Artificial Aphasias in Lesioned Language Models

2026-05-15

AI Summary

This study investigates whether language models exhibit symptoms analogous to human aphasia when subjected to parameter corruption, aiming to uncover their functional organization. By systematically ablating (zeroing) model parameters and evaluating outputs using the Textual Aphasia Battery (TAB), the work pioneers the application of clinical aphasia diagnostic frameworks to large language models. Experiments across five 1B-scale models reveal distinct symptom patterns: ablation of attention versus feedforward modules produces markedly different deficits, with early-layer damage primarily impairing syntactic and semantic processing, while late-layer lesions predominantly disrupt phonological accuracy and fluency. The findings establish causal links between specific architectural components and linguistic functions, while also highlighting fundamental differences in the distribution of aphasic symptoms between humans and artificial models.

Abstract

Aphasias, selective language impairments which can arise from brain damage, reveal the functional organization of human language by providing causal links between affected brain regions and specific symptom profiles. Drawing on this literature, we introduce an aphasia-inspired technique to characterize the emergent functional organization of language models (LMs). We ``lesion''(zero-out) model parameters and measure the effects of this intervention against clinical aphasia symptoms, as diagnosed by the Text Aphasia Battery (TAB). When applied to 112,426 outputs from five 1B-scale LMs, the full range of evaluated symptoms surface, but in distributions largely distinct from those of humans. Our method uncovers broad symptom-profile differences between attention components (query, key, value, output) and feed-forward components (up, gate, down), with weaker evidence for differences among components within the same mechanism. We also find an effect of depth, where lesions in early layers disproportionately cause syntactic and semantic symptoms while late-middle layers yield higher rates of phonological and fluency deficits. Although some LM lesions induce quantitatively more similar profiles to some human aphasia types than others, qualitative differences in symptom patterns between LMs and humans suggest that aphasia syndromes are heavily influenced by the details of learning and processing rather than being a domain-invariant consequence of disrupted language processing.

Problem

• aphasia
• language models
• lesion
• functional organization
• symptom profiles

Innovation

• artificial aphasia
• lesion study
• language models
• functional organization
• Text Aphasia Battery