AutoBCI: Forecast-Guided Agentic Neural Architecture Discovery for EEG-Based Brain--Computer Interfaces

Muyun Jiang; Yi Ding; Wei Zhang; Jinbo Chen; Chenyu Liu; Zhenjie Yang; Yuxin Li; Jingyuan Chen

cs.AISep 28, 2026

AutoBCI: Forecast-Guided Agentic Neural Architecture Discovery for EEG-Based Brain--Computer Interfaces

Authors: Muyun Jiang, Yi Ding, Wei Zhang, Jinbo Chen, Chenyu Liu, Zhenjie Yang, Yuxin Li, Jingyuan Chen, +4 more

Organizations: Nanyang Technological University, Singapore · The University of Hong Kong, Hong Kong, China · Southeast University, China

Abstract

EEG-based brain-computer interfaces support a broad range of applications, yet designing decoding architectures that perform well across diverse tasks remains challenging. We introduce AutoBCI, an agentic framework in which a Designer Agent and a Forecaster Agent support the discovery and selection of EEG decoding architectures across tasks. The Designer Agent performs Pool-Guided Architecture Discovery (PGAD), generating and refining architectures through training and validation across multiple EEG tasks, such as emotion recognition, motor imagery, and sleep staging. The Forecaster Agent performs Performance Estimation from Early Knowledge (PEEK), using architecture code, the training protocol, and early learning curves to predict full-budget validation performance and select promising candidates for continued training. Across 14 EEG datasets spanning motor imagery, emotion recognition, and sleep staging, we evaluate AutoBCI with six LLMs, including Opus 5.5 and GPT 5.6 Sol, and compare the architectures selected by the search procedure against ten baselines: six conventional EEG models and four foundation models. The architecture discovered by AutoBCI with Claude Opus 5.5 achieves 64.16% average test balanced accuracy (bAcc), compared with 63.87% for REVE, the strongest baseline on this metric. Using ten observed epochs, PEEK reduces mean absolute error in predicting average validation bAcc from 2.20 to 1.36 percentage points, a 38.1% reduction relative to the best-observed-score baseline.

Figures & tables

Figure 2: PGAD candidate generation protocol. Left: Architectures evaluated in previous discovery rounds. Middle: A shared candidate pool ranked by combined validation performance. Right: Top-ranked parents provide source code, learning curves, and task-specific validation scores for targeted refinements. These refinements are combined with fresh architecture proposals to form the next candidate batch.

Task	No.	Datasets	Unified input	Classes
Motor Imagery	1 2 3 4 5 6 7 8	BCIC-IV-2a ( Tangermann et al., 2012 ) OpenBMI-MI ( Lee et al., 2019 ) BCIC-Upperlimb ( Jeong et al., 2022 ) Cho2017 ( Cho et al., 2017 ) HighGamma ( Schirrmeister et al., 2017 ) PhysioNet-MI ( Schalk et al., 2004 ) SHU-MI ( Ma et al., 2022 ) Shin2017A ( Shin et al., 2016 )	65 ch $\times$ 4 s	Left, Right, Foot, Tongue, Cylin, Sphe, Lumbrical (7 classes)
Emotion Recognition	9 10 11	SEED ( Duan et al., 2013 ) SEED-IV ( Zheng et al., 2018 ) SEED-V ( Liu et al., 2021 )	65 ch $\times$ 4 s	Positive, Neutral, Negative, Sad, Fear, Happy, Disgust (7 classes)
Sleep Staging	12 13 14	Sleep-EDF ( Kemp et al., 2000 ) ISRUC ( Khalighi et al., 2016 ) HMC ( Alvarez-Estevez and Rijsman, 2021 )	8 ch $\times$ 30 s	Wake, N1, N2, N3, REM (5 classes)

Table 1: EEG datasets, task-specific inputs, and class counts.

