Interpreting Brain Responses to Language with Sparse Features from Language Models

A new encoding framework that replaces dense language-model hidden states with hierarchically-organized sparse features can interpret brain responses to language and suggests a nontrivial correspondence between brain and model representations, researchers report. The study, posted to arXiv on June 5, introduces Augmented Sparse Encoding Models, which project dense residual-stream language-model features into a sparse, hierarchically-organized sparse autoencoder basis and explicitly include surprisal — a measure of processing difficulty — as a predictor [1][3]. The work addresses a common criticism that studies relating biological and artificial representations merely relate one black box to another [2]. The framework was tested on a high-field 7T fMRI dataset collected from eight participants who listened to 200 linguistically diverse sentences [1][2]. The researchers first validated their approach by recovering previously known interpretations of voxel populations tuned to processing difficulty and meaning abstractness [1][3]. They then interpreted a previously-uncharacterized but reliable voxel population and found it is tuned to people-related content [2][3]. The team also examined the fronto-temporal human language network and found it is predicted by a common set of features across its constituent regions, though frontal regions were relatively well-explained by surprisal alone, even without language-model-based features [1][3]. A central finding is that brain responses during language processing are best explained by the features that capture the most general information encoded in language-model representations, rather than by an arbitrary set of features [1][2]. This result suggests a nontrivial correspondence between brain and language-model language representation [1][3]. The work builds on a growing body of research using sparse autoencoders to bridge mechanistic interpretability and neural encoding. A separate study found that semantic features extracted by sparse autoencoders from GPT-2 XL and Llama-3.1-8B recovered 94 percent of peak encoding performance and revealed a systematic cortical topography where specific semantic subcategories — concreteness, affect, social cognition, spatial relations, and event structure — mapped onto distinct brain regions in a pattern that matched predictions from three independent neuroscience programs [4]. That study reported a Spearman correlation of 0.72 and a hypergeometric p-value of 0.007 for the convergence test [4]. Other recent work has used explainable AI attribution methods to quantify how preceding words influence next-word predictions in language models, finding that attributions from early model layers align with early-stage language processing regions such as the auditory cortex and superior temporal gyrus, while deeper layers correspond to higher-order regions including the inferior frontal gyrus and angular gyrus [5]. The broader theoretical context for such findings includes predictive coding, a theory of brain function which postulates that the brain constantly generates and updates a mental model of the environment to predict sensory input [7]. The new Augmented Sparse Encoding Models framework, by explicitly incorporating surprisal alongside sparse features, provides a more interpretable bridge between these predictive processing theories and the internal representations of large language models [1][3].