Pith. sign in

REVIEW 4 major objections 4 minor 68 references

EasyBCI claims that a domain-specific LLM agent can replace manual neural-data preprocessing, preserving more task-relevant signal than general-purpose coding agents across six brain-signal modalities.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

An LLM agent with a text-only data fingerprint, quality-gated skill reuse, and expert checkpoints outperforms manual and general-purpose coding pipelines for EEG preprocessing and produces QC-passing pipelines for five other neural modalities.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Novel and well-engineered BCI preprocessing agent, but the key EEG comparison is statistically underpowered; the universal claim is premature. the 4 major comments →

arxiv 2607.29007 v2 pith:IQZ6WAO7 submitted 2026-07-31 q-bio.QM

EasyBCI Agent: Towards Universal Neural Data Preprocessing for Brain-Computer Interfaces

Yu Zhu , Runkai Zhao , Zhimin Zhou , Jinyu Cai , Zhouheng Yao , Chutian Zhang , Peiyuan Li , Xiaoxing Zhang
show 9 more authors
Qihao Zheng Jiamin Wu Mianxin Liu Chi Zhang Bin Min Lei Bai Chengyu Li Jingpeng Wu Chunfeng Song
This is my paper
classification q-bio.QM
keywords Brain-computer interfaceEEG preprocessingLLM agentneural datareproducibilityartifact removalmultimodalskill retrieval
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces EasyBCI, a two-phase LLM agent that plans and executes preprocessing pipelines for EEG, MEG, sEEG, ECoG, fNIRS, and unit-level spikes. Its central claim is that domain-specific orchestration—data fingerprinting, literature-grounded planning, quality-gated skill reuse, and expert gates—preserves more task-relevant linear separability than a manually designed pipeline or general-purpose coding agents using the same backbone model. On a 64-channel motor-imagery EEG dataset, all five EasyBCI backbones beat the manual pipeline on both classification tasks, and four of five beat general agents on binary classification while all five beat them on the four-class task. The same architecture produces complete, reproducible pipelines for all six modalities without per-run reconfiguration, suggesting that preprocessing can be standardized and audited without dedicated expert effort.

Core claim

Under a fixed-classifier protocol (CSP features plus LDA), EasyBCI preserves more task-relevant linear separability than both a manual pipeline and general-purpose coding agents. All five backbone variants outperform the manual pipeline on both action-type (binary) and object-type (4-class) labels. In same-backbone comparisons, EasyBCI with Opus 4.8 scores 66.85% versus 64.00% for Claude Code, and EasyBCI with GPT-5.5 scores 65.35% versus 63.29% for Codex; the gap widens on the 4-class task (44.70% vs. 32.50% and 40.65% vs. 35.54%). The paper attributes this advantage to domain-specific orchestration—modality-aware planning, quality-weighted skill retrieval with automatic deprecation, and ex

What carries the argument

The core mechanism is the data fingerprint: a text-only summary of the recording (format, channel count, sampling rate, spectral profile, artifact indicators, event structure) that is the only information about the raw signal ever passed to the LLM. This fingerprint feeds a plan agent that selects a literature-grounded operator sequence, which an execution agent carries out through a Code-Execute-Reflect loop with automatic QC checks and up to three repair attempts per step. Four supporting mechanisms do the work: modality-aware routing across six signal types, a three-layer Source Data Guard that keeps raw arrays off the model context and verifies file integrity, quality-weighted skill retr

Load-bearing premise

The evaluation treats LDA accuracy on CSP features from a single internal 64-channel motor-imagery dataset as the yardstick for 'task-relevant linear separability,' and the manual baseline was frozen without ICA or bad-channel removal; if that metric is not a faithful proxy, or the manual pipeline is unrepresentatively simple, the headline comparison is not established.