Model tier	Provider	LLM	Agent interface	Input (USD/1M)	Output (USD/1M)
Flash	Anthropic	Claude Sonnet 5	Claude Code	2.00	10.00
	DeepSeek	DeepSeek V4.1 Flash	Claude Code	0.30	1.20
	Google	Gemini 3.8 Flash	Gemini CLI	0.75	3.75
	Alibaba	Qwen3.8 Flash	Claude Code	0.15	0.47
Flagship	OpenAI	GPT-5.6 Sol	Codex CLI	4.00	20.00
	Anthropic	Claude Opus 5.5	Claude Code	4.00	20.00

Table 2: LLMs and execution settings for architecture discovery.

Model	Spec	Motor Imagery			Emotion			Sleep			Average
		bAcc	wF1	Kappa	bAcc	wF1	Kappa	bAcc	wF1	Kappa	bAcc	wF1	Kappa
EEGNet ( Lawhern et al., 2018 )	8.8K	57.83	57.12	0.233	37.97	34.48	0.164	69.09	70.60	0.640	54.96	54.07	0.345
TSception ( Ding et al., 2022 )	17.5K	54.58	53.63	0.167	43.93	45.91	0.256	66.75	70.74	0.648	55.09	56.76	0.357
DCN ( Schirrmeister et al., 2017 )	317K	66.70	66.36	0.396	42.73	38.03	0.215	73.89	76.68	0.697	61.10	60.35	0.436
FAST ( Jiang et al., 2026 )	475K	58.59	58.31	0.243	35.39	36.94	0.137	72.26	74.99	0.686	55.41	56.75	0.355
EEG-Conformer ( Song et al., 2022 )	2.13M	66.32	66.21	0.396	48.08	48.92	0.302	72.34	76.07	0.701	62.25	63.73	0.466

Table 3: Performance comparison of small-model baselines, foundation models, and AutoBCI-discovered architectures. Results are evaluated on held-out test sets for motor imagery, emotion recognition, and sleep staging. bAcc and wF1 are percentages. Per column: first , second , third .

Figure 3: Discovery trajectories across six LLM-guided searches. Dots show candidates completed on all three tasks; dashed lines mark round medians, and solid lines track the best combined validation bAcc so far. Labels give the incumbent scores. The vertical axis is truncated at 50%; round-by-round architecture identities appear in Supplementary Figure 6 .

Model	Round-best declines	Older-parent rounds	Older-parent link in champion lineage
Claude Opus 5.5	None	R3	Yes ( $1\rightarrow 3$ )
Claude Sonnet 5	R4 (0.221 pp)	R3, R4, R5, R6	Yes ( $2\rightarrow 5$ )
DeepSeek V4.1 Flash	R2 (0.030 pp)	R3, R4, R6	Yes ( $1\rightarrow 3$ )
Gemini 3.8 Flash	None	R3	None (fresh)
GPT-5.6 Sol	None	R3	Yes ( $1\rightarrow 3$ )
Qwen3.8 Flash	None	R5	None (fresh)

Table 4: Historical retention and parent reuse in recorded searches. Declines compare a round’s best combined validation bAcc with the prior incumbent. Older-parent rounds use at least one parent from before the immediately preceding round. Arrows identify parent-to-child generation rounds in the champion’s ancestry.

Figure 4: Early forecasts and candidate retention. (A–B) Ten-epoch forecasts versus actual combined validation bAcc for 269 architectures, averaging task-wise best-within-40 scores. (C–D) Top-one and top-two task-winner retention across observation windows; dark curves average the three tasks equally. Dotted lines mark ten epochs.

Figure 5: Forecast error across LLMs and observation windows. Combined validation bAcc MAE (pp), averaged over candidates within rounds and equally over paired rounds. Cohorts may vary by window; dotted lines mark ten epochs.

	Opus candidates		Sonnet candidates
Forecaster	Top-1	Top-2	Top-1	Top-2
Best observed score	50.0	83.3	50.0	55.6
Self	50.0	77.8	50.0	72.2
DeepSeek V4.1 Flash	50.0	77.8	50.0	66.7
GPT-5.6 Sol	44.4	77.8	50.0	66.7
Gemini 3.8 Flash	44.4	83.3	55.6	61.1

Table 5: Selection across Designer and Forecaster models. Top-1/Top-2 task-winner retention (%) at ten epochs, averaged equally over three tasks and six rounds per generator. Self denotes the generating model.

