Emergence and Localisation of Semantic Role Circuits in LLMs

2025-11-25

AI Summary

The internal representation of abstract semantic structures—particularly semantic roles—in large language models (LLMs) remains poorly understood. Method: We propose the “role-crossing minimal pair” localization method, integrating temporal emergence analysis, attribution tracing, and cross-architectural comparison across model scales to systematically identify and validate semantic role circuits. Contribution/Results: We discover a compact circuit comprising only 28 neurons that accounts for 89%–94% of semantic role attribution, exhibiting strong causal isolation and partial transferability across model sizes. Further analysis reveals the circuit operates via a progressive structural refinement mechanism, characterized by cross-scale conservation and high spectral similarity. This work provides the first neuron-level characterization of the dynamic formation pathway of semantic roles in LLMs, establishing a novel interpretability paradigm for semantic structure in foundation models.

Technology Category

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

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Large language models for search

Abstract

Despite displaying semantic competence, large language models' internal mechanisms that ground abstract semantic structure remain insufficiently characterised. We propose a method integrating role-cross minimal pairs, temporal emergence analysis, and cross-model comparison to study how LLMs implement semantic roles. Our analysis uncovers: (i) highly concentrated circuits (89-94% attribution within 28 nodes); (ii) gradual structural refinement rather than phase transitions, with larger models sometimes bypassing localised circuits; and (iii) moderate cross-scale conservation (24-59% component overlap) alongside high spectral similarity. These findings suggest that LLMs form compact, causally isolated mechanisms for abstract semantic structure, and these mechanisms exhibit partial transfer across scales and architectures.

Problem

• Characterizing internal mechanisms grounding semantic structure in LLMs
• Studying implementation of semantic roles through role-cross minimal pairs
• Analyzing circuit concentration and cross-scale conservation in models

Innovation

• Role-cross minimal pairs for semantic analysis
• Temporal emergence analysis of circuit development
• Cross-model comparison revealing structural conservation