2026-09-28 00:00 Papers Foundations & Methods Source language EN

AutoBCI Uses AI Agents to Design EEG Decoders That Edge Past the Best Existing Model

Summary Researchers have built AutoBCI, a framework in which AI agents design EEG decoding models, and report that it slightly outperformed the strongest existing model across 14 datasets. Building a decoder that works well across many EEG tasks usually takes repeated trial and error. AutoBCI splits the job between 2 agents: one generates and refines architectures, while the other forecasts final performance from early learning curves to pick out candidates worth training further. Across 14 EEG datasets covering motor imagery, emotion recognition and sleep staging, the architecture found with Claude Opus 5.5 reached an average test balanced accuracy of 64.16%, slightly above the 63.87% of REVE, the strongest baseline. Looking at only 10 training epochs, the forecasting agent cut prediction error from 2.20 to 1.36 percentage points. The study is a preprint and has not been peer reviewed.

Why it matters Benchmarked against 6 conventional models and 4 foundation models, the 0.29-point accuracy lead is marginal, so the more practical payoff may be the compute saved by judging candidates from early learning curves rather than training each one to completion.

BCIwiki (bciwiki.com) — An agentic framework called AutoBCI designed an EEG decoding architecture that reached 64.16% average test balanced accuracy across 14 datasets, slightly ahead of the 63.87% posted by REVE, the strongest baseline, according to a preprint posted to arXiv on September 28, 2026. The study has not been peer reviewed.

AutoBCI pairs 2 agents. A Designer Agent runs Pool-Guided Architecture Discovery (PGAD), generating and refining candidate architectures by training and validating them on tasks including emotion recognition, motor imagery and sleep staging. A Forecaster Agent runs Performance Estimation from Early Knowledge (PEEK), using the architecture code, the training protocol and early learning curves to predict full-budget validation performance and choose which candidates to keep training.

The researchers ran AutoBCI with 6 large language models, including Opus 5.5 and GPT 5.6 Sol, and compared the selected architectures with 10 baselines: 6 conventional EEG models and 4 foundation models. The 64.16% result came from the run using Claude Opus 5.5.

After observing 10 epochs, PEEK cut the mean absolute error in predicting average validation balanced accuracy from 2.20 to 1.36 percentage points, a 38.1% reduction relative to a best-observed-score baseline.

The paper lists 12 authors, including Muyun Jiang, Yi Ding and Cuntai Guan, and runs to 34 pages with 6 figures, including supplementary material.

Compiled by BCIwiki from public sources

Sources · 1

arxiv.org 2026-09-28

Decoding Algorithms EEG Non-invasive BCI

Read original ↗

Suggest a correction Revisions · none 中 / EN