Language Models Achieve Aphasia Phenotypes Via Component-Level Lesioning of Functional Units

Researchers are investigating whether large language models (LLMs) can accurately mimic the language impairments seen in aphasia following brain injury. Yifan Wang, Jichen Zheng from the State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, CAS, Beijing, China, Jingyuan Sun, and Yunhao Zhang et al. present a novel method for selectively ‘lesioning’ components within LLMs to replicate the effects of Broca’s and Wernicke’s aphasia. This work is significant because it establishes a scalable computational model for understanding language cognition and testing potential rehabilitation strategies , offering a controlled environment to explore how language functions break down under specific disruptions. By pinpointing components linked to aphasia subtypes and inducing graded impairments, the team demonstrate that these models can produce remarkably realistic, aphasia-like regressions, particularly within modular LLM architectures.

This work introduces a clinically grounded framework that selectively perturbs functional components within LLMs, offering a scalable proxy for testing rehabilitation hypotheses and probing the functional organisation of language itself. The team achieved this by developing a unified intervention interface applicable to both modular Mixture-of-Experts (MoE) models and dense Transformer architectures.

The study meticulously identifies components within LLMs linked to specific aphasia subtypes, interpreting these components using linguistic probing tasks to understand their function. This pipeline allowed for a controlled assessment of how targeted disruptions affect language production, mirroring the effects of focal brain lesions. Experiments revealed that subtype-targeted perturbations consistently produced more realistic, aphasia-like regressions in performance compared to random perturbations of equal size. Notably, the modularity of MoE models proved advantageous, supporting more localised and interpretable mappings between language phenotypes and the perturbed components.
This suggests that the inherent structure of these models aligns well with the partially specialised architecture of language processing in the human brain. This breakthrough opens new avenues for personalised rehabilitation strategies and a deeper understanding of the neural basis of language. This innovative approach moves beyond simply assessing LLM robustness, instead focusing on aligning simulated impairments with standardised clinical scoring and distinct aphasia subtypes. The team’s work builds upon the long-standing view of language as a modular system, where different brain regions contribute to specific linguistic functions. By “lesioning” specific components within LLMs, they successfully reproduced patient-like error patterns, albeit with a scale and linguistic expressivity previously unattainable with smaller models. The findings suggest that modular LLMs, particularly those based on the Mixture-of-Experts architecture, offer a powerful new tool for probing the complexities of human language and its disorders.

LLM Unit Ablation Simulating Aphasia via BLiMP

Scientists engineered a novel pipeline to simulate aphasia within large language models (LLMs), systematically perturbing functional components to replicate language-production impairments following brain lesions. The study employed two transformer LLMs, a dense baseline (OLMo) and its Mixture-of-Experts (OLMoE) variant, to enable architecture-independent comparison and robust evaluation. To ensure comparability, interventions operated on shared functional components: layer, expert pairs in OLMoE and hidden dimensions within the Feedforward Neural Network (FFN) for OLMo. Researchers first attributed fine-grained linguistic phenomena to individual units using the BLiMP probing suite, ablating each unit by zeroing its output and recording the resulting accuracy change, ∆t(u) = Acct(u) − Acct, to quantify causal importance.

In OLMoE, single experts were ablated while maintaining the router and other experts, and in the dense OLMo baseline, FFN hidden dimensions were treated as parallel neuron groups undergoing the same output-zeroing intervention. This approach yielded a task-specific attribution map, ∆t, for each BLiMP subtask, providing a foundation for subsequent alignment and thresholding analyses. The work then identified phenotype-linked units by fine-tuning both models on Broca and Wernicke subsets of the AphasiaBank dataset, ranking units by their training-time contribution using gradient-based statistics aggregated at the unit level. This process allowed the team to establish a link between specific units and the clinical manifestation of each aphasia subtype.

A crucial step involved selecting a stable top-p% threshold via a p-sweep, ensuring consistent and reliable perturbation across models and conditions. Subsequently, scientists aligned phenotype-linked units with linguistic phenomena using a rank-percentile heatmap, validated by an external CAP classifier, to confirm subtype separability. This progressive lesioning schedule, targeting 0.5%, 1.0%, 1.5%, and 2.0% of top units, enabled a matched comparison between MoE and dense models, revealing that subtype-targeted perturbations yielded more systematic, aphasia-like regressions than size-matched random perturbations.

LLMs Simulate Aphasia with Targeted Lesioning

Scientists achieved a breakthrough in simulating aphasia using large language models (LLMs), demonstrating the potential for these models to act as computational simulators of language cognition. The research team developed a clinically grounded framework to systematically manipulate LLMs, reproducing language-production impairments characteristic of aphasia following focal brain lesions. This innovative approach introduces a pipeline that identifies subtype-linked components within both modular Mixture-of-Experts (MoE) models and dense architectures, utilising a unified intervention interface. Experiments revealed that subtype-targeted perturbations consistently yield more aphasia-like regressions compared to size-matched random perturbations, highlighting the precision of the simulation.

Data shows that progressive impairment yields graded, clinically measurable declines on WAB subtests and AQ, demonstrating a strong correlation between model perturbation and aphasic symptoms. Measurements confirm that modular LLMs, particularly MoE architectures, support more localized and interpretable mappings between language phenotypes and underlying components. The study applied the framework to OLMoE-1B-7B-0924-Instruct and a dense baseline (OLMo), operating on shared functional components, layer-expert pairs in OLMoE and hidden dimensions in OLMo’s Feedforward Neural Network (FFN). Researchers attribute fine-grained linguistic phenomena to these units using BLiMP, then identified phenotype-linked units through AphasiaBank fine-tuning signals, validating subtype separability with an external CAP classifier.

Tests prove that a stable top-p% threshold was established via a p-sweep, and a rank-percentile heatmap aligned phenotype-linked units with linguistic phenomena. The breakthrough delivers a scalable experimental platform for modelling aphasia, enabling controlled studies of how distinct language functions degrade under targeted disruptions. The team’s methodology involved progressively lesioning the top-ranked units and quantifying clinical degradation using Functional Profiling and Phenotype, Phenomenon Alignment, ultimately providing a standardized target for assessing subtype-specific impairments.

LLMs Model Aphasia Through Targeted Disruption of language

Scientists have demonstrated that large language models (LLMs) can be systematically altered to replicate language impairments seen in aphasia following brain injury. Researchers introduced a clinically informed framework to simulate aphasia by selectively disrupting functional components within LLMs, applying this to both modular and dense models using a unified approach. The findings reveal that targeted disruptions produce more consistent, aphasia-like regressions compared to random disruptions, and modular LLMs offer clearer connections between specific components and observed language deficits. Broca-targeted lesions primarily reduced language production quality, while Wernicke-targeted lesions more often affected semantic coherence, these distinct effects were observed in both modular and dense models. The authors acknowledge that attributing specific functions to individual units is more challenging in dense models, resulting in less localized component signatures. Future research could explore the potential of this platform for testing rehabilitation strategies and further investigating the functional organisation of language.