Research map / Architectures and efficiency / Attention and transformers
MEG-XL: Data-Efficient Brain-to-Text via Long-Context Pre-Training
Dulhan Jayalath, Oiwi Parker Jones
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MEG-XL long-context pre-training instead of LaBraM at low data | MEG-XL · MEG-MASC (13% data) | top-10 balanced accuracy: 33.2 -> 47.0 | Improved |
| Use MEG-XL instead of best baseline at low data | MEG-XL · LibriBrain (13% data) | top-10 balanced accuracy: 40.3 -> 57.3 | Improved |
| Use MEG-XL instead of BrainOmni with full data | MEG-XL · Armeni (100% data) | top-10 balanced accuracy: 62.3 -> 61.2 | No clear effect |
| Use MEG-XL instead of BrainOmni with full data on shallow multi-subject data | MEG-XL · MEG-MASC (100% data) | top-10 balanced accuracy: 19.1 -> 46.4 | Improved |
| Pre-train MEG-XL versus random initialisation (ablation) | MEG-XL · word decoding across data amounts | top-10 balanced accuracy | Improved |
| Pre-trained MEG-XL versus supervised SOTA on MEG-MASC | MEG-XL · MEG-MASC | top-10 balanced accuracy: over 25% at certain points | Improved |
| Pre-trained MEG-XL versus supervised SOTA on Armeni | MEG-XL · Armeni | top-10 balanced accuracy: roughly 10% above until around 15 hours | Improved |
| Pre-trained MEG-XL versus supervised SOTA on deep single-subject LibriBrain | MEG-XL · LibriBrain | top-10 balanced accuracy: similar until 2.5 hours, then supervised ahead | Hurt |
| Increase pre-training context length (linear probing) (ablation) | MEG-XL · word decoding linear probe | top-10 balanced accuracy: diminishing returns after 100 seconds | Improved |
| Increase pre-training context length (zero-shot masked prediction) (ablation) | MEG-XL · unseen datasets masked token prediction | token prediction accuracy: log-linear trend through 150s | Improved |
| Provide longer inference context than pre-training context (ablation) | MEG-XL · word decoding linear probe | top-10 balanced accuracy: small difference between full and matched context | No clear effect |
| Resample to 100Hz (Nyquist-compliant) with 75s context (ablation) | MEG-XL · LibriBrain (100% data) | top-10 balanced accuracy: 63.0 -> 58.6 | Hurt |
Closest papers by meaning
- LibriBrain100: One Hundred Hours of Broad and Deep MEG Data for Neural Speech Decoding at Scale (NeurIPS 2026 · similarity 0.90)
- The Brain's Bitter Lesson: Scaling Speech Decoding With Self-Supervised Learning (ICML 2025 · similarity 0.88)
- Du-IN: Discrete units-guided mask modeling for decoding speech from Intracranial Neural signals (NeurIPS 2024 · similarity 0.86)
- BrainOmni: A Brain Foundation Model for Unified EEG and MEG Signals (NeurIPS 2025 · similarity 0.86)
- LibriBrain: Over 50 Hours of Within-Subject MEG to Improve Speech Decoding Methods at Scale (NeurIPS 2025 · similarity 0.86)
- NeuroLM: A Universal Multi-task Foundation Model for Bridging the Gap between Language and EEG Signals (ICLR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
