EEG foundation model that aligns brain signals with language, decoding BCI tasks by matching EEG embeddings to text label prototypes.
No providers recorded yet. Browse all providers
LEAF (Language-EEG Aligned Foundation Model) is an electroencephalography (EEG) foundation model for brain-computer interfaces that ties EEG representations to natural language. It was developed by Cuntai Guan and Yi Ding's group at Nanyang Technological University, with co-authors at Shanghai Jiao Tong University, University of Science and Technology Beijing and Southeast University. The first preprint version (arXiv, September 2025) introduced the model under the name ELASTIQ (EEG-Language Alignment with Semantic Task Instruction and Querying); the March 2026 revision renamed it LEAF and revised the evaluation.
The problem LEAF targets is how EEG models see labels. A motor-imagery dataset encodes "left hand" as class 0 and an emotion dataset encodes "happy" as class 0, and a conventional decoder, including the fine-tuned heads of EEG foundation models such as LaBraM, EEGPT and CBraMod, treats those integers as meaningless indices. LEAF replaces the integers with text: every class name is embedded by a frozen sentence encoder, and the EEG encoder is trained to land near the embedding of the correct label, conditioned on an instruction that names the task.
This differs from NeuroLM, which discretizes EEG into tokens, pushes them into a language model's vocabulary and decodes by generation. LEAF keeps the language model outside the network entirely. Instruction and label embeddings are computed once offline, so the deployed model is a compact EEG encoder that scores trials against text prototypes.
The EEG tokenizer cuts each trial into non-overlapping 0.5-second windows and applies a temporal convolution followed by a spatial convolution across all 65 channels. Each branch of the pretraining tower is a 12-layer transformer with 256-dimensional tokens and eight attention heads; the Q-Former has four layers. Self-supervised pretraining used nine unlabeled datasets totaling about 1,153 hours, among them Stieger2021, ChineseEEG, Chisco and ThinkOutLoud. The pretrained tower and Q-Former were then instruction-tuned jointly on the combined training splits of the downstream datasets, using all-mpnet-base-v2 as the frozen text encoder, which outperformed BERT, E5 and Qwen3 embedders in ablations. Training ran on four H100 GPUs.
Evaluation covers 16 datasets in five task families: eight motor-imagery sets, five emotion sets (the SEED series and FACED), SSVEP, covert speech and mental workload. Splits are cross-subject where participant counts allow and cross-trial otherwise, with balanced accuracy and Cohen's kappa reported over three seeds. After per-dataset fine-tuning, LEAF ranked first on 12 of 16 datasets against task-specific networks and the BIOT, EEGPT, LaBraM and CBraMod foundation models, and had the best average in every task family. The widest margin was on the SSVEP, covert-speech and workload group, at 71.36% balanced accuracy, more than 4.7 points above the next method. In the instruction ablation, average balanced accuracy rose from 47.4% with no instruction to 50.0% with task and target instructions. Deliberately wrong instructions reorganized the embedding space toward the stated task instead of the true one.
LEAF suits BCI and affective-computing groups that decode several paradigms (motor imagery for rehabilitation or device control, emotion recognition, SSVEP spellers, imagined speech, workload monitoring) and want one pretrained encoder rather than a model per task. Because labels are text, a new dataset can be added by writing its label names and a task instruction into the configuration, and the released preprocessing pipeline maps new electrode layouts onto the model's montage.
LEAF brings language into the EEG representation itself rather than into a decoder, and at 26 million parameters it is a lightweight alternative to generative EEG-language models such as NeuroLM. Its ablations show that task descriptions measurably restructure EEG embedding spaces, supporting language as a shared interface across heterogeneous BCI datasets. The code, both checkpoints and a documentation site are public under a non-commercial license from NTU's Centre for Brain Computing Research. All results are from a preprint that is under review, and all evaluation is offline decoding on public datasets.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.