Brain2Qwerty v2 was trained on approximately 22,000 sentences collected from nine volunteer participants. Each participant spent around 10 hours wearing a magnetoencephalography (MEG) device while actively typing, allowing the system to learn how raw brain signals correspond to written language. Rather than relying on manually engineered methods to identify neural events, the new pipeline applies end-to-end deep learning directly to the recorded brain activity.
According to Meta, the system combines multiple AI techniques to improve decoding quality. Large language models are fine-tuned using neural data so the decoder can use semantic context when reconstructing text from noisy brain recordings. The company also said AI agents were used to explore potential improvements to the decoding pipeline before engineers manually selected the final training configuration.
Meta reports that Brain2Qwerty v2 achieves an average word accuracy of 61%, compared with 8% for other non-invasive approaches cited by the company. For its highest-performing participant, the system reached 78% word accuracy, with more than half of decoded sentences containing one word error or less.
The researchers also observed that decoding performance improved as more training data was added, following what Meta describes as a log-linear relationship between data volume and accuracy. Based on those findings, the company says additional data alone could continue reducing the remaining performance difference between non-invasive systems and approaches that require surgery.
Beyond the new decoder, Meta positions Brain2Qwerty v2 as part of a broader effort to develop open foundational models for neuroscience. The company said the work complements projects including Tribev2 for perception encoding, NeuralSet for large-scale brain data processing, and NeuralBench for evaluating brain models. It also highlighted its recently announced $5 million Digital Brain Project fund, which is intended to encourage the creation of open neuroscience datasets.
Meta said it hopes that releasing its code and collaborating openly with the research community will help accelerate efforts to identify, diagnose, and treat neurological disorders while advancing technologies that could eventually restore communication for people affected by brain lesions.
This analysis is based on reporting from Meta.
Image courtesy of Meta Platforms.
This article was generated with AI assistance and reviewed for accuracy and quality.
