A Dynamical Systems Equivalence Model of Brain and Transformer-Based Language Models

This repository hosts the code, analysis notebooks, and supplementary materials for the study "Brain-Transformer Equivalence: A Dynamical Systems Equivalence Model of Brain and Transformer-Based Language Models." The research bridges cognitive neuroscience and AI mechanistic interpretability by evaluating how closely the internal representations of Large Language Models (LLMs) align with human brain activity (EEG) during naturalistic reading. The study demonstrates that while standard dense LLM representations hit a low biological alignment ceiling, extracting sparse, monosemantic features via Sparse Autoencoders (SAEs) yields a 4.3x improvement, effectively saturating the EEG noise ceiling. This project provides the framework and scripts to reproduce these layer-by-layer Representational Similarity Analysis (RSA) and linear encoding results, supporting the hypothesis that both biological and artificial networks rely on sparse distributed coding for semantic processing.