Activation-Guided Neuron Intervention to Induce Alzheimer's-Related Computational Language Phenotypes in a Large Language Model

2026-08-03

AI Summary

This study investigates whether Alzheimer’s disease (AD)-related linguistic representations in large language models exert a functional causal influence on cognitive-like behaviors, rather than merely serving as passive correlates. Leveraging the Qwen3-8B model, the authors identify feedforward neurons more strongly activated by AD patient corpora and perform activation-guided neuron editing by scaling their downstream projection weights, thereby generating multiple controlled output variants. Through systematic evaluation combining blinded human assessments and computational linguistic metrics, they demonstrate that amplifying AD-associated neurons significantly impairs performance on tasks such as story recall and verbal fluency, reducing linguistic complexity and information density—effectively recapitulating core features of human AD language phenotypes. Conversely, suppressing these neurons partially restores performance. This work establishes a controllable causal experimental framework linking linguistic anomalies to cognitive impairment.

Abstract

Changes in spontaneous speech provide an early signal of cognitive dysfunction in Alzheimer's disease (AD) that large language models (LLMs) can detect. However, detection alone cannot establish whether the underlying model representations contribute functionally to behavior. We introduce an activation-guided intervention framework using Qwen3-8B. The framework identifies feed-forward neurons with higher activation rates for AD than control transcripts and modulates their output contributions during generation by scaling the corresponding down-projection weights. This yielded nine edited variants differing in intervention direction, magnitude, and scope. The original and edited models completed the same 12-turn neuropsychological battery, assessed through blinded human ratings and computational linguistic measures. Amplifying AD-associated neurons produced graded impairments in story recall, verbal fluency, working memory, procedural discourse, scene construction, and coreference resolution. Attenuation largely preserved performance and selectively improved several outcomes. Amplification also reduced lexical surprisal, idea density, syntactic complexity, and discourse quantity, broadly paralleling changes reported in human AD speech. These findings show that neurons identified solely from clinical language differences can influence behavior across multiple cognitive domains, providing proof of concept for an AD-related computational phenotype and a controlled framework for experimentally examining links between language and broader cognitive dysfunction.

Problem

• Alzheimer's disease
• large language models
• neuron intervention
• computational phenotype
• language impairment

Innovation

• activation-guided intervention
• Alzheimer's-related computational phenotype
• neuron modulation
• large language model
• cognitive dysfunction