A Unified Single-Pass Decision Model for EEG
Ask an EEG window a question in plain language. One shared model returns a typed answer with probabilities: a choice among candidates, the probability that a statement holds, or a score on an ordered scale.
Four trained questions play below on recorded EEG. Dashed cards are the next questions.
One recording, end to end, with a saved answer for every window.
Question
Choice answer · a probability for each of the five sleep stages in this 30-s epoch. Noul answer · Model P(seizure overlap): the probability that an annotated seizure overlaps this 5-s window. Choice answer · a probability for each motor-imagery task offered in the question. Score answer · a probability for each rating from 0 to 10; the expected value is read from it.
Amber line: annotated onset · purple box: detail selection · purple trace: selected lead
Unplayed signal stays hidden.
A continuous excerpt of up to the last 20 seconds, following playback. The purple box on the main chart marks the same lead and time range. The onset line appears only while its time is in this excerpt. All displayed leads are available in recording order; no lead is designated as the seizure focus. One voltage scale is shared across leads.
The dataset provides seizure start and end times, not labels for individual waves or onset channels. This replay uses saved model answers; an answer appears only after its full 5-s window has played.
Recorded replaySaved model outputData & methods
02 · Window by window
Curated examples from other recordings, with EEG traces, dataset annotations and saved model answers.
EEG: voltage (µV) over time. The axis is seconds within this window.
Loading recorded material…
How it works
A frozen REVE-L encoder turns the EEG window into features, and a trainable MLP connector maps them into the embedding space of Qwen3.5-2B, whose frozen base weights are adapted by LoRA. The EEG, the question and the answer specification pass through the model once. At a learned readout marker, the readout selected by the answer type returns probabilities; no text is generated.
The research
EEG-Jev is evaluated on selected NeuralBench datasets, not on the whole benchmark. Data & methods