← News

Models & toolsSources: arXiv · 19 Nov 2026

RABBiT model predicts speech-evoked brain activity from fMRI with little participant data

Researchers introduce RABBiT, a compact audio-to-fMRI foundation model that predicts brain responses to natural speech with no or minimal participant-specific training data.

In plain words

A research team built an AI model called RABBiT that predicts which brain regions activate when a person listens to natural speech, using fMRI scans. It works without any data from a new participant, and improves with as little as 10 minutes of that person's scan data. The authors say this could remove the need for lengthy per-person model fitting in language-brain studies.

What was announced

RABBiT (Rapidly Adaptive BOLD foundation model via BraIn-Tuning) is described in an arXiv abstract as a compact encoder that maps audio to fMRI activity. The authors report that it predicts responses to natural speech across auditory and language-selective brain regions in a zero-shot setting, meaning without participant-specific training, and that it outperforms a state-of-the-art fMRI foundation model and predictions based on group averages. The evaluation covered 324 participants across multiple unseen fMRI datasets.

When participant-specific data is available, the model can be adapted with parameter-efficient tuning, and the authors state that roughly 10 minutes of such data substantially improves performance over per-participant linear models. They attribute the results to two design choices: learned region-specific attention and a decomposition separating shared from subject-specific response components, built on a brain-tuned speech backbone. The structured representations also support interpretability, and the code has been released on GitHub. These claims come from the paper's abstract and have not been independently verified.

What it does not show

  • The evidence is the paper's abstract only; full methods, benchmarks and caveats in the article were not available for review.
  • Performance figures, statistical tests and dataset details are not quantified in the abstract, so the size of reported improvements cannot be assessed.
  • The claims are self-reported by the authors and have not been independently confirmed.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.