A Roadmap for MEG Foundation Models
Philipp Thölke, Hamza Abdelhedi, Yorguin Mantilla-Ramos, Fouad Lbakali, Oumayma Gharbi, Catherine Duclos, Annalisa Pascarella, Vanessa Hadid, et al.
q-bio.NC · cs.AI
Quick verdict
A roadmap for MEG foundation models: thorough and honest about open questions, but purely qualitative—no new experiments, self-referential coverage.
Summary
This perspective article maps the emerging field of MEG foundation models and proposes a research roadmap. The authors catalog current MEG-specific models (MEG-GPT, Headache-MEG-FM, GPT2MEG) and MEG-inclusive multimodal models (BrainOmni, Brain-OF, Ferrante et al.) in a comparison table covering data scale, architecture, pretraining objective and tokenization. They structure five development pathways — native MEG pretraining, transfer from EEG models, continued pretraining on MEG, cross-modal joint training, and adaptation of generic time-series models — and analyze design choices in tokenization, sensor- versus source-space representation, positional/sensor-geometry encoding, self-supervised objectives, and pretraining data. On evaluation, they review the PNPL/LibriBrain competitions (over 50 hours from one subject in 2025; about 104 hours across 33 subjects in 2026) and NeuralBench (two MEG tasks), specifying requirements for benchmarks that can actually test the foundation-model claim. Key observations: MEG pretraining corpora remain modest; larger EEG models have not reliably outgeneralized smaller ones; JEPA-style latent objectives, spectral reconstruction and contrastive predictive coding lack MEG instances; and no multimodal model has been ablated against its uni-modal counterpart. Limitations: purely qualitative with no new experiments, heavily reliant on arXiv preprints, no systematic literature-search methodology, and several surveyed models and benchmarks involve the authors themselves. A curated public dataset list accompanies the paper.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- MEG warrants a dedicated foundation-model program because millisecond temporal resolution combined with weak skull distortion (closer cortical source link) supports source-resolved questions that EEG and fMRI cannot address alone.
- No single development pathway dominates: MEG-native pretraining, EEG-model transfer or continued pretraining, cross-modal joint training and generic time-series adaptation each trade MEG-specificity against data and compute, so the choice must follow the scientific question.
- Current MEG foundation models are pretrained on single datasets, so corpus breadth, composition and curation — not raw volume alone — are the binding constraint on generalization.
- Existing benchmarks (PNPL/LibriBrain speech tasks; NeuralBench's two MEG tasks) cannot yet test the foundation-model claim; valid evaluation requires open diverse data, participant-level exclusion from pretraining, standardized probing protocols and matched task-specific baselines.
- High decoding accuracy does not establish scientific value; MEG foundation models must be shown to reveal verifiable latent structure in cortical dynamics beyond what classical analyses already expose.
Strengths
Systematic landscape catalog
Table I compares seven MEG-relevant efforts (MEG-GPT, Headache-MEG-FM, BrainOmni, Brain-OF, Ferrante et al., GPT2MEG, Jayalath et al.) across data scale, architecture, objective, tokenization and downstream evaluation — a consolidated view of a fast-moving niche, backed by a public companion dataset list (github.com/thecocolab/meg-fm-resources).
Epistemic honesty about open questions
Frames EEG-to-MEG transfer as 'an empirical hypothesis rather than an inherently advantageous starting point' with explicit negative-transfer risk (§II-A Strategy 2), and flags that Défossez et al.'s geometry-aware attention was never isolated experimentally (§II-B) — rare candor for a roadmap.
Benchmark critique with operational teeth
Precise gap identification
Didactic scaffolding with checkable numbers
Weaknesses
No new empirical contribution
Undisclosed author overlap with surveyed work
Preprint-heavy evidence base
No systematic review methodology
Roadmap breadth without prioritization or cost
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a perspective/roadmap article (arXiv 2609.04461v2, q-bio.NC), not an empirical study: it contributes synthesis, taxonomy and argument rather than experiments. Stated aims (§I): situate MEG within brain-foundation-model work, survey early MEG-specific and MEG-inclusive systems (Table I), analyze design choices (§III), review benchmarks (§IV), and lay out open problems (§V). Central claims: (1) MEG merits a dedicated foundation-model program because millisecond resolution plus weak skull distortion supports source-resolved questions neither EEG nor fMRI answers alone (§I-C); (2) no development pathway dominates — MEG-native pretraining, EEG-model transfer, continued pretraining, cross-modal joint training and generic time-series adaptation trade MEG-specificity against data and compute (§II); (3) corpus breadth and composition, not just volume, is the binding constraint (§III-E, §V-A); (4) current benchmarks cannot yet test the foundation-model claim (§IV); (5) decoding accuracy alone does not establish scientific value (§V-E).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is a MEG foundation model?
A model pretrained with self-supervised objectives on large collections of magnetoencephalography (MEG) recordings — rather than trained for one task — whose learned representations can then be probed or fine-tuned for many downstream uses such as speech decoding, cognitive-state classification or clinical inference. Examples in this 2026 perspective include MEG-GPT (autoregressive, trained on Cam-CAN) and multimodal models like BrainOmni and Brain-OF that include MEG alongside EEG/fMRI.
Do MEG foundation models already exist?
Yes, but at an early stage. The paper catalogs a handful: MEG-GPT (Cam-CAN, ~612 subjects), a clinically oriented Headache-MEG-FM (~416 participants), precursors like GPT2MEG, plus MEG-inclusive multimodal models (BrainOmni, Brain-OF, Ferrante et al.). Most are pretrained on a single dataset, corpora are modest compared with EEG, and standardized benchmarks are only beginning to appear (PNPL/LibriBrain competitions, NeuralBench).
Can EEG foundation models be reused for MEG?
Possibly, but the authors stress this is an unproven hypothesis, not a default advantage: EEG and MEG differ in sensor geometry, spatial sensitivity, signal scale and noise, so direct transfer risks negative transfer. Early evidence is encouraging — in NeuralBench, the EEG-pretrained REVE scored best among tested models on a MEG typing task — but no rigorous matched comparison against MEG-native pretraining exists. The paper recommends continued pretraining of EEG models on MEG as the most practical near-term route.
What benchmarks exist for evaluating MEG foundation models?
The PNPL competitions built on LibriBrain (2025: over 50 hours from one subject, speech detection and phoneme classification; 2026: ~104 hours across 33 subjects, word classification and between-subject evaluation) and NeuralBench, whose MEG part currently has only two tasks (THINGS-MEG image decoding and a keystroke dataset available on request). The paper argues these are too narrow and that proper benchmarking needs open diverse data, participant-level exclusion from pretraining, and matched baselines.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- No truncation detected: the extraction runs from the title page through Acknowledgments and reference [45]; body text, Table I, Table II and the Fig. 1 caption are all present.
- Only the visual rendering of Fig. 1 (the pathway diagram) is absent from the text extraction; its content is fully described in the caption and §II, so no judgment in this review depends on it.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.