# EEG Foundation Models

6 related articles

Subscribe to updates

arXiv Nov 4, 2026
General vision-language models read EEG directly: Qwen3-VL hits 0.835 balanced accuracy on 4-second discharge detection without EEG pretraining
Clinical EEG reading has long depended on experts paging through recordings, while recent EEG foundation models first pretrain on large unlabeled EEG corpora. This work renders EEG segments as the bipolar montage waveforms or spectrograms clinicians already inspect, pairs each image with a fixed text prompt, and hands it to a general vision-language model, Qwen3-VL, with no EEG pretraining at all. On the 4-second interictal epileptiform discharge task vEpiSet, Qwen3-VL-4B adapted with LoRA reaches 0.835 balanced accuracy against 0.763 for fully fine-tuned BIOT; on the TUEV EEG-Bench split, Qwen3-VL-8B reaches 0.750 against 0.739 for the second-best single-task model.
Clinical EEG reading has long depended on experts paging through recordings, while recent EEG foundation models first pretrain on large unlabeled EEG corpora. This work renders EEG segments as the bipolar montage waveforms or spectrograms clinicians already inspect, pairs each image with a fixed text prompt, and hands it to a general vision-language model, Qwen3-VL, with no EEG pretraining at all. On the 4-second interictal epileptiform discharge task vEpiSet, Qwen3-VL-4B adapted with LoRA reaches 0.835 balanced accuracy against 0.763 for fully fine-tuned BIOT; on the TUEV EEG-Bench split, Qwen3-VL-8B reaches 0.750 against 0.739 for the second-best single-task model.
Binghamton University Harvard University
Xin Zhou（First author）
Vision-Language Models EEG Foundation Models Qwen3-VL
bioRxiv Nov 6, 2026
A frozen EEG foundation model needs no fine-tuning: one first-layer attention head's connectivity matrix reaches 42.13 balanced accuracy in cross-subject emotion recognition
The EEG foundation model REVE no longer needs full fine-tuning or all-layer pooled features to recognize emotion. Such models had not consistently beaten task-specific architectures on cross-subject tasks, and their large representations are costly to compute and store. This approach aggregates the first-layer attention weights of a single head over each trial into a weighted channel-to-channel matrix, then uses a label-free symmetry score to pick the most informative head; on SEED-IV, classical classifiers reach 42.13 balanced accuracy with about 176 times fewer calculations than all-layer attention pooling.
The EEG foundation model REVE no longer needs full fine-tuning or all-layer pooled features to recognize emotion. Such models had not consistently beaten task-specific architectures on cross-subject tasks, and their large representations are costly to compute and store. This approach aggregates the first-layer attention weights of a single head over each trial into a weighted channel-to-channel matrix, then uses a label-free symmetry score to pick the most informative head; on SEED-IV, classical classifiers reach 42.13 balanced accuracy with about 176 times fewer calculations than all-layer attention pooling.
University of Patras National University of Singapore
Λέανδρος Νικόλαος Αρβανιτόπουλος（First author）Christos A. Fidas（Corresponding author）
EEG Foundation Models Self-Attention Emotion Recognition Linear Probing
arXiv Oct 14, 2026
EEG foundation model fine-tuning sheds subject variation: content and style split, cross-subject decoding steadier
EEG foundation model fine-tuning gains a plug-and-play adapter, DEFT. Pretrained embeddings previously retained substantial subject-related variation, adding burden to the downstream classifier in cross-subject settings. DEFT splits pretrained tokens into a task-oriented content branch and a subject-consistent style branch, then predicts from the content branch, using recursive Gaussian refinement, temporal-variance regularization, subject-aware contrastive learning, and reconstruction to instantiate this factorize-before-classify design. Across multiple EEG datasets and four foundation model backbones it improves matched fine-tuning baselines, and ablations support the intended content/style branch organization.
EEG foundation model fine-tuning gains a plug-and-play adapter, DEFT. Pretrained embeddings previously retained substantial subject-related variation, adding burden to the downstream classifier in cross-subject settings. DEFT splits pretrained tokens into a task-oriented content branch and a subject-consistent style branch, then predicts from the content branch, using recursive Gaussian refinement, temporal-variance regularization, subject-aware contrastive learning, and reconstruction to instantiate this factorize-before-classify design. Across multiple EEG datasets and four foundation model backbones it improves matched fine-tuning baselines, and ablations support the intended content/style branch organization.
Nanyang Technological University Peking University Chinese Academy of Sciences
Yuqiu Deng（First author）
DEFT EEG Foundation Models Fine-tuning
arXiv Oct 9, 2026
Personalized EEG models of epilepsy and non-epilepsy show similar predictive fit but markedly different learned latent dependency structure
