BCIwiki (bciwiki.com) — A collaboration involving The Chinese University of Hong Kong, Shenzhen, and Harbin Institute of Technology, Shenzhen, China, posted OneEEG to arXiv on October 26, 2026. The EEG multimodal language-model framework is paired with the OneEEG-Instruct training dataset and OneEEG-Bench. Evaluation uses closed-set questions with finite candidate answers.
OneEEG jointly models numerical and visual EEG representations and is trained on the team's curated OneEEG-Instruct multi-task EEG instruction-tuning dataset. It introduces a task-aware Mixture of Projectors (MoP) to align heterogeneous EEG features from different task families with the large language model's embedding space, the internal space used to represent and process information.
The team also established OneEEG-Bench to evaluate existing multimodal large language models and OneEEG on multi-task EEG understanding. Experiments showed that the existing models evaluated had limited ability to interpret EEG inputs. OneEEG achieved stronger performance by integrating numerical and visual EEG representations with task-aware alignment.
Ablation experiments and further analyses examined EEG visual renderings, MoP variants, router behavior and feature dispersion across EEG tasks. The work focuses on model understanding and evaluation of EEG inputs; these findings do not establish clinical diagnostic capability or effectiveness in brain-computer interface (BCI) applications.
OneEEG uses numerical EEG and waveform images for tasks mapped to the standard 19-channel setting, but only images for non-standard channel settings. The benchmark covers motor, state, clinical, event, sleep and vigilance tasks plus basic EEG attributes. Data with participant identifiers were split by participant.
The full text reports 65.67% Overall and 68.24% Macro accuracy on this closed-set benchmark, compared with 50.39% Overall for GPT-5.5, the strongest baseline. OneEEG received EEG-specific training, while baselines received only waveform images and text questions. Differences in training and input capabilities limit interpretation as a general model ranking.