What would settle it

Run the same same-backbone comparison on two or more public EEG datasets (e.g., BCI Competition IV) with a manual baseline that includes ICA and bad-channel removal; if EasyBCI's accuracy advantage shrinks to within noise, the claim that domain-specific orchestration preserves more separability than a reasonable manual pipeline is not general.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Laboratories without dedicated preprocessing expertise can generate reproducible, auditable preprocessing pipelines with recorded decision provenance for any of six neural modalities.
  • Raw neural recordings can be processed without ever exposing the signal values to the model, reducing privacy review to inspection of a text fingerprint.
  • Quality-gated skill accumulation means the system improves with use while automatically discarding strategies whose performance degrades, countering the known failure mode of naive skill accumulation.
  • The fixed-classifier protocol isolates preprocessing as the sole experimental variable, offering a methodology that could be applied to other modalities and datasets to compare preprocessing choices.
  • The architecture suggests design principles for AI agents in other scientific domains where preprocessing decisions shape downstream conclusions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's claim is about preserving linear separability, not about end-to-end decoding performance; coupling preprocessing with downstream decoder accuracy as an acceptance criterion could reveal whether the preserved separability translates into better BCI performance in practice.
  • The rule-based QC metrics used for multi-modal validation certify signal-level properties but not task-level utility; a natural testable extension would be to run downstream decoders on the produced AI-ready arrays for each modality and check whether the pipelines improve classification.
  • Because the skill library was empty in the reported experiments, the architecture's benefit is shown in cold-start mode; the quality-weighted retrieval mechanism may yield further gains after repeated use on similar data, which the paper does not directly measure.
  • The privacy-by-architecture approach of exposing only fingerprints is a general template for other sensitive biomedical data types, where raw arrays could be replaced by summary statistics before any model sees them.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper introduces EasyBCI, a two-phase LLM agent that plans and executes preprocessing pipelines for six neural signal modalities (EEG, MEG, sEEG, ECoG, fNIRS, and unit-level spikes). A Plan Agent produces a text-only Data Fingerprint and a literature-grounded pipeline plan; an Execution Agent generates, runs, and repairs code under a Source Data Guard, quality-gated skill retrieval, and two expert-in-the-loop gates. The central quantitative claim is that, on a 64-channel motor-imagery EEG dataset, all five EasyBCI backbones preserve more task-relevant linear separability than a manually designed pipeline, and that EasyBCI outperforms general-purpose coding agents when using the same backbone LLM. Qualitative case studies on five additional modalities report QC-passing pipelines. The paper argues that domain-specific orchestration, rather than backbone capability, drives preprocessing quality.

Significance. The architecture is thoughtfully designed: raw-data isolation via a text-only fingerprint, AST-linted operator contracts, deterministic seeding, a quality-weighted skill library with deprecation, and explicit expert gates are concrete and transferable contributions. The output mini-repository with standalone code, provenance, and SHA-256 input hashes is a strong reproducibility feature. If the same-backbone advantage were statistically established, the paper would make a useful case for domain-specific agent orchestration in scientific preprocessing. However, the current evidence base is too thin to support the headline claims: the central comparison rests on six cross-validation-mean values from a single unreleased internal dataset, with no significance tests, confidence intervals, or effect sizes, and the Manual baseline omits standard operations such as ICA and bad-channel removal. The multi-modal demonstrations are valuable as engineering case studies, but their QC metrics are self-assessments and do not establish task-relevant validity. The manuscript is suitable for major revision rather than rejection because the architecture and claims are plausible and the statistical gaps

