Large Language Models are Locally Linear Mappings

• URL: http://arxiv.org/abs/2505.24293v2
• Date: Wed, 04 Jun 2025 03:50:57 GMT
• Title: Large Language Models are Locally Linear Mappings
• Abstract summary: We map the inference operations of several open-weight large language models to an exactly equivalent linear system for an input sequence.<n>Despite their power and global nonlinearity, modern LLMs can be interpreted through nearly-exact locally linear decompositions.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: We demonstrate that the inference operations of several open-weight large language models (LLMs) can be mapped to an exactly equivalent linear system for an input sequence without modifying the model weights or altering output predictions. Extending techniques from image diffusion models that exhibit local or piecewise linearity, we strategically alter the gradient computation with respect to a given input sequence for a next-token prediction such that the Jacobian of the model nearly exactly reproduces the forward prediction with a linear system. We demonstrate this approach across models (Llama 3, Gemma 3, Qwen 3, Phi 4, Mistral Ministral and OLMo 2, up to Llama 3.3 70B Q4) and show through the singular value decomposition of the detached Jacobian that these LLMs operate in extremely low-dimensional subspaces where many of the largest singular vectors decode to concepts related to the most-likely output token. This approach also allows us to examine the operation of each successive layer (and its attention and MLP components) as nearly-exact linear systems and observe the emergence of semantic concepts. Despite their expressive power and global nonlinearity, modern LLMs can be interpreted through nearly-exact locally linear decompositions that provide insights into their internal representations and reveal interpretable semantic structures in the next-token prediction process.