Method	Winner retention	Full runs	Task-epochs	Epoch savings	GPU job-hours
Full training	100.0	807	32,280	0.0	68.58
PEEK ( $k=1$ )	61.1	108	11,310	65.0	24.52
PEEK ( $k=2$ )	83.3	216	14,550	54.9	31.86
PEEK ( $k=3$ )	91.7	324	17,790	44.9	38.85
PEEK ( $k=4$ )	94.4	429	20,940	35.1	45.53
PEEK ( $k=5$ )	97.2	534	24,090	25.4	52.18

Table 6: Winner retention and estimated screening cost. Epoch-10 forecasts retrospectively select up to

k

candidates per round across 36 rounds. Completed 40-epoch runs provide ground truth; cost assumes only selected candidates continue. Retention and epoch savings are percentages.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Architecture discovery with round incumbents. Dots show individual evaluated candidates; dashed and solid lines show round medians and cumulative best bAcc. Each inset names the incumbent after rounds R1–R6; bold text marks the final architecture.

Task	Dataset	Train	Validation	Test
Emotion	SEED	Trials 1–9	Trials 10–12	Trials 13–15
	SEED-IV	Trials 1–16	Trials 17–20	Trials 21–24
	SEED-V	Trials 1–5	Trials 6–10	Trials 11–15
MI	BCI Competition IV-2a	Subjects 0–6	Subjects 0–6	Subjects 7–8
	BCI Upper Limb	Subjects 0–10	Subjects 0–10	Subjects 11–14
	Cho2017	Subjects 0–39	Subjects 0–39	Subjects 40–48

Appendix

Table 7: Dataset split construction. Emotion trial numbers are one-based and apply within each subject and session; subjects occur in all three splits. MI subject ranges are zero-based positions in HDF5 group iteration order, not original subject IDs. MI training and validation use approximately 80% and 20% of examples from the same development subjects, respectively; test subjects are held out. Sleep entries give numbers of mutually disjoint subjects.

Task	Dataset	Classes	Train	Validation	Test
Emotion	SEED (3 classes)	3	22,545	7,905	7,620
	SEED-IV	4	26,025	5,340	6,210
	SEED-V	5	8,512	10,672	9,984
MI	BCI Competition IV-2a	4	3,148	788	1,152
	BCI Upper Limb	3	2,640	660	1,200
	Cho2017	2	6,464	1,616	1,800

Appendix

Table 8: Dataset composition and split sizes. Counts refer to EEG examples, including 30-second epochs for sleep. Classes are the retained local classes in each dataset. Emotion and MI inputs have shape

65\times 800

; sleep inputs have shape

8\times 3000

Setting	Value
Full training budget	40 epochs per task
Batch size	512
Optimizer	Fused AdamW
Initial learning rate	$10^{-3}$
Weight decay	$10^{-4}$
Adam coefficients $(\beta_{1},\beta_{2})$	$(0.9,0.999)$

Appendix

Table 9: Common candidate-training settings. The same settings apply to emotion, MI, and sleep in all six completed searches.

Model	Precision	GFLOPs	Time (s)	MiB	Test bAcc (%)
EEGNet	FP32	0.126	$0.650\pm 0.001$	644	$55.47\pm 0.69$
	BF16 mixed	0.126	$0.819\pm 0.009$	371	$54.48\pm 0.90$
	BF16 full	0.126	$0.810\pm 0.006$	345	$53.79\pm 1.11$
DeepConvNet	FP32	0.328	$0.380\pm 0.008$	1336	$58.91\pm 3.82$
	BF16 mixed	0.328	$0.427\pm 0.001$	1099	$57.96\pm 3.04$
	BF16 full	0.328	$0.423\pm 0.001$	1083	$55.90\pm 2.91$

Appendix

Table 10: Training precision on BCI Competition IV-2a. Means and standard deviations over three seeds. GFLOPs are counted per training example (forward and backward); time is per training epoch; memory is mean peak allocation. Test bAcc uses validation-selected checkpoints.

