Differential syntactic and semantic encoding in LLMs

2026/01/08 by Santiago Acevedo, Alessandro Laio, Marco Baroni · 2 voices
Computer Science · Physics and Astronomy · Social Sciences · #Computational and Text Analysis Methods #Differential coding #Encoding (memory) #Generative Adversarial Networks and Image Synthesis #Semantics (computer science) #Sentence #Similarity (geometry) #Syntactic structure #Syntax #Topic Modeling #cs.AI #cs.CL #cs.LG #physics.comp-ph

paper · pdf · open access · doi:10.48550/arxiv.2601.04765

published in arXiv (Cornell University) (Cornell University)

openalex publication_date 2026/01/08 · arxiv published 2026/01/08 · openalex created_date 2026/01/10 · arxiv updated 2026/05/28 · openalex updated_date 2026/07/28

Abstract

We study how syntactic and semantic information is encoded in inner layer representations of Large Language Models (LLMs), focusing on the very large DeepSeek-V3. We find that, by averaging hidden-representation vectors of sentences sharing syntactic structure or meaning, we obtain vectors that capture a significant proportion of the syntactic and semantic information contained in the representations. In particular, subtracting these syntactic and semantic ``centroids'' from sentence vectors strongly affects their similarity with syntactically and semantically matched sentences, respectively, suggesting that syntax and semantics are, at least partially, linearly encoded. We also find that the cross-layer encoding profiles of syntax and semantics are different, and that the two signals can to some extent be decoupled, suggesting differential encoding of these two types of linguistic information in LLM representations.

Citations

• Learning the Wrong Lessons: Syntactic-Domain Spurious Correlations in Language Models
• Probing Syntax in Large Language Models: Successes and Remaining Challenges
• On the Emergence of Linear Analogies in Word Embeddings
• A quantitative analysis of semantic information in deep representations of text and images
• Layers at Similar Depths Generate Similar Activations Across LLM Architectures
• Shared Global and Local Geometry of Language Model Embeddings
• Gemma 3 Technical Report
• Layer by Layer: Uncovering Hidden Representations in Language Models
• You can’t fight in here! This is BBS!
• How linguistics learned to stop worrying and love the language models
• Large Language Models Share Representations of Latent Grammatical Concepts Across Typologically Diverse Languages
• Concept Space Alignment in Multilingual LLMs
• Evidence from fMRI Supports a Two-Phase Abstraction Process in Language Models
• Qwen2 Technical Report
• The Geometry of Categorical and Hierarchical Concepts in Large Language Models
• Emergence of a High-Dimensional Abstraction Phase in Language Transformers
• A Primer on the Inner Workings of Transformer-based Language Models
• Decoding Probing: Revealing Internal Linguistic Structures in Neural Language Models using Minimal Pairs
• Characterizing Truthfulness in Large Language Model Generations with Local Intrinsic Dimension
• Massive Activations in Large Language Models
• SGLang: Efficient Execution of Structured Language Model Programs
• Improving Activation Steering in Language Models with Mean-Centring
• The Linear Representation Hypothesis and the Geometry of Large Language Models
• Getting aligned on representational alignment
• Explainability for Large Language Models: A Survey
• Linearity of Relation Decoding in Transformer Language Models
• Quantizable Transformers: Removing Outliers by Helping Attention Heads Do Nothing
• Similarity of Neural Network Models: A Survey of Functional and Representational Measures
• A Note on Normalized Emergence Timing (in Pythia Language Model Evaluations)
• Revisiting Model Stitching to Compare Neural Representations
• BERT Busters: Outlier Dimensions that Disrupt Transformers
• Language Through a Prism: A Spectral Approach for Multiscale Language Representations
• Designing and Interpreting Probes with Control Tasks
• BERT Rediscovers the Classical NLP Pipeline
• Similarity of Neural Network Representations Revisited
• Linguistic Knowledge and Transferability of Contextual Representations
• All-but-the-Top: Simple and Effective Postprocessing for Word Representations
• Scikit-learn: Machine Learning in Python

Discussions

• #cs.CL #cs.AI #cs.LG #physics.comp-ph Origin | Interest | Match [bsky, 0 points, 0 comments]
• #cs.CL #cs.AI #cs.LG #physics.comp-ph Origin | Interest | Match [bsky, 0 points, 0 comments]

Related

• Semantic Structure in Large Language Model Embeddings
• Probing LLMs for Joint Encoding of Linguistic Categories
• Large Language Models Encode Semantics and Alignment in Linearly Separable Representations
• Understanding LLM Behaviors via Compression: Data Generation, Knowledge Acquisition and Scaling Laws
• Explaining Large Language Model-Based Neural Semantic Parsers (Student Abstract)