Reproduce “Sparse Autoencoders Map Brain-LLM Alignment onto Cortical Semantic Topography” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Dongxin Guo, Jikun Wu, Siu Ming Yiu
Researchers found that Sparse Autoencoders (SAEs) decompose large language model (LLM) representations into interpretable features, revealing that the intermediate-layer advantage in brain-LLM alignment is driven by rich semantic content. These SAE-derived semantic features formally align with established cortical semantic topography, showing a Spearman correlation of 0.72 with a priori neuroanatomical predictions.
Key insights
- The intermediate-layer advantage in brain-LLM alignment is primarily driven by rich semantic features, which emerge and peak in these layers, accounting for a significant portion of explained neural variance.
- SAE-derived semantic features accurately recapitulate known cortical semantic organization, providing a granular, feature-level mapping that strongly converges with *a priori* neuroscientific accounts.
- The methodology offers a novel framework for using mechanistic interpretability tools to generate and test specific neuroscientific hypotheses about human language processing, bridging AI and cognitive neuroscience.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.