What happens when you damage a language model's brain

Aphasia—selective language loss from brain damage—provides a window into how human brains organize language function. This work adapts that clinical framework to language models by zeroing out parameters and measuring the resulting impairments using the Text Aphasia Battery, a diagnostic tool. Testing five 1B-scale models revealed that while all aphasia symptom types can emerge, their distributions differ markedly from human patterns. The study found that attention and feed-forward components produce distinct symptom profiles, and that lesioning early layers causes syntactic/semantic deficits while middle-late layers produce phonological and fluency problems. The qualitative differences between LM and human aphasia suggest that symptom patterns arise from learning and architecture details rather than being universal consequences of language disruption.