major comments (4)
  1. [§4.2.1, Table 4] The central claim that EasyBCI outperforms general-purpose coding agents on the same backbone is not statistically supported. The table reports means over 2 sessions × 3 seeds (n=6). The binary differences EasyBCI/Opus vs. Claude Code (66.85 vs. 64.00, 2.85 points) and EasyBCI/GPT-5.5 vs. Codex (65.35 vs. 63.29, 2.06 points) are well within noise under a two-sample t-test (p≈0.36 and p≈0.18, respectively). The 4-class Opus difference is larger (p≈0.04 unadjusted), but the GPT-5.5 4-class difference is not (p≈0.27), and no multiple-comparison correction is reported. Similarly, the claim that all five backbones outperform Manual fails for GPT-4o binary (61.75 vs. 55.25, p≈0.19). Please report per-session and per-seed values, paired tests across the common split seeds, confidence intervals, and effect sizes, or temper the claims to 'numerically higher in this exploratory dataset'.
  2. [§4.2.2, Table 5] The Manual baseline is not a representative literature-based motor-imagery pipeline: it omits ICA-based artifact removal, bad-channel rejection, and common average reference, all of which are standard in the cited motor-imagery literature. The 23-point gap between Manual and EasyBCI/DeepSeek-V4-Pro may therefore reflect pipeline completeness rather than the agent's planning or orchestration. A stronger baseline—for example, a fixed pipeline using PREP or Autoreject, or a manually designed pipeline that includes ICA with the same downstream classifier—would make the comparison meaningful. If the omission is intentional, the paper should state that the Manual pipeline is deliberately minimal and justify this choice.
  3. [§4.1.1 / §4.2.1] The quantitative evaluation uses a single internal 64-channel dataset with only 40 trials per session and two sessions. The reported mean and SD are over six CV-mean values, not over an independent sample of recordings. The claim that 'agent architecture accounts for more variance than the choice of language model' is extrapolated from this one dataset. Without an external dataset with more subjects/sessions, or at least a clear statement that the finding is dataset-specific, the generalization is unsupported. Please add replication on a public dataset (e.g., BCI Competition IV 2a) or explicitly limit the conclusion to this internal recording.
  4. [§4.3, Table 8] The multi-modal case studies are presented as demonstrations that the system 'produces complete reproducible pipelines,' which is supported by the described outputs. However, the claim of cross-modal generalization is not validated: all five non-EEG cases are evaluated only by rule-based auto-QC and PSD SNR. PSD SNR gains can be trivially inflated by bandpass filtering, as the Neuropixel case itself acknowledges (SNR decreases). No downstream metric, ground-truth comparison, or independent human review is reported. Please reframe these as engineering feasibility studies and, ideally, add at least one quantitative downstream validation per modality (e.g., preserved evoked-response latency for MEG, spike-sorting quality for Neuropixel).
minor comments (4)
  1. [§4.2.1] The sentence 'The top and bottom backbone positions are stable across both tasks (DeepSeek-V4-Pro highest, GPT-4o lowest), while the middle ranks of Opus 4.8 and GPT-5.5-Pro swap' overstates differences of 3–4 points on n=6; these rank changes are within noise and should be described as numerical, not stable.
  2. [§3.6, Eq. (4)] The quality weight λ uses ρ* ('reference threshold') but the value of ρ* is never defined. Please specify the reference threshold or state that it is a fixed hyperparameter.
  3. [Table 1 / §4.3] Table 1 lists fNIRS under the 'Skill' tier, but §4.3 and Appendix A.4 describe fNIRS as handled by the second tier; this is consistent but should be stated explicitly when the table is first invoked. Also, Table 8's 'Key validation' column includes statements like 'HFO band preserved' that are not quantitatively shown in the main text; please point to the appendix figures.
  4. [Appendix A.1] The Neuropixel QC report reports a decrease in PSD SNR (0.770 to 0.109) yet marks the case PASS. The reasoning is plausible, but the inconsistency between 'better SNR' as a QC criterion elsewhere and this case should be acknowledged in the main text to avoid confusion.

Circularity Check

0 steps flagged

No significant circularity: pipelines are generated before evaluation from fingerprints and literature, with empty skill library; the single self-citation is non-load-bearing.

full rationale

The central claims are not derived from fitted inputs or from the paper's own conclusions. The EEG comparison evaluates EasyBCI in cold-start mode with an empty skill library (§4.1.3: "The third category is EasyBCI itself, evaluated with five backbone LLMs ... in cold-start mode with an empty skill library"), so the reported predictions cannot be forced by accumulated self-derived skills. Pipeline plans are produced before any downstream accuracy is measured: the Plan Agent uses only a text-only Data Fingerprint and web-retrieved literature evidence (§3.4), and the fingerprint statistics are computed from the raw signal, not from test labels. The fixed-classifier probe (CSP + LDA) is applied identically across all conditions, so the comparison isolates preprocessing rather than encoding the result into the method. The quality-gating thresholds and skill-retrieval weights are hand-set design choices (Eq. 4, the 40% deprecation rule in §3.6), not parameters fitted to the reported accuracies. The multi-modal case studies are explicitly evaluated by signal-level QC and domain diagnostics rather than by circular reuse of the claimed outcome (§4.1.1: "Evaluation for these cases relies on signal-level QC metrics and domain-appropriate diagnostics rather than downstream classification accuracy"). The only self-citation, ref. [42], appears in a list of general agent surveys in §2.1 and is not load-bearing for any derivation. The statistical fragility of Table 4's same-backbone differences noted in the skeptic attack is a correctness/evidence concern, not a circularity, and does not raise the circularity score.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claims are empirical rather than derived, so no parameter is fitted to the reported accuracies. Several system and evaluation thresholds are hand-selected, and some are not reported with exact values; the analysis also depends on domain assumptions that LDA/CSP accuracy and rule-based QC capture preprocessing quality. No new physical or mathematical entities are introduced.

