Meta’s Non‑Invasive Brain‑to‑Text Decoder Hits New Accuracy Milestone

Background

Brain‑computer interfaces (BCIs) are classified as invasive (e.g., Neuralink) or non‑invasive. Invasive approaches implant micro‑electrode arrays directly into cortex, achieving decoding accuracies above 90 % but requiring surgery. Non‑invasive approaches record brain activity from sensors placed on the scalp, avoiding surgery but suffering from signal attenuation and lower spatial resolution, which historically limited decoding performance.

Brain2Qwerty v2 Overview

Meta released Brain2Qwerty v2, an end‑to‑end non‑invasive BCI that decodes complete sentences directly from magnetoencephalography (MEG) recordings. Compared with Brain2Qwerty v1, which decoded one character at a time with a character error rate (CER) of ~32 %, v2 shifts to word‑level decoding and outputs full sentences in a single step.

Experimental Results

Data were collected from nine volunteers, each providing roughly 2,400 sentences while typing on a keyboard while wearing a MEG helmet (≈10 h of recording per participant, ≈22,000 sentences total). The system achieved:

Average word‑accuracy of 61 % across participants.

Best‑subject word‑accuracy of 78 %.

More than 50 % of decoded sentences deviated by at most one word from the ground‑truth.

These results represent an ~8 % absolute improvement in word accuracy over previously reported non‑invasive decoding methods.

Model Architecture

The pipeline consists of two main components:

Brain Encoder : Raw MEG waveforms are processed by a Conformer module for character‑level detection, then passed to an Aligner that builds word‑level embeddings. The Encoder compresses MEG signals using a BrainModule that maps sensor coordinates to 270 virtual channels, applies per‑subject affine compensation, followed by four layers of dilated convolution (hidden size 1,500) with 4× temporal down‑sampling, and four Conformer layers (dimension 1,024, 4 heads). Supervision uses CTC loss with heads attached after down‑sampling and after the Conformer to provide gradients at shallow layers.

NeuroLLM : A Qwen3‑4B large language model fine‑tuned with LoRA (rank 128) receives both CTC‑derived text tokens and “neural word embeddings”. Neural embeddings are obtained by mean‑pooling Conformer features within CTC‑identified word boundaries, passing them through an MLP, and aligning them to LLM word‑embedding space with a SigLIP contrastive loss (Hard DTW alignment + sigmoid BCE). During inference the LLM consumes both streams, with 10 % random token dropout applied to each stream for robustness, and generates the final sentence using beam search (beam 16).

Training Procedure

Hyper‑parameter optimization was performed by three independent AI coding agents (Cursor + Claude Opus 4.6). Each agent operated in its own Git worktree, launching 50 SLURM training jobs per round for 10 rounds (1,500 experiments total). Engineers manually inspected the agents’ results and selected the best configuration.

Open‑Source Resources

All training code for v1 and v2 is publicly available at https://github.com/facebookresearch/brain2qwerty. The v1 dataset is hosted on Hugging Face at https://huggingface.co/datasets/bcbl190626/SpanishBCBL. The original Nature Neuroscience paper can be accessed at https://www.nature.com/articles/s41593-026-02303-2, and the accompanying blog post is at

https://ai.meta.com/blog/brain2qwerty-brain-ai-human-communication/

.

Remaining Challenges

Decoding accuracy remains insufficient for everyday communication; word‑level and character‑level errors persist.

The MEG system used is large, expensive, and requires SQUID‑based cryogenic cooling and magnetic shielding, making bedside or portable deployment infeasible. Emerging optically pumped magnetometer (OPM)‑based portable MEG systems may eventually reduce size and cost, but are still in early development.

The researchers observed a near log‑linear relationship between decoding accuracy and training data size, suggesting that scaling the dataset could further close the performance gap with invasive BCIs without changing the model architecture.