Transformers learn factored representations

• URL: http://arxiv.org/abs/2602.02385v1
• Date: Mon, 02 Feb 2026 17:49:06 GMT
• Title: Transformers learn factored representations
• Authors: Adam Shai, Loren Amdahl-Culleton, Casper L. Christensen, Henry R. Bigelow, Fernando E. Rosas, Alexander B. Boyd, Eric A. Alt, Kyle J. Ray, Paul M. Riechers,
• Abstract summary: Transformers pretrained via next token prediction learn to factor their world into parts.<n>We formalize two representational hypotheses: (1) a representation in the product space of all factors, whose dimension grows exponentially with the number of parts.<n>We derive precise predictions about the geometric structure of activations for each, including the number of subspaces, their dimensionality, and the arrangement of context embeddings within them.<n>This provides a principled explanation for why transformers decompose the world into parts, and suggests that interpretable low dimensional structure may persist even in models trained on complex data.
• Score: 62.86679034549244
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Transformers pretrained via next token prediction learn to factor their world into parts, representing these factors in orthogonal subspaces of the residual stream. We formalize two representational hypotheses: (1) a representation in the product space of all factors, whose dimension grows exponentially with the number of parts, or (2) a factored representation in orthogonal subspaces, whose dimension grows linearly. The factored representation is lossless when factors are conditionally independent, but sacrifices predictive fidelity otherwise, creating a tradeoff between dimensional efficiency and accuracy. We derive precise predictions about the geometric structure of activations for each, including the number of subspaces, their dimensionality, and the arrangement of context embeddings within them. We test between these hypotheses on transformers trained on synthetic processes with known latent structure. Models learn factored representations when factors are conditionally independent, and continue to favor them early in training even when noise or hidden dependencies undermine conditional independence, reflecting an inductive bias toward factoring at the cost of fidelity. This provides a principled explanation for why transformers decompose the world into parts, and suggests that interpretable low dimensional structure may persist even in models trained on complex data.

Related papers

• On the Emergence of Induction Heads for In-Context Learning [121.64612469118464]
  We study the emergence of induction heads, a previously identified mechanism in two-layer transformers.<n>We explain the origin of this structure using a minimal ICL task formulation and a modified transformer architecture.
  arXiv  Detail & Related papers  (2025-11-02T18:12:06Z)
• Transformers Are Universally Consistent [14.904264782690639]
  We show that Transformers equipped with softmax-based nonlinear attention are uniformly consistent when tasked with executing Least Squares regression.<n>We derive upper bounds on the empirical error which, in the regime, decay at a provable rate of $mathcalO(t-1/2d)$, where $t$ denotes the number of input tokens and $d$ the embedding dimensionality.
  arXiv  Detail & Related papers  (2025-05-30T12:39:26Z)
• 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)
• Nonparametric Partial Disentanglement via Mechanism Sparsity: Sparse Actions, Interventions and Sparse Temporal Dependencies [58.179981892921056]
  This work introduces a novel principle for disentanglement we call mechanism sparsity regularization. We propose a representation learning method that induces disentanglement by simultaneously learning the latent factors. We show that the latent factors can be recovered by regularizing the learned causal graph to be sparse.
  arXiv  Detail & Related papers  (2024-01-10T02:38:21Z)
• Why bother with geometry? On the relevance of linear decompositions of Transformer embeddings [5.151529346168568]
  We study representations from machine-translation decoders using two of such embedding decomposition methods. Our results indicate that, while decomposition-derived indicators effectively correlate with model performance, variation across different runs suggests a more nuanced take on this question.
  arXiv  Detail & Related papers  (2023-10-10T19:56:10Z)
• CausalX: Causal Explanations and Block Multilinear Factor Analysis [3.087360758008569]
  We propose a unified multilinear model of wholes and parts. We introduce an incremental bottom-up computational alternative, the Incremental M-mode Block SVD. The resulting object representation is an interpretable choice of intrinsic causal factor representations related to an object's hierarchy of wholes and parts.
  arXiv  Detail & Related papers  (2021-02-25T13:49:01Z)
• Learning Disentangled Representations with Latent Variation Predictability [102.4163768995288]
  This paper defines the variation predictability of latent disentangled representations. Within an adversarial generation process, we encourage variation predictability by maximizing the mutual information between latent variations and corresponding image pairs. We develop an evaluation metric that does not rely on the ground-truth generative factors to measure the disentanglement of latent representations.
  arXiv  Detail & Related papers  (2020-07-25T08:54:26Z)
• CausalVAE: Structured Causal Disentanglement in Variational Autoencoder [52.139696854386976]
  The framework of variational autoencoder (VAE) is commonly used to disentangle independent factors from observations. We propose a new VAE based framework named CausalVAE, which includes a Causal Layer to transform independent factors into causal endogenous ones. Results show that the causal representations learned by CausalVAE are semantically interpretable, and their causal relationship as a Directed Acyclic Graph (DAG) is identified with good accuracy.
  arXiv  Detail & Related papers  (2020-04-18T20:09:34Z)

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.