free parameters (6)
  • repair_attempt_cap = 3
    r_max=3 automatic repair attempts per step (§3.5). Hand-set, not fitted; controls how often the expert gate triggers.
  • skill_retrieval_top_k = 3
    Top k=3 skills returned to the code-generation prompt (§3.6). Hand-set design choice.
  • skill_deprecation_threshold = pass rate <40% after ≥5 uses
    Skills are deprecated below this threshold (§3.6, §3.9). Hand-set; not used in the reported cold-start EEG experiments.
  • retrieval_dimension_weights = modality 0.3, goal 0.1, paradigm 0.25, channels 0.15, sfreq 0.1, duration 0.1 (Fig. 6A)
    Fixed weights in Eq. (3), chosen by the authors, values partly shown in Fig. 6A. Not learned and not reported as tuned to the evaluation data.
  • bad_channel_variance_threshold = 3× montage median variance
    Used in EEG drop_bads and ECoG bad-channel detection (§4.2.2, Table 6; Appendix A.5). Domain heuristic, not fitted to downstream accuracy.
  • manual_baseline_parameters = notch 50 Hz, bandpass 0.4–40 Hz, resample 200 Hz, epoch 4 s
    Manual pipeline fixed before execution (§4.1.2, Table 5). It omits ICA and artifact rejection, which affects the fairness of the headline comparison.
axioms (5)
  • domain assumption CSP features + LDA accuracy is a sufficient proxy for task-relevant linear separability preserved by preprocessing.
    Stated in §4.1.2; the entire EEG evaluation rests on this measurement assumption.
  • domain assumption The text-only Data Fingerprint contains enough information to choose valid preprocessing parameters.
    The Plan Agent never sees raw samples (§3.3). If the fingerprint misses decisive signal properties, the generated plans can be wrong.
  • domain assumption Rule-based automatic QC metrics (PSD SNR, NaN fraction, channel counts) certify that a pipeline is correct.
    Multi-modal verdicts in §4.3 and Table 8 rely on these rules without external or downstream validation for five modalities.
  • standard math Standard library operators (MNE, ICA, CSP, Beer-Lambert) behave as documented.
    All generated pipelines call these libraries (§3.2, §4.1.4); the results inherit their correctness.
  • domain assumption LLMs can generate and repair contract-conforming operators from fingerprints within three attempts.
    Central to the Execution Agent (§3.5); supported only by the runs shown, not by a formal guarantee or broad benchmark.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of EasyBCI Agent: Towards Universal Neural Data Preprocessing for Brain-Computer Interfaces." pith.science (2026). https://pith.science/paper/IQZ6WAO7

@misc{pith2026260729007,
  author       = {Pith},
  title        = {Pith review of: EasyBCI Agent: Towards Universal Neural Data Preprocessing for Brain-Computer Interfaces},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IQZ6WAO7}},
  note         = {Machine review of arXiv:2607.29007}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Brain-computer interfaces translate neural activity into device commands, yet their performance hinges on preprocessing that remains manual, expert-dependent and poorly reproducible. Large language model agents can automate scientific coding, but existing systems lack the modality coverage, raw-data isolation, experience accumulation and domain oversight that neural preprocessing requires. We introduce EasyBCI, a two-phase LLM agent that plans and executes preprocessing pipelines for six signal types. A Plan Agent profiles each recording into a text-only Data Fingerprint that never exposes raw data to the model and selects a literature-grounded operator sequence. An Execution Agent generates, runs and self-corrects code until quality criteria are met, while a quality-gated experience system retains validated strategies as reusable skills. A domain expert intervenes at two decision gates, retaining human judgement where undetected error can invalidate downstream analyses. Evaluation on EEG with a fixed linear classifier shows that all five EasyBCI backbones preserve more task-relevant separability than the manual pipeline. Under same-backbone comparison, EasyBCI outperforms general-purpose coding agents on both label schemes for four of five configurations. EasyBCI extends to five additional modalities spanning nearly three orders of magnitude in sampling rate, producing complete reproducible pipelines with recorded decision provenance. These results indicate that domain-specific orchestration can bring auditable preprocessing within reach of laboratories lacking dedicated expertise, illustrating design principles applicable to AI agents in other scientific domains.

