Neural feature alignment between large language models and brain activities: A knowledge-based framework for...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Cognitive Science

Background:

• Large Language Models (LLMs) show representational alignment with brain activity.
• Existing studies lack detailed analysis of shared cognitive mechanisms between LLMs and the brain.
• Bridging artificial and biological intelligence requires interpretable frameworks.

Purpose of the Study:

• Introduce Feature Alignment (FA), an interpretable framework to map LLM representations to neural signatures.
• Enable systematic analysis of cross-modal correspondences between computational and biological intelligence.
• Provide a neuroscience-grounded methodology for evaluating AI cognitive plausibility.

Main Methods:

• Developed Feature Alignment (FA) to transform LLM eigenspace projections into semantically interpretable feature vectors.
• Mapped LLM features to cognitively linked electroencephalography (EEG) signatures.
• Conducted layer-wise and region-specific analyses across diverse LLMs and EEG features.

Main Results:

• FA scores showed strong predictive validity for LLM capabilities (r=0.736-0.886).
• FA revealed multi-scale correspondences between computational and neural knowledge processing.
• A decline in alignment was observed in deeper LLM layers, with distinct influences from computational factors like knowledge distillation.

Conclusions:

• FA provides an interpretable and extensible framework for characterizing shared representational primitives across AI and biological intelligence.
• The framework offers a knowledge-based approach to evaluate AI cognitive plausibility.
• FA can systematically guide the development of AI systems with more human-aligned knowledge representations.