BCIwiki (bciwiki.com) — Authors affiliated with Tongji University, Ant Group, Shanghai Jiao Tong University and Shanghai Yangzhi Rehabilitation Hospital introduced NeuroAct, an EEG research agent. In an experience-reuse experiment, BCIC-IV-2a balanced accuracy rose from 34.66% to 50.81% while large language model token use fell 64.02%. The preprint was uploaded to arXiv on October 18, 2026 and has not been peer reviewed.
The framework uses two loops. Research, development and review roles test candidate changes using evidence from signals, model representations, predictions and training. An outer loop turns results into recommendations with explicit task and model conditions, then tests and revises them on later tasks.
Four evaluation settings were reported. General EEG analysis scored 80.88 on a macro-average measure, 17.75 points above BrainAgent’s 63.13. Cross-task optimization raised EEGPT’s mean balanced accuracy by 7.34 percentage points and HBN Pearson correlation from 0.6894 to 0.7362. Single-task open research improved ADFTD balanced accuracy by 16.33 percentage points; collaborative prediction increased Workload balanced accuracy from 79.69% to 85.02%.
The experience-reuse experiment compared prediction performance and model-use cost. Tokens measure the text processed by an LLM, so the 64.02% reduction concerns text-processing volume in that setup. Broader transfer beyond the tested analysis and prediction tasks requires independent evaluation.