NeuroAct: Advancing EEG Analysis and Decoding via Dual-Loop Autonomous Research

Wei Xiong; Jiangtong Li; Han Zhao; Linxiao Yang; Jie Li; Changjun Jiang

cs.LGOct 18, 2026

NeuroAct: Advancing EEG Analysis and Decoding via Dual-Loop Autonomous Research

Authors: Wei Xiong, Jiangtong Li, Han Zhao, Linxiao Yang, Jie Li, Changjun Jiang

Organizations: Tongji University · Ant Group · Shanghai Jiao Tong University · Translational Research Center, Shanghai Yangzhi Rehabilitation Hospital (Shanghai Sunshine Rehabilitation Center), Tongji University

Abstract

EEG analysis and decoding underpin brain-computer interfaces, cognitive assessment, and sleep monitoring. Fixed pipelines and reward-driven search do not explicitly determine how signal structure, representations, predictor behavior, and training dynamics contribute to performance or generalization gaps, leaving intervention choice underdetermined. We introduce NeuroAct, an autonomous research agent framework for EEG analysis and decoding. NeuroAct implements dual-loop autonomous research across EEG foundation models, task-specific predictors, and model combinations. In an inner loop, research, development, and review roles use evidence on signal structure, representations, predictor behavior, and training dynamics to test hypotheses, select analyses or interventions, and refine candidates using performance, stability, and cost feedback. An outer loop distills outcomes into conditional, testable recommendations with task and model conditions, then validates and revises them on subsequent tasks.We evaluate four settings: (A) EEG General Analysis, (B) Cross-task Model Optimization, (C) Single-task Open Research, and (D) Collaborative Prediction. In Setting A, NeuroAct's 80.88 macro-average exceeds BrainAgent's 63.13 by 17.75 points. In Setting B, NeuroAct raises EEGPT's mean balanced accuracy (BA) by 7.34 points and HBN Pearson's (r) from 0.6894 to 0.7362; in Setting C, it improves ADFTD BA by 16.33 points; and in Setting D, it raises Workload BA from 79.69% to 85.02%. Experience reuse raises BCIC-IV-2a BA from 34.66% to 50.81% while reducing LLM tokens by 64.02%, indicating recommendations improve cross-task performance. NeuroAct enables evidence-guided EEG research improved by cross-task experience.

Figures & tables

Scope / mechanism	CodeAct	BrainAgent	BrainPilot	NeuroWeaver	NeuroAct
Primary scope	General coding	Signal analysis	Brain research	Pipeline search	Analysis & decoding
Research actions
Code generation & execution	✓	—	✓	✓	✓
EEG analysis workflows	—	✓	✓	◐	✓
Model editing & training	✓	—	✓	✓	✓
EEG-FM adaptation & fine-tuning	—	—	—	—	✓

Table 1: EEG analysis, decoding, and research guidance across agent frameworks ( Wang et al., 2024c ; Zhou et al., 2026b ; Li et al., 2026a ; Wang et al., 2026 ) . ✓: supported; ◐: partial; —: unspecified.

Figure 1: NeuroAct’s dual-loop autonomous research framework. The Coordinator decomposes an EEG objective into parallel questions and directs separate cycles involving the Researcher, Developer, and Reviewer. The inner loop selects actions from reviewed evidence and experimental feedback within budget. The outer loop distills, matches, applies, and revises conditional experience across tasks. Capabilities are disclosed on demand; the lower panels show four evaluation settings.

Agent	Found.	Sleep	Neuro.	Phys.	Macro
CodeAct	68.16	61.01	60.38	54.52	61.02
BrainAgent	67.38	67.97	67.21	49.98	63.13
BrainPilot	64.82	58.43	50.83	61.52	58.90
NeuroAct	87.26	75.36	68.73	92.16	80.88

Table 3: Agent comparison with DeepSeek. Setting A includes source-reported CodeAct and BrainAgent scores ( Zhou et al., 2026a ) . Setting B uses LaBraM. Settings B, C, and D report BA (%) and HBN Pearson

r

. Metrics are detailed in App. D .