Related papers

• Factual recall in linear associative memories: sharp asymptotics and mechanistic insights [11.11958822145404]
  We study the limits of storing and retrieving input--output associations with neural networks.<n>A linear associative memory maps $p$ input embeddings in $mathbbRd$ to their corresponding$d$-dimensional targets.<n>We show that a decoupled model can store up to $p_c log p_c / d2 = 1 / 2$ associations.
  arXiv  Detail & Related papers  (2026-05-11T16:24:02Z)
• Learning to Recall with Transformers Beyond Orthogonal Embeddings [42.18876773867171]
  We analyze a transformer with random embeddings trained with gradient descent on a simple token-retrieval task.<n>Our analysis tracks the early phase'' of gradient descent and yields explicit formulas for the model's storage capacity.
  arXiv  Detail & Related papers  (2026-03-16T21:17:01Z)
• Task Matrices: Linear Maps for Cross-Model Finetuning Transfer [4.243926243206826]
  We develop the concept of a task matrix, a linear transformation from a base to finetuned embedding state.<n>We show that for vision and text models and ten different datasets, a base model augmented with a task matrix results surpassing linear probes.<n>Our results validate the existence of cross-layer linear encodings between pretrained and finetuned architectures.
  arXiv  Detail & Related papers  (2025-12-16T19:51:29Z)
• Sequential-Parallel Duality in Prefix Scannable Models [68.39855814099997]
  Recent developments have given rise to various models, such as Gated Linear Attention (GLA) and Mamba.<n>This raises a natural question: can we characterize the full class of neural sequence models that support near-constant-time parallel evaluation and linear-time, constant-space sequential inference?
  arXiv  Detail & Related papers  (2025-06-12T17:32:02Z)
• How can representation dimension dominate structurally pruned LLMs? [17.953689537875377]
  Pruning assumes a subnetwork exists in the original deep neural network.<n>It is unclear how the model performance varies with the different subnetwork extractions.
  arXiv  Detail & Related papers  (2025-03-06T12:28:59Z)
• Large Language-Geometry Model: When LLM meets Equivariance [53.8505081745406]
  We propose EquiLLM, a novel framework for representing 3D physical systems.<n>We show that EquiLLM delivers significant improvements over previous methods across molecular dynamics simulation, human motion simulation, and antibody design.
  arXiv  Detail & Related papers  (2025-02-16T14:50:49Z)
• Demystifying Singular Defects in Large Language Models [61.98878352956125]
  In large language models (LLMs), the underlying causes of high-norm tokens remain largely unexplored.<n>We provide both theoretical insights and empirical validation across a range of recent models.<n>We showcase two practical applications of these findings: the improvement of quantization schemes and the design of LLM signatures.
  arXiv  Detail & Related papers  (2025-02-10T20:09:16Z)
• Predicting the Performance of Black-box LLMs through Self-Queries [60.87193950962585]
  Large language models (LLMs) are increasingly relied on in AI systems, predicting when they make mistakes is crucial.<n>In this paper, we extract features of LLMs in a black-box manner by using follow-up prompts and taking the probabilities of different responses as representations.<n>We demonstrate that training a linear model on these low-dimensional representations produces reliable predictors of model performance at the instance level.
  arXiv  Detail & Related papers  (2025-01-02T22:26:54Z)
• Pushing the Limits of Large Language Model Quantization via the Linearity Theorem [71.3332971315821]
  We present a "line theoremarity" establishing a direct relationship between the layer-wise $ell$ reconstruction error and the model perplexity increase due to quantization. This insight enables two novel applications: (1) a simple data-free LLM quantization method using Hadamard rotations and MSE-optimal grids, dubbed HIGGS, and (2) an optimal solution to the problem of finding non-uniform per-layer quantization levels.
  arXiv  Detail & Related papers  (2024-11-26T15:35:44Z)
• Learning Linear Attention in Polynomial Time [127.14106124669645]
  We provide the first results on learnability of single-layer Transformers with linear attention.<n>We show that linear attention may be viewed as a linear predictor in a suitably defined RKHS.<n>We show how to efficiently identify training datasets for which every empirical riskr is equivalent to the linear Transformer.
  arXiv  Detail & Related papers  (2024-10-14T02:41:01Z)
• Unsupervised Representation Learning from Sparse Transformation Analysis [79.94858534887801]
  We propose to learn representations from sequence data by factorizing the transformations of the latent variables into sparse components. Input data are first encoded as distributions of latent activations and subsequently transformed using a probability flow model.
  arXiv  Detail & Related papers  (2024-10-07T23:53:25Z)
• Unveiling Induction Heads: Provable Training Dynamics and Feature Learning in Transformers [54.20763128054692]
  We study how a two-attention-layer transformer is trained to perform ICL on $n$-gram Markov chain data. We prove that the gradient flow with respect to a cross-entropy ICL loss converges to a limiting model.
  arXiv  Detail & Related papers  (2024-09-09T18:10:26Z)
• Transformer Block Coupling and its Correlation with Generalization in LLMs [3.007031501305338]
  We analyze the trajectories of token embeddings as they pass through transformer blocks, linearizing the system along these trajectories through their Jacobian matrices.<n>We uncover the phenomenon of textbftransformer block coupling in a multitude of Large Language Models, characterized by the coupling of their top singular vectors across tokens and depth.<n>We further investigate how these properties emerge during training, observing a progressive development of coupling, increased linearity, and layer-wise exponential growth in token trajectories.
  arXiv  Detail & Related papers  (2024-07-10T16:30:27Z)
• Optimal Matrix-Mimetic Tensor Algebras via Variable Projection [0.0]
  Matrix mimeticity arises from interpreting tensors as operators that can be multiplied, factorized, and analyzed analogous to matrices. We learn optimal linear mappings and corresponding tensor representations without relying on prior knowledge of the data. We provide original theory of uniqueness of the transformation and convergence analysis of our variable-projection-based algorithm.
  arXiv  Detail & Related papers  (2024-06-11T04:52:23Z)
• Weight-based Decomposition: A Case for Bilinear MLPs [0.0]
  Gated Linear Units (GLUs) have become a common building block in modern foundation models. Bilinear layers drop the non-linearity in the "gate" but still have comparable performance to other GLUs. We develop a method to decompose the bilinear tensor into a set of interacting eigenvectors.
  arXiv  Detail & Related papers  (2024-06-06T10:46:51Z)
• Unlocking the Secrets of Linear Complexity Sequence Model from A Unified Perspective [26.479602180023125]
  The Linear Complexity Sequence Model (LCSM) unites various sequence modeling techniques with linear complexity. We segment the modeling processes of these models into three distinct stages: Expand, Oscillation, and Shrink. We perform experiments to analyze the impact of different stage settings on language modeling and retrieval tasks.
  arXiv  Detail & Related papers  (2024-05-27T17:38:55Z)
• On the Origins of Linear Representations in Large Language Models [51.88404605700344]
  We introduce a simple latent variable model to formalize the concept dynamics of the next token prediction. Experiments show that linear representations emerge when learning from data matching the latent variable model. We additionally confirm some predictions of the theory using the LLaMA-2 large language model.
  arXiv  Detail & Related papers  (2024-03-06T17:17:36Z)
• Unfolding Projection-free SDP Relaxation of Binary Graph Classifier via GDPA Linearization [59.87663954467815]
  Algorithm unfolding creates an interpretable and parsimonious neural network architecture by implementing each iteration of a model-based algorithm as a neural layer. In this paper, leveraging a recent linear algebraic theorem called Gershgorin disc perfect alignment (GDPA), we unroll a projection-free algorithm for semi-definite programming relaxation (SDR) of a binary graph. Experimental results show that our unrolled network outperformed pure model-based graph classifiers, and achieved comparable performance to pure data-driven networks but using far fewer parameters.
  arXiv  Detail & Related papers  (2021-09-10T07:01:15Z)
• Eigendecomposition-Free Training of Deep Networks for Linear Least-Square Problems [107.3868459697569]
  We introduce an eigendecomposition-free approach to training a deep network. We show that our approach is much more robust than explicit differentiation of the eigendecomposition. Our method has better convergence properties and yields state-of-the-art results.
  arXiv  Detail & Related papers  (2020-04-15T04:29:34Z)
• Semi-Supervised Learning with Normalizing Flows [54.376602201489995]
  FlowGMM is an end-to-end approach to generative semi supervised learning with normalizing flows. We show promising results on a wide range of applications, including AG-News and Yahoo Answers text data.
  arXiv  Detail & Related papers  (2019-12-30T17:36:33Z)

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.