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

68 extracted references · 8 linked inside Pith

  1. [1]

    An accurate and rapidly calibrating speech neuroprosthesis

    Nicholas S. Card, Maitreyee Wairagkar, Carrina Iacobacci, Xianda Hou, Tyler Singer-Clark, et al. “An accurate and rapidly calibrating speech neuroprosthesis”. In:New England Journal of Medicine391.7 (2024), pp. 609–618

    2024

  2. [2]

    Long-term independent use of an intracortical brain–computer interface for speech and cursor control

    Nicholas S. Card, Tyler Singer-Clark, Hamza Peracha, Carrina Iacobacci, Xianda Hou, et al. “Long-term independent use of an intracortical brain–computer interface for speech and cursor control”. In:Nature Medicine32 (2026), pp. 2504–2510

    2026

  3. [3]

    A high-performance speech neuroprosthesis

    Francis R. Willett, Erin M. Kunz, Chaofei Fan, Donald T. Avansino, Guy H. Wilson, et al. “A high-performance speech neuroprosthesis”. In:Nature620.7976 (2023), pp. 1031–1036

    2023

  4. [4]

    A high-performance neuroprosthesis for speech decoding and avatar control

    Sean L. Metzger, Kaylo T. Littlejohn, Alexander B. Silva, David A. Moses, Margaret P. Seaton, et al. “A high-performance neuroprosthesis for speech decoding and avatar control”. In:Nature 620.7976 (2023), pp. 1037–1046

    2023

  5. [5]

    Walking naturally after spinal cord injury using a brain–spine interface

    Henri Lorach, Andrea Galvez, Valeria Spagnolo, Felix Martel, Serpil Karakas, et al. “Walking naturally after spinal cord injury using a brain–spine interface”. In:Nature618.7963 (2023), pp. 126–133

    2023

  6. [6]

    Multimodal functional neuroimaging: integrating functional MRI and EEG/MEG

    Bin He and Zhongming Liu. “Multimodal functional neuroimaging: integrating functional MRI and EEG/MEG”. In:IEEE Reviews in Biomedical Engineering1 (2008), pp. 23–40

    2008

  7. [7]

    Current status and issues regarding pre-processing of fNIRS neuroimaging data: an investigation of diverse signal filtering methods within a general linear model framework

    Paola Pinti, Felix Scholkmann, Antonia Hamilton, Paul Burgess, and Ilias Tachtsidis. “Current status and issues regarding pre-processing of fNIRS neuroimaging data: an investigation of diverse signal filtering methods within a general linear model framework”. In:Frontiers in Human Neuroscience12 (2019), p. 505

    2019

  8. [8]

    fNIRS reproducibility varies with data quality, analysis pipelines, and researcher experience

    Meryem A. Yücel, Robert Luke, Rickson C. Mesquita, Alexander von Lühmann, David M. A. Mehler, et al. “fNIRS reproducibility varies with data quality, analysis pipelines, and researcher experience”. In:Communications Biology8.1 (2025), p. 1149

    2025

  9. [9]

    Dengzhe Hou, Zihao Wu, Lingyu Jiang, Zirui Li, Fangzhou Lin, et al.Same Brain, Different Prediction: How Preprocessing Choices Undermine EEG Decoding Reliability. 2026. arXiv:2605. 07212[q-bio.NC]

    2026

  10. [10]

    How EEG preprocessing shapes decoding performance

    Roman Kessler, Alexander Enge, and Michael A. Skeide. “How EEG preprocessing shapes decoding performance”. In:Communications Biology8.1 (2025), p. 1039

    2025

  11. [11]

    An open multi- center MEG-EEG dataset for studying conscious visual perception

    Ling Liu, Oscar Ferrante, Tara Ghafari, Dorottya Hetenyi, Shujun Yang, et al. “An open multi- center MEG-EEG dataset for studying conscious visual perception”. In:Scientific Data13.1 (2026), p. 799

    2026

  12. [12]

    Open multi-center intracranial electroencephalography dataset with task probing conscious visual perception

    Alia Seedat, Alex Lepauvre, Jay Jeschke, Urszula Gorska-Klimowska, Marcelo Armendariz, et al. “Open multi-center intracranial electroencephalography dataset with task probing conscious visual perception”. In:Scientific Data12.1 (2025), p. 854

    2025

  13. [13]

    Large brain model for learning generic representations with tremendous EEG data in BCI

    Wei-Bang Jiang, Liming Zhao, and Bao-Liang Lu. “Large brain model for learning generic representations with tremendous EEG data in BCI”. In:International Conference on Learning Representations. Vol. 2024. 2024, pp. 16405–16426

    2024

  14. [14]