	Workload		BCIC-IV-2a		ADFTD
Ablation	BA $\uparrow$	AUC $\uparrow$	BA $\uparrow$	AUC $\uparrow$	BA $\uparrow$	AUC $\uparrow$
Full NeuroAct	77.73	0.886	56.34	0.805	62.94	0.831
w/o specialist personas	77.11	0.853	51.65	0.778	61.29	0.723
w/o EEG diagnostics	72.63	0.752	52.17	0.798	41.80	0.579
w/o code editing	66.22	0.722	52.60	0.789	48.70	0.576
w/o runtime skills	70.60	0.783	52.60	0.788	48.95	0.633

Table 4: Mechanism ablations with DeepSeek and a fixed CBraMod family. Full reports the best complete-framework BA within this model constraint. App. E.2 gives complete metrics.

Study	Experience	BA $\uparrow$	AUC $\uparrow$	Tokens (M) $\downarrow$	Min $\downarrow$
PhysioMI $\to$ BCIC-IV-2a	w/o	34.66	0.592	29.21	236.7
	with	50.81	0.775	10.51	165.3
CBraMod $\to$ BENDR	w/o	58.65	0.806	29.00	264.1
	with	58.28	0.811	21.96	219.2
CBraMod $\to$ BIOT	w/o	47.92	0.730	10.78	209.7
	with	50.26	0.749	9.45	163.4

Table 5: Experience transfer and revision with DeepSeek. All rows report full research costs.

Figure 2: Performance-cost tradeoffs. (a) Absolute outcomes by runtime-skill condition. (b) Changes from each study’s w/o or initial-advice reference at the origin. Token changes are percentages. Upward-leftward movement indicates better prediction at lower token cost.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Target	Channels	Window (s)	Training	Validation	Test
HBN	Age regression	129	2	121,740	28,620	19,740
Workload	Binary classification	19	4	1,537	300	297
PhysioMI	4-class classification	64	4	6,210	1,734	1,893
ADFTD	3-class classification	19	10	4,743	1,115	1,155
HMC	5-class classification	4	30	91,681	22,804	22,440
BCIC-IV-2a	4-class classification	22	4	2,784	1,152	1,152

Appendix

Table 6: Dataset characteristics and prediction-unit counts.

Capability	Principal input	Returned evidence or operation
Task and workspace context	Current task or selected candidate	Data/model specifications, research state, run summaries and selected detail
Literature and resources	Research query or named source	Retrieved papers or implementations, source content and compatibility information
EEG signal analysis	Recordings, event annotations and analysis configuration	Signal statistics and temporal, spectral or spatial structure, including event responses and connectivity
Model analysis	Candidate, run or prediction outputs	Input checks, optimization evidence, errors, perturbations and complementarity
Custom analysis	Developer-authored program and permitted inputs	Executed measurements with associated artifacts
Experiment execution	Implemented candidate and resource allowance	Training or prediction job, progress, outcomes and measured resource use

Appendix

Table 7: Research capabilities and their inputs and outputs. Access is filtered by persona.

Skill	Trigger
Source investigation	An unfamiliar paper, model implementation or data source is encountered.
Dataset onboarding	The study introduces new data or unclear recording and label semantics.
Model onboarding	A model is unregistered, or its implementation needs revision.
EEG signal characterization	The study examines EEG properties or a signal-related bottleneck.
Sleep recording analysis	The question involves sleep stages or events in a sleep recording.
Event and behavioral analysis	The task links EEG responses to events or behavioral measurements.

Appendix

Table 8: EEG runtime skills and the situations that trigger them.

	Workload			BCIC-IV-2a			ADFTD
Ablation	BA	F1	AUC	BA	F1	AUC	BA	F1	AUC
Full NeuroAct	77.73	0.674	0.886	56.34	0.564	0.805	62.94	0.636	0.831
w/o specialist personas	77.11	0.694	0.853	51.65	0.504	0.778	61.29	0.600	0.723
w/o EEG diagnostics	72.63	0.649	0.752	52.17	0.513	0.798	41.80	0.415	0.579
w/o code editing	66.22	0.620	0.722	52.60	0.523	0.789	48.70	0.485	0.576
w/o runtime skills	70.60	0.622	0.783	52.60	0.512	0.788	48.95	0.469	0.633

Appendix

Table 22: Complete predictive metrics of mechanism ablations for the CBraMod with DeepSeek.

