Sparse Autoencoders Reveal How LLM Representations Align with Human Brain Semantics

Summary

A mechanistic interpretability study published on arXiv demonstrates that sparse autoencoders (SAEs) can decompose large language model representations into interpretable semantic features that closely align with human cortical organization. Researchers applied SAEs to GPT-2 XL and Llama-3.1-8B, extracting 16K-32K interpretable features per layer and creating a human-validated semantic taxonomy (Cohen's kappa ≥ 0.74). Remarkably, semantic features alone recovered 94% of peak neural encoding performance during natural language comprehension tasks (r = 0.285), substantially outperforming variance-matched baselines (p < 0.001, d = 1.31).

The research connects LLM internals to human neuroscience through neural encoding analyses of fMRI responses during naturalistic reading. A formal cortical topography convergence test revealed strong alignment between SAE-derived semantic features and distinct brain regions (Spearman rho = 0.72, p < 0.001; hypergeometric p = 0.007). The authors report that SAE features predict human reading times beyond traditional lexical controls (delta log-likelihood = 38.4, p < 0.001) and demonstrate cross-linguistic generalization across English, Chinese, and French, validating the universality of the discovered semantic structure.

This work establishes mechanistic interpretability as a bridge between model internals and cognitive neuroscience, offering a methodological pathway for understanding how language models represent meaning in ways that parallel human brain organization.

• Cross-linguistic generalization across English, Chinese, and French suggests discovered semantic axes reflect universal language representation principles
• SAE features predict human reading times and capture prediction-error signals for unexpected semantic content beyond lexical baselines

Editorial Opinion

This research demonstrates that sparse autoencoders unlock interpretability at a granularity that maps directly onto neuroscientific data—a significant advance for mechanistic interpretability. If independently replicated on additional participant cohorts and datasets, the cortical topography findings would establish a fundamental connection between model internals and human language processing. The cross-linguistic generalization particularly strengthens the claim that these semantic structures reflect universal organizational principles, positioning interpretability methods as tools not just for model debugging but for cognitive science itself.