{"slug": "sparse-autoencoders-map-brain-llm-alignment-onto-cortical-semantic-topography", "title": "Sparse Autoencoders Map Brain-LLM Alignment onto Cortical Semantic Topography", "summary": "Researchers decomposed GPT-2 XL and Llama-3.1-8B into 16,000 to 32,000 interpretable features per layer using sparse autoencoders, finding that semantic features alone recovered 94% of peak brain-response encoding performance. The semantic subcategories mapped onto distinct cortical regions with high statistical significance, and the features predicted human reading times beyond lexical controls. The findings, which generalized across English, Chinese, and French, provide a mechanistic explanation for why intermediate LLM layers best predict human brain responses to language.", "body_md": "arXiv:2605.23035v1 Announce Type: new\nAbstract: Intermediate layers of large language models (LLMs) best predict human brain responses to language, one of the most robust findings in computational neurolinguistics, yet why remains mechanistically unexplained. We address this gap by bridging sparse autoencoders (SAEs) from mechanistic interpretability with neural encoding models, decomposing GPT-2 XL and Llama-3.1-8B into 16K-32K interpretable features per layer. A human-validated taxonomy ($\\kappa \\geq 0.74$) reveals that semantic features alone recover 94% of peak encoding performance ($r=0.285$), substantially exceeding variance-matched baselines ($p<0.001$, $d=1.31$). Beyond this aggregate dominance, we test a novel cortical topography prediction: five semantic subcategories derived a priori from three independent neuroscience programs should map onto distinct brain regions. A formal convergence test confirms this alignment (Spearman $\\rho=0.72$, $p<0.001$; hypergeometric $p=0.007$), demonstrating that SAE-discovered features recapitulate known cortical semantic organization at a granularity inaccessible to prior methods. SAE features further predict human reading times beyond lexical controls ($\\Delta\\mathrm{logLik}=38.4$, $p<0.001$), and an exploratory prediction-error analysis provides preliminary evidence that the brain additionally encodes unexpected semantic content. Results generalize across English, Chinese, and French.", "url": "https://wpnews.pro/news/sparse-autoencoders-map-brain-llm-alignment-onto-cortical-semantic-topography", "canonical_source": "https://arxiv.org/abs/2605.23035", "published_at": "2026-05-25 04:00:00+00:00", "updated_at": "2026-05-25 15:26:21.874582+00:00", "lang": "en", "topics": ["large-language-models", "natural-language-processing", "neural-networks", "artificial-intelligence", "machine-learning"], "entities": ["GPT-2 XL", "Llama-3.1-8B"], "alternates": {"html": "https://wpnews.pro/news/sparse-autoencoders-map-brain-llm-alignment-onto-cortical-semantic-topography", "markdown": "https://wpnews.pro/news/sparse-autoencoders-map-brain-llm-alignment-onto-cortical-semantic-topography.md", "text": "https://wpnews.pro/news/sparse-autoencoders-map-brain-llm-alignment-onto-cortical-semantic-topography.txt", "jsonld": "https://wpnews.pro/news/sparse-autoencoders-map-brain-llm-alignment-onto-cortical-semantic-topography.jsonld"}}