Dataset	Skills	BA $\uparrow$	F1 $\uparrow$	AUC $\uparrow$	Tokens (M)	GPUh	Min
Workload	w/o	70.60	0.622	0.783	25.38	0.159	140.3
Workload	with	74.68	0.689	0.803	10.10	0.174	70.10
PhysioMI	w/o	51.13	0.515	0.759	14.88	0.965	173.3
PhysioMI	with	62.18	0.621	0.838	9.863	0.687	118.7
ADFTD	w/o	48.95	0.469	0.633	21.95	0.320	289.1
ADFTD	with	57.52	0.568	0.693	22.21	0.525	276.5

Appendix

Table 23: Predictive performance and research cost with and without runtime skills. ADFTD time sums active execution intervals.

Ablation	BA $\uparrow$	F1 $\uparrow$	AUC $\uparrow$	Tokens (M)	GPUh	Min
Full NeuroAct	56.34	0.564	0.805	35.11	0.671	299.1
w/o specialist personas	51.65	0.504	0.778	12.52	0.730	127.0
w/o EEG diagnostics	52.17	0.513	0.798	18.19	1.274	377.9
w/o code editing	52.60	0.523	0.789	19.85	1.413	386.3
w/o runtime skills	52.60	0.512	0.788	29.62	0.767	369.6
w/o auxiliary metrics	49.74	0.475	0.786	29.64	0.709	368.9

Appendix

Table 24: BCIC-IV-2a mechanism ablations and research cost.

Source	Target	Experience	BA	F1	AUROC	Tokens (M)	Min
PhysioMI	BCIC-IV-2a	w/o	34.66	0.278	0.592	29.21	236.7
PhysioMI	BCIC-IV-2a	with	50.81	0.495	0.775	10.51	165.3

Appendix

Table 25: PhysioMI-to-BCIC-IV-2a experience transfer with CBraMod and DeepSeek.

Experience	BA $\uparrow$	F1 $\uparrow$	AUC $\uparrow$	Tokens (M)	GPUh	Min
Original advice	62.36	0.625	0.840	13.90	1.096	149.8
Revised advice	63.13	0.632	0.847	15.43	0.850	139.1

Appendix

Table 26: Experience revision on PhysioMI with CBraMod and DeepSeek.

Experience	BA $\uparrow$	F1 $\uparrow$	AUC $\uparrow$	Tokens (M)	GPUh	Min
w/o	58.65	0.587	0.806	29.00	0.950	264.1
with	58.28	0.584	0.811	21.96	0.818	219.2

Appendix

Table 27: BENDR research on PhysioMI with DeepSeek. Scores describe the best evaluated predictors; costs cover each complete target study.

Experience	BA $\uparrow$	F1 $\uparrow$	AUC $\uparrow$	Tokens (M)	GPUh	Min
w/o	47.92	0.459	0.730	10.78	1.679	209.7
with	50.26	0.504	0.749	9.45	1.243	163.4

Appendix

Table 28: CBraMod experience transferred to BIOT on PhysioMI with DeepSeek.

Figure 3: ADFTD spectral development guided by montage and prediction-error analysis.

Figure 4: Amplitude-scaled training perturbations in the Workload study.

Figure 5: Workload research shifts from encoder learning-rate reduction to a class-weighted loss.

Figure 6: Transferred experience guides a patch-aggregation experiment on BCIC-IV-2a.

Figure 7: Workload collaborative prediction through a revised decision rule.

Figure 8: ADFTD router development from prediction analysis to a sample-dependent combination.

Dataset	BA (%) $\uparrow$	Macro-F1 $\uparrow$	AUROC $\uparrow$	Tokens (M) $\downarrow$	Min $\downarrow$
Workload	$77.73\pm 02.21$	$0.674\pm 0.056$	$0.886\pm 0.017$	$19.17\pm 08.09$	$106.4\pm 25.1$
BCIC-IV-2a	$56.34\pm 01.76$	$0.564\pm 0.033$	$0.805\pm 0.021$	$35.11\pm 13.90$	$209.2\pm 54.5$

Appendix

Table 29: Stability estimates (means

\pm

standard deviations).

