Language Impairment as a Lens onto Large Language Models

Key points

• Although large language models (LLMs) encode language in highly distributed spaces, they remain vulnerable to focal disruptions---a phenomenon strikingly similar to how localized brain damage causes clinical aphasia in humans.1
• To investigate this theoretical tension, we ask whether LLMs possess a functional structure analogous to the human language system, by introducing a novel mechanistic interpretability framework inspired by aphasia.1
• Furthermore, while steering generalizes robustly across downstream tasks, structural lesioning remains asymmetrically constrained.1
• By bridging neuropsychology and mechanistic interpretability, we demonstrate that LLMs, functionally analogous to biological brains, rely on a densely entangled architecture that can be systematically mapped through clinically-informed interventions.1

Sources (1)

• [1]Language Impairment as a Lens onto Large Language Models
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 11:42 PM

  Through two complementary approaches, classifier-guided lesioning and continuous activation steering, we systematically induce aphasia-like breakdowns in LLMs. We find that language functions can be successfully isolated to discrete structural circuits and continuous representational vectors, yet these functions remain densely entangled with significant circuit overlap and vector similarities acro

  Although large language models (LLMs) encode language in highly distributed spaces, they remain vulnerable to focal disruptions---a phenomenon strikingly similar to how localized brain damage causes clinical aphasia in humans.

Extractive summary: sentences quoted from the sources.