    Brant: Foundation model for intracranial neural signal

    Daoze Zhang, Zhizhang Yuan, Yang Yang, Junru Chen, Jingjing Wang, et al. “Brant: Foundation model for intracranial neural signal”. In:Advances in Neural Information Processing Systems36 (2023), pp. 26304–26321

    2023

  15. [15]

    Brain– computer interface devices for patients with paralysis and amputation: a meeting report

    Kristen Bowsher, EF Civillico, James Coburn, J Collinger, JL Contreras-Vidal, et al. “Brain– computer interface devices for patients with paralysis and amputation: a meeting report”. In: Journal of neural engineering13.2 (2016), p. 023001. 25

    2016

  16. [16]

    Peter Mitchell, Sarah CM Lee, Peter E Yoo, Andrew Morokoff, Rahul P Sharma, et al. “Assess- ment of safety of a fully implanted endovascular brain-computer interface for severe paralysis in 4 patients: the stentrode with thought-controlled digital switch (SWITCH) study”. In:JAMA neurology80.3 (2023), pp. 270–278

    2023

  17. [17]

    Long- term unsupervised recalibration of cursor-based intracortical brain–computer interfaces using a hidden Markov model

    Guy H Wilson, Elias A Stein, Foram Kamdar, Donald T Avansino, Tsam Kiu Pun, et al. “Long- term unsupervised recalibration of cursor-based intracortical brain–computer interfaces using a hidden Markov model”. In:Nature Biomedical Engineering(2025), pp. 1–19

    2025

  18. [18]

    EEG Agent:A UnifiedFramework for Automated EEG Analysis Using Large Language Models

    ShaZhao, MingyiPeng, HaitengJiang, TaoLi, andShijianLi. “EEG Agent:A UnifiedFramework for Automated EEG Analysis Using Large Language Models”. In:Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 40. 21. 2026, pp. 18063–18071

    2026

  19. [19]

    EEG-AI: An agentic system for AI-assisted semi-automated EEG preprocessing and artifact removal

    Abdelrahman Abdou, Martin Ivanov, Sarmed Shaya, Alice Rueda, Fatemeh Gholamali Nezhad, et al. “EEG-AI: An agentic system for AI-assisted semi-automated EEG preprocessing and artifact removal”. In:Journal of Neuroscience Methods432 (2026), p. 110759

    2026

  20. [20]

    Executable code actions elicit better LLM agents

    Xingyao Wang, Yangyi Chen, Lifan Yuan, Yizhe Zhang, Yunzhu Li, et al. “Executable code actions elicit better LLM agents”. In:Proceedings of the 41st International Conference on Machine Learning. 2024

    2024

  21. [21]

    Advocating for neurodata privacy and neurotechnology regulation

    Rafael Yuste. “Advocating for neurodata privacy and neurotechnology regulation”. In:Nature Protocols18.10 (2023), pp. 2869–2875

    2023

  22. [22]

    Considerations for patient privacy of Large Language Models in health care: scoping review

    Xiaoying Zhong, Siyi Li, Zhao Chen, Long Ge, Dongdong Yu, et al. “Considerations for patient privacy of Large Language Models in health care: scoping review”. In:Journal of Medical Internet Research27 (2025), e76571

    2025

  23. [23]

    Haochen Shi, Xingdi Yuan, and Bang Liu.Evolving Programmatic Skill Networks. 2026. arXiv: 2601.03509

    Pith/arXiv arXiv 2026

  24. [24]

    Human-in-the-loop software development agents

    Wannita Takerngsaksiri, Jirat Pasuksmit, Patanamon Thongtanunam, Chakkrit Tantithamtha- vorn, Ruixiong Zhang, et al. “Human-in-the-loop software development agents”. In:2025 IEEE/ACM 47th International Conference on Software Engineering: Software Engineering in Practice (ICSE-SEIP). IEEE, 2025, pp. 342–352

    2025

  25. [25]

    ReAct:SynergizingReasoning and Acting in Language Models

    ShunyuYao,JeffreyZhao,DianYu,NanDu,IzhakShafran,etal.“ReAct:SynergizingReasoning and Acting in Language Models”. In:International Conference on Learning Representations. 2023

    2023

  26. [26]

    Reflexion: Language Agents with Verbal Reinforcement Learning

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. “Reflexion: Language Agents with Verbal Reinforcement Learning”. In:Advances in Neural Information Processing Systems. Vol. 36. 2023, pp. 8634–8652

    2023

  27. [27]

