Language Impairment as a Lens onto Large Language Models
Muxuan He · Wang Bill Zhu · Yue Chen · Elsi Kaiser · Weizhi Du · Yao Du · Samuel Nastase · Yanjun Gao
Uses classifier-guided lesioning and activation steering derived from human aphasic transcripts to induce and map Broca-like and Wernicke-like breakdowns across 11 language models.
Mapping the brain signature of dyslexia risk onto histamine, acetylcholine and glutamate receptor systems
M. C. Celik · Y. Cicek · M. A. Sakiroglu · D. Aslan · A. T. Altunc · H. A. Velioglu
Links polygenic structural-brain signatures of dyslexia risk to normative neurotransmitter maps and reading skill, identifying a testable biological axis for impairment.
Are Linguistic and Arithmetic Circuits Dissociated in LLMs?
Riku Kisako · Tatsuki Kuribayashi · Ryohei Sasano
Uses neurolinguistic dissociation as a hypothesis for mechanistic interpretability and finds largely separate linguistic and arithmetic circuits across five LLMs.