Assessing Cognitive Performance in Language Models: A Framework for Cognitive Impairment Simulation
Author(s)
Publisher
Springer Nature Switzerland
Date issued
November 17, 2026
In
Lecture Notes in Computer Science
Progress in Artificial Intelligence
Number of pages
13
Serie
Lecture Notes in Computer Science; 17198
Subjects
Language Models Cognitive Assessment Cognitive Impairment Alzheimer’s Disease Brain Alignment Model Alteration
Abstract
We introduce a framework for assessing the cognitive performance of language models as a baseline for in silico Alzheimer’s impairment simulation. We adapt the Montreal Cognitive Assessment into structured prompts and pairwise comparison metrics capturing a model’s ability to prefer correct over wrong responses. Applying neuron ablation and margin-loss fine-tuning across four models spanning 82M to 7B parameters, we measure co-evolving changes in cognitive scores and brain score, the alignment of model representations with human fMRI recordings. Our results reveal an asymmetric impairment profile: attention and memory degrade faster than abstraction and language. Brain score trajectories are heterogeneous and model-size-dependent, with no simple mapping between cognitive decline and neural misalignment. These findings suggest that cognitive and neural signals capture partially independent
dimensions of model deterioration, and that our framework offers a controlled setting for systematic evaluation toward eventual clinical calibration.
dimensions of model deterioration, and that our framework offers a controlled setting for systematic evaluation toward eventual clinical calibration.
Event name
EPIA 2026
Location
Madeira, Portugal
ISSN
0302-9743
1611-3349
Publication type
conference paper
File(s)![Thumbnail Image]()
Loading...
Name
Assessing Cognitive Performance in Language.pdf
Type
Main Article
Size
473.06 KB
Format
Adobe PDF
