Context Staircase: Signature-Aligned Dynamics of Token Embeddings under Small Initialization

• URL: http://arxiv.org/abs/2608.30315v1
• Date: Mon, 31 Aug 2026 06:33:09 GMT
• Title: Context Staircase: Signature-Aligned Dynamics of Token Embeddings under Small Initialization
• Abstract summary: Token embeddings are the basic representational units that connect discrete tokens with continuous computation in language models.<n>We identify that the evolving embedding structures are closely related to token-conditioned label and contextual distributions.<n>We show that these embedding structures play an important role in both task learning and the incorporation of semantic structure into the embedding space.
• Score: 7.898646648931469
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Token embeddings are the basic representational units that connect discrete tokens with continuous computation in language models. Although modern language models learn embeddings from random initialization through gradient-based training, the dynamical mechanism by which meaningful embedding structures emerge remains unclear. In this work, we identify that the evolving embedding structures are closely related to token-conditioned label and contextual distributions, which we formalize as probability signatures. We observe a progressive learning process, which we term Context Staircase: embeddings learn the low-order statistic signatures of the data before the high-order ones. More specifically, we observe that early in training they align with the simplest, context-free signature linking a token to its label, and as training proceeds, they progressively reflect signatures involving more and more context tokens. We then analyze the gradient flow of embeddings under small initialization to explain this phenomenon, deriving embedding evolution equations for feed-forward and self-attention architectures. We further extend these observations to real language-model training. Finally, we show that these embedding structures play an important role in both task learning and the incorporation of semantic structure into the embedding space. Overall, our results provide a dynamic explanation of how data statistics and architecture jointly shape token embeddings in language models, and reveal an implicit bias in the space of data statistics: training proceeds from simpler, low-order statistical relations toward increasingly complex, context-dependent ones.

Related papers

• Probability Signature: Bridging Data Semantics and Embedding Structure in Language Models [8.87728727154868]
  We propose a set of probability signatures that reflect the semantic relationships among tokens.<n>We generalize our work to large language models (LLMs) by training the Qwen2.5 architecture on the subsets of the Pile corpus.
  arXiv  Detail & Related papers  (2025-09-24T13:49:44Z)
• Scaling Laws and Representation Learning in Simple Hierarchical Languages: Transformers vs. Convolutional Architectures [49.19753720526998]
  We derive theoretical scaling laws for neural network performance on synthetic datasets.<n>We validate that convolutional networks, whose structure aligns with that of the generative process through locality and weight sharing, enjoy a faster scaling of performance.<n>This finding clarifies the architectural biases underlying neural scaling laws and highlights how representation learning is shaped by the interaction between model architecture and the statistical properties of data.
  arXiv  Detail & Related papers  (2025-05-11T17:44:14Z)
• From Text to Graph: Leveraging Graph Neural Networks for Enhanced Explainability in NLP [3.864700176441583]
  This study proposes a novel methodology to achieve explainability in natural language processing tasks.<n>It automatically converts sentences into graphs and maintains semantics through nodes and relations.<n>Experiments delivered promising results in determining the most critical components within the text structure for a given classification.
  arXiv  Detail & Related papers  (2025-04-02T18:55:58Z)
• The Foundations of Tokenization: Statistical and Computational Concerns [51.370165245628975]
  Tokenization is a critical step in the NLP pipeline.<n>Despite its recognized importance as a standard representation method in NLP, the theoretical underpinnings of tokenization are not yet fully understood.<n>The present paper contributes to addressing this theoretical gap by proposing a unified formal framework for representing and analyzing tokenizer models.
  arXiv  Detail & Related papers  (2024-07-16T11:12:28Z)
• LabelPrompt: Effective Prompt-based Learning for Relation Classification [31.291466190218912]
  This paper presents a novel prompt-based learning method, namely LabelPrompt, for the relation classification task. Motivated by the intuition to GIVE MODEL CHOICES!'', we first define additional tokens to represent relation labels, which regard these tokens as the verbaliser with semantic initialisation. Then, to mitigate inconsistency between predicted relations and given entities, we implement an entity-aware module with contrastive learning.
  arXiv  Detail & Related papers  (2023-02-16T04:06:25Z)
• Compositional Generalization in Grounded Language Learning via Induced Model Sparsity [81.38804205212425]
  We consider simple language-conditioned navigation problems in a grid world environment with disentangled observations. We design an agent that encourages sparse correlations between words in the instruction and attributes of objects, composing them together to find the goal. Our agent maintains a high level of performance on goals containing novel combinations of properties even when learning from a handful of demonstrations.
  arXiv  Detail & Related papers  (2022-07-06T08:46:27Z)
• Dependency Induction Through the Lens of Visual Perception [81.91502968815746]
  We propose an unsupervised grammar induction model that leverages word concreteness and a structural vision-based to jointly learn constituency-structure and dependency-structure grammars. Our experiments show that the proposed extension outperforms the current state-of-the-art visually grounded models in constituency parsing even with a smaller grammar size.
  arXiv  Detail & Related papers  (2021-09-20T18:40:37Z)
• Imposing Relation Structure in Language-Model Embeddings Using Contrastive Learning [30.00047118880045]
  We propose a novel contrastive learning framework that trains sentence embeddings to encode the relations in a graph structure. The resulting relation-aware sentence embeddings achieve state-of-the-art results on the relation extraction task.
  arXiv  Detail & Related papers  (2021-09-02T10:58:27Z)
• Prototypical Representation Learning for Relation Extraction [56.501332067073065]
  This paper aims to learn predictive, interpretable, and robust relation representations from distantly-labeled data. We learn prototypes for each relation from contextual information to best explore the intrinsic semantics of relations. Results on several relation learning tasks show that our model significantly outperforms the previous state-of-the-art relational models.
  arXiv  Detail & Related papers  (2021-03-22T08:11:43Z)
• Predicting What You Already Know Helps: Provable Self-Supervised Learning [60.27658820909876]
  Self-supervised representation learning solves auxiliary prediction tasks (known as pretext tasks) without requiring labeled data. We show a mechanism exploiting the statistical connections between certain em reconstruction-based pretext tasks that guarantee to learn a good representation. We prove the linear layer yields small approximation error even for complex ground truth function class.
  arXiv  Detail & Related papers  (2020-08-03T17:56:13Z)
• Temporal Embeddings and Transformer Models for Narrative Text Understanding [72.88083067388155]
  We present two approaches to narrative text understanding for character relationship modelling. The temporal evolution of these relations is described by dynamic word embeddings, that are designed to learn semantic changes over time. A supervised learning approach based on the state-of-the-art transformer model BERT is used instead to detect static relations between characters.
  arXiv  Detail & Related papers  (2020-03-19T14:23:12Z)

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.