Generative causal testing to bridge data-driven models and scientific theories in language neuroscience

A new framework called generative causal testing (GCT) can produce concise explanations of how the brain processes language, using large language models to bridge the gap between opaque predictive models and formal scientific theory, according to a paper posted to arXiv [1]. The framework, detailed in a submission last revised on 13 Jun 2026, addresses a core limitation in neuroimaging research: while representations from large language models (LLMs) predict BOLD fMRI responses to language stimuli with high accuracy, the features driving those responses in specific brain areas have remained unclear [1]. GCT generates explanations of language selectivity and then tests them in follow-up experiments using stimuli produced by the LLM itself [2]. The authors report that the method successfully explains selectivity in individual voxels and in cortical regions of interest, including newly identified microROIs in the prefrontal cortex [2]. Explanatory accuracy, they find, is closely tied to the predictive power and stability of the underlying models [1]. The work also shows that GCT can dissect fine-grained functional differences between brain areas that exhibit similar selectivity profiles [2]. This capability marks a step toward resolving long-standing questions about the division of linguistic labor across the cortex. The paper argues that LLMs can serve as a conduit between purely data-driven models and the kind of formal, testable theories that have traditionally been harder to extract from high-dimensional neural networks [1]. LLMs are a class of machine learning model trained with self-supervised learning on vast text corpora to perform natural language processing tasks such as generation [10]. Their internal representations have become a dominant tool in computational neuroscience, even though current neural networks are generally regarded as low-quality models of biological brain function [3]. The GCT framework attempts to move beyond correlation by introducing a causal testing step, generating stimuli designed to confirm or refute a candidate explanation [2]. The paper appeared on arXiv, the preprint server that, through collaborations such as arXivLabs, now integrates interactive demos from platforms like Hugging Face Spaces directly alongside paper abstract pages [6][7]. This integration allows readers to explore open-source demos without writing code, a feature aimed at increasing the reproducibility and accessibility of machine learning research [6][8]. The GCT paper's submission history shows three versions, with the file size growing from 28,458 KB in the initial October 2024 upload to 40,253 KB by the June 2026 revision [1]. The lead author is Richard Antonello [1].