Sparse Autoencoders Map Brain-LLM Alignment onto Cortical Semantic Topography

• URL: http://arxiv.org/abs/2605.23035v1
• Date: Thu, 21 May 2026 21:00:24 GMT
• Title: Sparse Autoencoders Map Brain-LLM Alignment onto Cortical Semantic Topography
• Abstract summary: We show that semantic features alone recover 94% of peak encoding performance.<n>We also show that SAE-discovered features recapitulate known cortical semantic organization.<n>Results generalize across English, Chinese, and French.
• Score: 13.891522069967507
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: 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 ($κ\geq 0.74$) 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 $ρ=0.72$, $p<0.001$; 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 ($Δ\mathrm{logLik}=38.4$, $p<0.001$), and an exploratory prediction-error analysis provides preliminary evidence that the brain additionally encodes unexpected semantic content. Results generalize across English, Chinese, and French.

Related papers

• PhysSAE: Mechanistic Interpretability of PINNs with Sparse Autoencoders [0.0]
  PhysSAE is a mechanistic interpretability framework that trains overcomplete sparse autoencoders on PINN penultimate-layer activations.<n>We show that PINNs develop sparse, physically structured latent representations that can be identified and causally interrogated post-hoc.
  arXiv  Detail & Related papers  (2026-09-07T05:32:03Z)
• Learn from your own latents and not from tokens: A sample-complexity theory [53.5821824211418]
  We study networks trained to predict their emphown latent representations of related views or masked regions.<n>We show that latent prediction achieves this with a number of samples emphconstant in $L$, up to logarithmic factors.<n>This suggests that explicit stacking such as H-JEPA is largely redundant.
  arXiv  Detail & Related papers  (2026-05-26T22:16:42Z)
• Language Model Circuits Are Sparse in the Neuron Basis [50.460651620833055]
  We show that textbfMLP neurons are as sparse a feature basis as SAEs.<n>This work advances automated interpretability of language models without additional training costs.
  arXiv  Detail & Related papers  (2026-01-30T05:41:19Z)
• Dynamic Large Concept Models: Latent Reasoning in an Adaptive Semantic Space [56.37266873329401]
  Large Language Models (LLMs) apply uniform computation to all tokens, despite language exhibiting highly non-uniform information density.<n>We propose $textbfDynamic Large Concept Models (DLCM)$, a hierarchical language modeling framework that learns semantic boundaries from latent representations and shifts from tokens to a compressed concept space where reasoning is more efficient.
  arXiv  Detail & Related papers  (2025-12-31T04:19:33Z)
• Neural Networks Learn Generic Multi-Index Models Near Information-Theoretic Limit [66.20349460098275]
  We study the gradient descent learning of a general Gaussian Multi-index model $f(boldsymbolx)=g(boldsymbolUboldsymbolx)$ with hidden subspace $boldsymbolUin mathbbRrtimes d$.<n>We prove that under generic non-degenerate assumptions on the link function, a standard two-layer neural network trained via layer-wise gradient descent can agnostically learn the target with $o_d(1)$ test error.
  arXiv  Detail & Related papers  (2025-11-19T04:46:47Z)
• Real-time nonlinear inversion of magnetic resonance elastography with operator learning [0.06797079068199119]
  The oNLI framework enables real-time MRE inversion (30,000x speedup) of elastograms with comparable spatial accuracy to NLI.<n>A structural prior mechanism, analogous to Soft Prior Regularization in the MRE literature, was incorporated to improve spatial accuracy.
  arXiv  Detail & Related papers  (2025-10-03T08:55:40Z)
• Information Locality as an Inductive Bias for Neural Language Models [52.92279412466086]
  We show that $m$local entropy are difficult for Transformer and LSTM LMs to learn languages.<n>These results suggest that neurals are highly sensitive to the statistical structure of a language.
  arXiv  Detail & Related papers  (2025-06-05T15:21:05Z)
• Probing Neural Topology of Large Language Models [12.298921317333452]
  We introduce graph probing, a method for uncovering the functional connectivity of large language models.<n>By probing models across diverse LLM families and scales, we discover a universal predictability of next-token prediction performance.<n>Strikingly, probing on topology outperforms probing on activation by up to 130.4%.
  arXiv  Detail & Related papers  (2025-06-01T14:57:03Z)
• Explanations of Large Language Models Explain Language Representations in the Brain [5.7916055414970895]
  We propose a novel approach using explainable AI (XAI) to strengthen link between language processing and brain neural activity.<n>Applying attribution methods, we quantify the influence of preceding words on predictions.<n>We find stronger attributions suggest brain alignment for assessing the biological explanation methods.
  arXiv  Detail & Related papers  (2025-02-20T16:05:45Z)
• Interpretable Next-token Prediction via the Generalized Induction Head [59.500195503897764]
  Generalized Induction-Head Model (GIM) is an interpretable model for next-token prediction.<n>In language modeling, GIM improves next-token prediction by up to 25%p over interpretable baselines.<n>In an fMRI setting, GIM improves neural response prediction by 20%.
  arXiv  Detail & Related papers  (2024-10-31T12:33:26Z)
• Residual Learning of Neural Text Generation with $n$-gram Language Model [41.26228768053928]
  We learn a neural LM that fits the residual between an $n$-gram LM and the real-data distribution. Our approach attains additional performance gains over popular standalone neural models consistently.
  arXiv  Detail & Related papers  (2022-10-26T02:42:53Z)
• Long-range and hierarchical language predictions in brains and algorithms [82.81964713263483]
  We show that while deep language algorithms are optimized to predict adjacent words, the human brain would be tuned to make long-range and hierarchical predictions. This study strengthens predictive coding theory and suggests a critical role of long-range and hierarchical predictions in natural language processing.
  arXiv  Detail & Related papers  (2021-11-28T20:26:07Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.