    MetaGPT: Meta Programming for a Multi-Agent Collaborative Framework

    Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, et al. “MetaGPT: Meta Programming for a Multi-Agent Collaborative Framework”. In:International Conference on Learning Representations. 2024

    2024

  28. [28]

    AutoGen: Enabling Next- Gen LLM Applications via Multi-Agent Conversation

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, et al. “AutoGen: Enabling Next- Gen LLM Applications via Multi-Agent Conversation”. In:Conference on Language Modeling. 2024

    2024

  29. [29]

    SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering

    John Yang, Carlos Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, et al. “SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering”. In:Advances in Neural Information Processing Systems. Vol. 37. 2024, pp. 50528–50652

    2024

  30. [30]

    OpenHands: An Open Platform for AI Software Developers as Generalist Agents

    Xingyao Wang, Boxuan Li, Yufan Song, Frank F. Xu, Xiangru Tang, et al. “OpenHands: An Open Platform for AI Software Developers as Generalist Agents”. In:International Conference on Learning Representations. 2025. 26

    2025

  31. [31]

    Teaching Large Language Models to Self-Debug

    Xinyun Chen, Maxwell Lin, Nathanael Schärli, and Denny Zhou. “Teaching Large Language Models to Self-Debug”. In:International Conference on Learning Representations. 2024

    2024

  32. [32]

    Voyager: An Open-Ended Embodied Agent with Large Language Models

    Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, et al. “Voyager: An Open-Ended Embodied Agent with Large Language Models”. In:Transactions on Machine Learning Research(2024)

    2024

  33. [33]

    Bodhisattwa Prasad Majumder, Bhavana Dalvi Mishra, Peter Jansen, Oyvind Tafjord, Niket Tandon, et al.CLIN: A Continually Learning Language Agent for Rapid Task Adaptation and Generalization. 2023. arXiv:2310.10134

    Pith/arXiv arXiv 2023

  34. [34]

    TroVE: Inducing Verifiable and Efficient Toolboxes for Solving Programmatic Tasks

    Zhiruo Wang, Graham Neubig, and Daniel Fried. “TroVE: Inducing Verifiable and Efficient Toolboxes for Solving Programmatic Tasks”. In:Proceedings of the 41st International Conference on Machine Learning. PMLR, 2024, pp. 51177–51191

    2024

  35. [35]

    WentaoHu,ZhendongChu,YimingZhang,JundaWu,MingJin,etal.SkillBrew:Multi-Objective Curation of Skill Banks for LLM Agents. 2026. arXiv:2605.29440

    Pith/arXiv arXiv 2026

  36. [36]

    Hongji Pu, Xinyuan Song, and Liang Zhao.SkillOps: Managing LLM Agent Skill Libraries as Self-Maintaining Software Ecosystems. 2026. arXiv:2605.13716

    Pith/arXiv arXiv 2026

  37. [37]

    Xunjian Yin, Xinyi Wang, Liangming Pan, Li Lin, Xiaojun Wan, et al.Gödel Agent: A Self- Referential Agent Framework for Recursive Self-Improvement. 2024. arXiv:2410.04444

    Pith/arXiv arXiv 2024

  38. [38]

    Hangfan Zhang, Shao Zhang, Kangcong Li, Chen Zhang, Yang Chen, et al.Self-Harness: Harnesses That Improve Themselves. 2026. arXiv:2606.09498

    Pith/arXiv arXiv 2026

  39. [39]

    The Rise and Potential of Large Language Model Based Agents: A Survey

    Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, et al. “The Rise and Potential of Large Language Model Based Agents: A Survey”. In:Science China Information Sciences68.2 (2025), p. 121101

    2025

  40. [40]

    A Survey on Large Language Model Based Autonomous Agents

    Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, et al. “A Survey on Large Language Model Based Autonomous Agents”. In:Frontiers of Computer Science18.6 (2024), p. 186345

    2024

  41. [41]

    Shuo Ren, Can Xie, Pu Jian, Zhenjiang Ren, Chunlin Leng, et al.Towards Scientific Intelligence: A Survey of LLM-Based Scientific Agents. 2025. arXiv:2503.24047

    arXiv 2025

  42. [42]

    SciDataSailor: Deep Scientific Data Exploring

    Jiyong Rao, Yicheng Qiu, Chi Zhang, Chunfeng Song, and Runkai Zhao. “SciDataSailor: Deep Scientific Data Exploring”. In:arXiv preprint arXiv:2607.28098(2026)

    Pith/arXiv arXiv 2026

  43. [43]

