Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Component-Level Lesioning of Language Models Reveals Clinically Aligned Aphasia Phenotypes

Yifan Wang, Jichen Zheng, Jingyuan Sun, Yunhao Zhang, Chunyu Ye, Jixing Li, Chengqing Zong, Shaonan Wang

TL;DR

This study investigates whether large language models can model human aphasia by perturbing clinically relevant internal components. It introduces a clinically grounded framework that targets subtype-linked units in both Mixture-of-Experts and dense Transformer LLMs, linking perturbations to observable aphasia profiles through BLiMP probing, AphasiaBank supervision, and Western Aphasia Battery metrics summarized as the Aphasia Quotient. Across architectures, targeted, top-% perturbations yield graded, aphasia-like impairments, with MoE models providing more localized and interpretable mappings between disrupted components and language deficits. The work demonstrates a scalable approach to simulating aphasia, enabling systematic study of how distinct language functions degrade under targeted disruptions and informing rehabilitation research and theories of language organization.

Abstract

Large language models (LLMs) increasingly exhibit human-like linguistic behaviors and internal representations that they could serve as computational simulators of language cognition. We ask whether LLMs can be systematically manipulated to reproduce language-production impairments characteristic of aphasia following focal brain lesions. Such models could provide scalable proxies for testing rehabilitation hypotheses, and offer a controlled framework for probing the functional organization of language. We introduce a clinically grounded, component-level framework that simulates aphasia by selectively perturbing functional components in LLMs, and apply it to both modular Mixture-of-Experts models and dense Transformers using a unified intervention interface. Our pipeline (i) identifies subtype-linked components for Broca's and Wernicke's aphasia, (ii) interprets these components with linguistic probing tasks, and (iii) induces graded impairments by progressively perturbing the top-k subtype-linked components, evaluating outcomes with Western Aphasia Battery (WAB) subtests summarized by Aphasia Quotient (AQ). Across architectures and lesioning strategies, subtype-targeted perturbations yield more systematic, aphasia-like regressions than size-matched random perturbations, and MoE modularity supports more localized and interpretable phenotype-to-component mappings. These findings suggest that modular LLMs, combined with clinically informed component perturbations, provide a promising platform for simulating aphasic language production and studying how distinct language functions degrade under targeted disruptions.

Component-Level Lesioning of Language Models Reveals Clinically Aligned Aphasia Phenotypes

TL;DR

This study investigates whether large language models can model human aphasia by perturbing clinically relevant internal components. It introduces a clinically grounded framework that targets subtype-linked units in both Mixture-of-Experts and dense Transformer LLMs, linking perturbations to observable aphasia profiles through BLiMP probing, AphasiaBank supervision, and Western Aphasia Battery metrics summarized as the Aphasia Quotient. Across architectures, targeted, top-% perturbations yield graded, aphasia-like impairments, with MoE models providing more localized and interpretable mappings between disrupted components and language deficits. The work demonstrates a scalable approach to simulating aphasia, enabling systematic study of how distinct language functions degrade under targeted disruptions and informing rehabilitation research and theories of language organization.

Abstract

Large language models (LLMs) increasingly exhibit human-like linguistic behaviors and internal representations that they could serve as computational simulators of language cognition. We ask whether LLMs can be systematically manipulated to reproduce language-production impairments characteristic of aphasia following focal brain lesions. Such models could provide scalable proxies for testing rehabilitation hypotheses, and offer a controlled framework for probing the functional organization of language. We introduce a clinically grounded, component-level framework that simulates aphasia by selectively perturbing functional components in LLMs, and apply it to both modular Mixture-of-Experts models and dense Transformers using a unified intervention interface. Our pipeline (i) identifies subtype-linked components for Broca's and Wernicke's aphasia, (ii) interprets these components with linguistic probing tasks, and (iii) induces graded impairments by progressively perturbing the top-k subtype-linked components, evaluating outcomes with Western Aphasia Battery (WAB) subtests summarized by Aphasia Quotient (AQ). Across architectures and lesioning strategies, subtype-targeted perturbations yield more systematic, aphasia-like regressions than size-matched random perturbations, and MoE modularity supports more localized and interpretable phenotype-to-component mappings. These findings suggest that modular LLMs, combined with clinically informed component perturbations, provide a promising platform for simulating aphasic language production and studying how distinct language functions degrade under targeted disruptions.
Paper Structure (23 sections, 7 equations, 4 figures, 2 tables)

This paper contains 23 sections, 7 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Aphasia and the Neural Basis of Language
  4. Simulating Language Disorders with Computational Models
  5. Mixture-of-Experts Models and Component-Level Lesioning
  6. Method
  7. Model Architecture
  8. Unit Attribution via BLiMP Zero-Ablation
  9. Aphasia Phenotype-Specific Unit Contribution
  10. Threshold Selection and Robustness
  11. Alignment of Language Phenomena with Clinical Phenotypes
  12. Damage Model for Progressive Aphasia Lesions
  13. WAB-Based Clinical Evaluation and Aphasia Quotient
  14. Experimental Setup
  15. Dataset
  16. ...and 8 more sections

Figures (4)

  • Figure 1: Details of the analysis pipeline. Functional Profiling and Phenotype–Phenomenon Alignment: BLiMP probing and AphasiaBank fine-tuning yield unit-importance rankings; a heatmap links linguistic phenomena to Broca/Wernicke units, with CAP as an external phenotype check. Progressive Lesioning and Clinical Validation: Top-k% units are progressively lesioned and evaluated on WAB/AQ, enabling a matched comparison between MoE and dense models.
  • Figure 2: Rank-percentile heatmaps and pie chart for top-2% subtype-relevant units. [a–b] Dense (OLMo) neurons (Wernicke, Broca). [c–d] MoE (OLMoE) experts (Wernicke, Broca). Color shows rank percentile (lower = more important) across BLiMP sub-tasks and summary columns. [e] Proportion of dominant BLiMP sub-tasks among the top-2% units (by each unit’s maximum contribution).
  • Figure 3: Robustness of subtype task profiles around the 2% lesion choice. [a] MoE (OLMoE) and [b] dense (OLMo) show Spearman similarity between profiles at lesion ratio $p \in (0.5\%, 1\%, 2\%, 3\%, 5\%, 10\%)$ and the 2% profile for Broca- and Wernicke-targeted units.
  • Figure 4: WAB Test Score. Progressive lesioning (Xavier) yields graded clinical degradation on the WAB evaluation (Aphasia Quotient, AQ). [a] Dense OLMo and [b] MoE OLMoE.