Scaling Next-Brain-Token Prediction for MEG

2026-01-28

AI Summary

This study addresses the challenge of long-context modeling and generation of source-space magnetoencephalography (MEG) signals across diverse datasets and MEG acquisition systems. To this end, we develop a large-scale autoregressive model based on the Qwen2.5-VL architecture, incorporating a modified SEANet-style vector quantizer to compress multichannel MEG into flat token sequences. Trained on over 500 hours of data spanning thousands of sessions, the model recursively generates minute-long neural signals. We innovatively extend large language model architectures to neurophysiological signal generation and introduce a comprehensive evaluation framework for long-range MEG synthesis, assessing manifold stability, condition specificity, and cross-dataset generalization. On the MOUS test set, our generated signals maintain stable long-horizon rollouts and significantly outperform prompt-swap baselines, demonstrating both efficacy and robust generalization.

Technology Category

Machine Learning: Deep Generative Models & AutoencodersNatural Language Processing: GenerationCognitive Modeling & Cognitive Systems: Neural Spike Coding

Application Category

Web Mining and Content Analysis: Large pretrained models with web dataSearch and Retrieval-Augmented AI: Large language models for searchSocial Networks and Social Media: Generative AI / large language models and their impact on social systems

Abstract

We present a large autoregressive model for source-space MEG that scales next-token prediction to long context across datasets and scanners: handling a corpus of over 500 hours and thousands of sessions across the three largest MEG datasets. A modified SEANet-style vector-quantizer reduces multichannel MEG into a flattened token stream on which we train a Qwen2.5-VL backbone from scratch to predict the next brain token and to recursively generate minutes of MEG from up to a minute of context. To evaluate long-horizon generation, we introduce task-matched tests: (i) on-manifold stability via generated-only drift compared to the time-resolved distribution of real sliding windows, and (ii) conditional specificity via correct context versus prompt-swap controls using a neurophysiologically grounded metric set. We train on CamCAN and Omega and run all analyses on held-out MOUS, establishing cross-dataset generalization. Across metrics, generations remain relatively stable over long rollouts and are closer to the correct continuation than swapped controls. Code available at: https://github.com/ricsinaruto/brain-gen.

Problem

• MEG
• next-token prediction
• long-context modeling
• cross-dataset generalization
• brain signal generation

Innovation

• next-token prediction
• vector-quantized MEG
• long-context autoregressive modeling
• cross-dataset generalization
• neurophysiologically grounded evaluation