On the Emergence and Test-Time Use of Structural Information in Large Language Models

2026-01-25

AI Summary

This study investigates how large language models learn abstract structural information from observational data and leverage it for compositional generation and complex reasoning during inference. To this end, we construct a controlled natural language dataset grounded in linguistic structural transformations and systematically analyze the mechanisms by which models acquire structural knowledge and generalize it to novel contexts. Our work provides the first empirical evidence of a positive correlation between structural learning and performance on complex reasoning tasks, while also revealing significant limitations in current models’ ability to achieve compositional generalization at test time. Through carefully designed structured data, transformation modeling, and behavioral analysis, this research offers a novel perspective on the structural inductive biases and reasoning capabilities of large language models.

Technology Category

• Machine Learning: Structured Learning
• Natural Language Processing: (Large) Language Models
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Learning structural information from observational data is central to producing new knowledge outside the training corpus. This holds for mechanistic understanding in scientific discovery as well as flexible test-time compositional generation. We thus study how language models learn abstract structures and utilize the learnt structural information at test-time. To ensure a controlled setup, we design a natural language dataset based on linguistic structural transformations. We empirically show that the emergence of learning structural information correlates with complex reasoning tasks, and that the ability to perform test-time compositional generation remains limited.

Problem

• structural information
• large language models
• compositional generation
• complex reasoning
• test-time utilization

Innovation

• structural information
• large language models
• compositional generation
• test-time reasoning
• linguistic transformations