How Do Language Models Understand Tables? A Mechanistic Analysis of Cell Location

2026-02-09

AI Summary

This study addresses the unclear internal mechanisms by which large language models comprehend linearized two-dimensional tables, particularly regarding the fundamental task of cell localization. By integrating interpretability techniques—including activation patching, vector arithmetic, and attention head analysis—the work systematically reveals that table understanding unfolds in three stages: semantic binding, coordinate localization, and information extraction. The authors find that models localize cells by counting delimiters and encode column indices within linear subspaces. Moreover, specific attention heads are reused across multi-cell tasks. This mechanism not only elucidates the model’s ordinal reasoning capabilities but also enables precise control over its focus positions, demonstrating strong generalization across diverse tasks.

Technology Category

• Natural Language Processing: (Large) Language Models
• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Large Vision Models

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

While Large Language Models (LLMs) are increasingly deployed for table-related tasks, the internal mechanisms enabling them to process linearized two-dimensional structured tables remain opaque. In this work, we investigate the process of table understanding by dissecting the atomic task of cell location. Through activation patching and complementary interpretability techniques, we delineate the table understanding mechanism into a sequential three-stage pipeline: Semantic Binding, Coordinate Localization, and Information Extraction. We demonstrate that models locate the target cell via an ordinal mechanism that counts discrete delimiters to resolve coordinates. Furthermore, column indices are encoded within a linear subspace that allows for precise steering of model focus through vector arithmetic. Finally, we reveal that models generalize to multi-cell location tasks by multiplexing the identical attention heads identified during atomic location. Our findings provide a comprehensive explanation of table understanding within Transformer architectures.

Problem

• language models
• table understanding
• cell location
• mechanistic analysis
• transformer architectures

Innovation

• table understanding
• mechanistic interpretability
• activation patching
• coordinate localization
• attention heads