Lie Group Transformations [50.010924231754856]
  Adapting pre-trained foundation models for diverse downstream tasks is a core practice in artificial intelligence.<n>To overcome this, parameter-efficient fine-tuning (PEFT) methods like LoRA have emerged and are becoming a growing research focus.<n>We propose a generalization that extends matrix-based PEFT methods to higher-dimensional parameter spaces without compromising their structural properties.
  arXiv  Detail & Related papers  (2025-04-01T14:36:45Z)
• On the Emergence of Cross-Task Linearity in the Pretraining-Finetuning Paradigm [47.55215041326702]
  We discover an intriguing linear phenomenon in models that are from a common pretrained checkpoint and finetuned on different tasks, termed as Cross-Task Linearity (CTL) We show that if we linearly interpolate the weights of two finetuned models, the features in the weight-interpolated model are often approximately equal to the linearities of features in two finetuned models at each layer. We conjecture that in the pretraining-finetuning paradigm, neural networks approximately function as linear maps, mapping from the parameter space to the feature space.
  arXiv  Detail & Related papers  (2024-02-06T03:28:36Z)
• Analysis of Catastrophic Forgetting for Random Orthogonal Transformation Tasks in the Overparameterized Regime [9.184987303791292]
  We show that in permuted MNIST image classification tasks, the performance of multilayer perceptrons trained by vanilla gradient descent can be improved. We provide a theoretical explanation of this effect by studying a qualitatively similar two-task linear regression problem. We show that when a model is trained on the two tasks in sequence without any additional regularization, the risk gain on the first task is small.
  arXiv  Detail & Related papers  (2022-06-01T18:04:33Z)
• Deformation Robust Roto-Scale-Translation Equivariant CNNs [10.44236628142169]
  Group-equivariant convolutional neural networks (G-CNNs) achieve significantly improved generalization performance with intrinsic symmetry. General theory and practical implementation of G-CNNs have been studied for planar images under either rotation or scaling transformation.
  arXiv  Detail & Related papers  (2021-11-22T03:58:24Z)

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.