Explore similar work

EvoBrain: Continual Learning of EEG Foundation Models Across Heterogeneous BCI Tasks
Jun 1, 2026Yangxuan Zhou, Sha Zhao, Jiquan Wang +2Brain-Computer Interface Electroencephalography Foundation Models
State Key Laboratory of Brain-machine Intelligence and College of Computer Science and Technology, Zhejiang University, Hangzhou, Zhejiang, China · MOE Frontier Science Center for Brain Science and Brain-Machine Integration, Zhejiang University, Hangzhou, Zhejiang, China
ThinkNet: Compact Architecture Selection and Validation-Gated Ensembles for Subject-Independent MI-EEG Decoding
Sep 27, 2026Abdul Basit, Saim Rehman, Muhammad ShafiqueBrain-Computer Interface Controlled Comparison
eBRAIN Lab, Division of Engineering New York University (NYU) Abu Dhabi, Abu Dhabi, UAE
STEAM: A Spatio-TEmporal Alignment Mixture-of-Experts Model with Hierarchical Pre-training for EEG Decoding
Aug 3, 2026Zhu Chen, Dingkun Liu, Yuheng Chen +1Electroencephalography Foundation Models Electroencephalography
Ministry of Education Key Laboratory of Image Processing and Intelligent Control, School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan 430074, China · Zhongguancun Academy, Beijing 100094, China

Jun 1, 2026, cs.AIJ/K move · Enter open · S save

EvoBrain: Continual Learning of EEG Foundation Models Across Heterogeneous BCI Tasks

Yangxuan Zhou, Sha Zhao, Jiquan Wang, Shijian Li+1

State Key Laboratory of Brain-machine Intelligence and College of Computer Science and Technology, Zhejiang University, Hangzhou, Zhejiang, China · MOE Frontier Science Center for Brain Science and Brain-Machine Integration, Zhejiang University, Hangzhou, Zhejiang, China

Brain-Computer Interface Electroencephalography Foundation Models Electroencephalography Replay-Based Continual Learning Continual Learning

Open paper AskPDF arXiv

Electroencephalography (EEG) is the cornerstone of non-invasive brain-computer interfaces (BCIs), yet conventional decoding relies on fragmented, task-specific architectures that severely limit cross-task scalability. While EEG foundation models pre-trained on massive corpora promise universal brain decoding, current post-training depends on task-isolated fine-tuning. This static paradigm restricts knowledge transfer across heterogeneous tasks, hinders model scalability, and incurs computational and storage overheads that scale linearly with task count. To overcome these bottlenecks, we formulate downstream adaptation as a cross-task continual learning problem and propose EvoBrain, a dynamic, task-aware continual learning framework for unified EEG decoding. EvoBrain addresses the plasticity-stability trade-off via two complementary components: (1) Neuro-Spectral Task Normalization (NSN) aligns incoming tasks with historical statistics while recalibrating spectral responses to handle distributional and neuro-spectral shifts; and (2) Response-Affinity Distillation (RAD), combined with time-dependent replay, preserves old-task response geometry and promotes selective knowledge transfer between spectrally compatible tasks, effectively mitigating forgetting. Extensive evaluations across six distinct BCI tasks demonstrate that EvoBrain consistently surpasses state-of-the-art methods across diverse foundation backbones, optimally balancing plasticity and stability. To our knowledge, this work pioneers cross-task continual learning in the EEG domain, advancing the realization of a unified, one-for-all brain decoding system.

Abstract

Explore similar work

EvoBrain: Continual Learning of EEG Foundation Models Across Heterogeneous BCI Tasks

ThinkNet: Compact Architecture Selection and Validation-Gated Ensembles for Subject-Independent MI-EEG Decoding

STEAM: A Spatio-TEmporal Alignment Mixture-of-Experts Model with Hierarchical Pre-training for EEG Decoding