EEG Decoding

Electroencephalography (EEG) decoding maps brain-signal patterns to mental states, stimuli, speech, or intended actions.

See all 208 papers

Sep 28, 2026

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

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.

Muyun Jiang, Yi Ding, Wei Zhang +9

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

Brain-Computer Interfaces Large Language Model-Guided NAS EEG Decoding Neural Architecture Search

Ask PDF

Aug 4, 2026

BrainBench: Benchmarking Large Language Models for Comprehensive EEG Understanding

Electroencephalography (EEG) analysis extends beyond assigning predefined labels to recordings; it requires workflows connecting natural-language instructions, signal processing, quantitative evidence, and scientific interpretation. We term this capability comprehensive EEG understanding. Existing evaluations, however, primarily target isolated decoding tasks or system-specific demonstrations, leaving the competence of large language models (LLMs) insufficiently quantified. We introduce \benchmarkname{}, a unified benchmark for comprehensive, instruction-conditioned EEG understanding. It comprises four subsets---Foundational Analysis, Sleep Assessment, Neurocognitive Assessment, and Physiological Integration---covering 17 datasets, \numcases{} tasks, and over \numinstances{} real-data instances. Given an instruction and EEG recordings with optional physiological signals, a system must perform the analysis and produce a scientifically grounded report and, when required, artifacts. Outputs are assessed through numerical, categorical, set, sequence, semantic, and artifact validation. We evaluate \nummodels{} representative LLMs across more than 100K executions under two paradigms: autonomous code execution with CodeAct and structured agentic analysis with BrainAgent. Results vary substantially across models, subsets, difficulty levels, and execution paradigms, showing that EEG competence depends on the model and its operationalization. \benchmarkname{} provides a reproducible testbed for advancing LLM-based EEG understanding. The code and benchmark will be released soon, with evaluation results continuously updated.

Yangxuan Zhou, Sha Zhao, Yuning Chen +4

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

Electroencephalography LLM Evaluation

Ask PDF

Feb 13, 2026

NeuroWeaver: An Autonomous Evolutionary Agent for Exploring the Programmatic Space of EEG Analysis Pipelines

Although foundation models have achieved remarkable success in general domains, applying them to electroencephalography (EEG) analysis is constrained by substantial data requirements and large parameter counts, which incur prohibitive computational costs and impede deployment in resource-constrained clinical environments. General-purpose automated machine learning frameworks are likewise ill-suited to this domain, since exploration within an unbounded programmatic space fails to incorporate essential neurophysiological priors and frequently yields neuroscientifically implausible solutions. We therefore propose NeuroWeaver, a unified autonomous evolutionary agent that generalizes across diverse EEG datasets and tasks by reformulating pipeline engineering as a discrete constrained optimization problem solved through large language model (LLM)-driven generation of executable code. A Domain-Informed Subspace Initialization confines the search to a neuroscientifically plausible manifold, while a Multi-Objective Evolutionary Optimization dynamically balances performance, novelty, and efficiency via self-reflective refinement. Across five heterogeneous benchmarks, NeuroWeaver synthesizes lightweight pipelines that outperform state-of-the-art task-specific methods on nearly all metrics and attain accuracy comparable to large-scale foundation models, even surpassing them on the HMC and Workload benchmarks with only

0.18

M and

0.011

M parameters, respectively.

Guoan Wang, Shihao Yang, Feng Liu

Department of Industrial and Systems Engineering, Rutgers, The State University of New Jersey, 96 Frelinghuysen Road, Piscataway, NJ 08854, USA

Evolutionary Optimization EEG Decoding Large Language Model-Guided Optimization LLM-Based Program Synthesis Automated Machine Learning

Ask PDF

Abstract

EEG Decoding

Explore similar work

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

BrainBench: Benchmarking Large Language Models for Comprehensive EEG Understanding

NeuroWeaver: An Autonomous Evolutionary Agent for Exploring the Programmatic Space of EEG Analysis Pipelines