Using frozen representations from a lightweight CNN–Transformer EEG foundation model pretrained on TUEG, the study maps TUEP segments into a shared latent-state space and fits sparse multinomial logistic transition distributions independently per subject to obtain personalized latent transition-dependency graphs; it finds that epilepsy and non-epilepsy subjects are comparably predictable within-subject yet epilepsy models retain substantially denser learned dependency structure, with graph-derived features retaining moderate group discrimination.
Using frozen representations from a lightweight CNN–Transformer EEG foundation model pretrained on TUEG, the study maps TUEP segments into a shared latent-state space and fits sparse multinomial logistic transition distributions independently per subject to obtain personalized latent transition-dependency graphs; it finds that epilepsy and non-epilepsy subjects are comparably predictable within-subject yet epilepsy models retain substantially denser learned dependency structure, with graph-derived features retaining moderate group discrimination.
University of Illinois Urbana-Champaign
Rita Huan-Ting Peng（First author）
Multinomial Logistic Transition Distributions Latent Transition-Dependency Graph EEG Foundation Models Epilepsy
arXiv Oct 5, 2026
ZeroMAG generates multimodal adapters zero-shot from unlabeled calibration signals, raising balanced accuracy of frozen EEG foundation models by 7.22 points on average across six held-out datasets
ZeroMAG introduces a zero-shot multimodal adapter generation framework that, with a frozen EEG encoder and prediction head, uses only unlabeled target calibration recordings and task context to generate adapter weights through a configuration-invariant 1+N adapter, a modality–subject–task condition, and a function-constrained latent space learned from source adapters; across six held-out target datasets and three EEG foundation model backbones it improves balanced accuracy by 7.22 percentage points over EEG-only inference and 4.89 points over direct weight regression, coming within 0.50 points of supervised multimodal adaptation on average, with no target labels or target-side optimization.
ZeroMAG introduces a zero-shot multimodal adapter generation framework that, with a frozen EEG encoder and prediction head, uses only unlabeled target calibration recordings and task context to generate adapter weights through a configuration-invariant 1+N adapter, a modality–subject–task condition, and a function-constrained latent space learned from source adapters; across six held-out target datasets and three EEG foundation model backbones it improves balanced accuracy by 7.22 percentage points over EEG-only inference and 4.89 points over direct weight regression, coming within 0.50 points of supervised multimodal adaptation on average, with no target labels or target-side optimization.
Zhejiang University Stanford University National University of Singapore Nanyang Technological University Chinese Academy of Sciences University of Oxford Brown University
Yubo Wang（First author）
ZeroMAG Adapter Generation EEG Foundation Models Conditional Latent Diffusion
arXiv Sep 29, 2026
Treating EEG masking geometry as the only variable: 58 pre-trained models point to a moderate spatial radius with short temporal blocks, and expose a JEPA-specific bias-inflation collapse
The study unifies EEG self-supervised masking strategies into a three-parameter framework of spatial radius, temporal length and mask ratio, trains 58 models under a fixed REVE-Small backbone and corpus across MAE and JEPA, evaluates them with a linear probe on the 12 OpenEEGBench datasets, and finds that both frameworks agree on a moderate spatial radius with short temporal blocks as the optimum while identifying a JEPA-specific bias-inflation collapse at full-channel masking.
The study unifies EEG self-supervised masking strategies into a three-parameter framework of spatial radius, temporal length and mask ratio, trains 58 models under a fixed REVE-Small backbone and corpus across MAE and JEPA, evaluates them with a linear probe on the 12 OpenEEGBench datasets, and finds that both frameworks agree on a moderate spatial radius with short temporal blocks as the optimum while identifying a JEPA-specific bias-inflation collapse at full-channel masking.
Commissariat à l'Énergie Atomique et aux Énergies Alternatives University of California San Diego Radboud University Nijmegen Centre National de la Recherche Scientifique Université Paris-Saclay
Pierre Guetschel（First author）
Masking Strategy EEG Foundation Models Bias-Inflation Collapse Joint-Embedding Predictive Architecture OpenEEGBench

You’re all caught up

General vision-language models read EEG directly: Qwen3-VL hits 0.835 balanced accuracy on 4-second discharge detection without EEG pretraining

A frozen EEG foundation model needs no fine-tuning: one first-layer attention head's connectivity matrix reaches 42.13 balanced accuracy in cross-subject emotion recognition

EEG foundation model fine-tuning sheds subject variation: content and style split, cross-subject decoding steadier

Personalized EEG models of epilepsy and non-epilepsy show similar predictive fit but markedly different learned latent dependency structure

ZeroMAG generates multimodal adapters zero-shot from unlabeled calibration signals, raising balanced accuracy of frozen EEG foundation models by 7.22 points on average across six held-out datasets

Treating EEG masking geometry as the only variable: 58 pre-trained models point to a moderate spatial radius with short temporal blocks, and expose a JEPA-specific bias-inflation collapse