    Kim, Ahmed Alaa, and Danilo Bernardo.EEG-GPT: Exploring Capabilities of Large Language Models for EEG Classification and Interpretation

    Jonathan W. Kim, Ahmed Alaa, and Danilo Bernardo.EEG-GPT: Exploring Capabilities of Large Language Models for EEG Classification and Interpretation. 2024. arXiv:2401.18006

    Pith/arXiv arXiv 2024

  44. [44]

    Jiazhen Hong, Weinan Wang, and Laleh Najafizadeh.ChatBCI: A P300 Speller BCI Leveraging Large Language Models for Improved Sentence Composition in Realistic Scenarios. 2024. arXiv: 2411.15395

    arXiv 2024

  45. [45]

    EEGUnity: Open- Source Tool in Facilitating Unified EEG Datasets Towards Large-Scale EEG Model

    Chengxuan Qin, Rui Yang, Wenlong You, Zhige Chen, Longsheng Zhu, et al. “EEGUnity: Open- Source Tool in Facilitating Unified EEG Datasets Towards Large-Scale EEG Model”. In:IEEE Transactions on Neural Systems and Rehabilitation Engineering33 (2025), pp. 1653–1663

    2025

  46. [46]

    EEGLAB: An Open Source Toolbox for Analysis of Single- Trial EEG Dynamics Including Independent Component Analysis

    Arnaud Delorme and Scott Makeig. “EEGLAB: An Open Source Toolbox for Analysis of Single- Trial EEG Dynamics Including Independent Component Analysis”. In:Journal of Neuroscience Methods134.1 (2004), pp. 9–21

    2004

  47. [47]

    FieldTrip: Open Source Software for Advanced Analysis of MEG, EEG, and Invasive Electrophysiological Data

    Robert Oostenveld, Pascal Fries, Eric Maris, and Jan-Mathijs Schoffelen. “FieldTrip: Open Source Software for Advanced Analysis of MEG, EEG, and Invasive Electrophysiological Data”. In:Computational Intelligence and Neuroscience2011.1 (2011), p. 156869

    2011

  48. [48]

    Brainstorm: A User-Friendly Application for MEG/EEG Analysis

    François Tadel, Sylvain Baillet, John C. Mosher, Dimitrios Pantazis, and Richard M. Leahy. “Brainstorm: A User-Friendly Application for MEG/EEG Analysis”. In:Computational Intelligence and Neuroscience2011.1 (2011), p. 879716. 27

    2011

  49. [49]

    MEG/EEG Group Analysis with Brainstorm

    François Tadel, Elizabeth Bock, Guiomar Niso, John C. Mosher, Martin Cousineau, et al. “MEG/EEG Group Analysis with Brainstorm”. In:Frontiers in Neuroscience13 (2019), p. 76

    2019

  50. [50]

    MEG and EEG Data Analysis with MNE-Python

    Alexandre Gramfort, Martin Luessi, Eric Larson, Denis A. Engemann, Daniel Strohmeier, et al. “MEG and EEG Data Analysis with MNE-Python”. In:Frontiers in Neuroscience7 (2013), p. 267

    2013

  51. [51]

    Neo: An Object Model for Handling Electrophysiology Data in Multiple Formats

    Samuel Garcia, Domenico Guarino, Florent Jaillet, Todd Jennings, Robert Pröpper, et al. “Neo: An Object Model for Handling Electrophysiology Data in Multiple Formats”. In:Frontiers in Neuroinformatics8 (2014), p. 10

    2014

  52. [52]

    SpikeInterface, a Unified Framework for Spike Sorting

    Alessio P. Buccino, Cole L. Hurwitz, Samuel Garcia, Jeremy Magland, Joshua H. Siegle, et al. “SpikeInterface, a Unified Framework for Spike Sorting”. In:eLife9 (2020), e61834

    2020

  53. [53]

    NeuroKit2: A Python Toolbox for Neurophysiological Signal Processing

    Dominique Makowski, Tam Pham, Zen J. Lau, Jan C. Brammer, François Lespinasse, et al. “NeuroKit2: A Python Toolbox for Neurophysiological Signal Processing”. In:Behavior Research Methods53.4 (2021), pp. 1689–1696

    2021

  54. [54]

    Software; source code available at https://github.com/brainflow- dev/brainflow

    BrainFlow contributors.BrainFlow, a Library to Obtain, Parse and Analyze EEG, EMG, ECG and Other Biosensor Data. Software; source code available at https://github.com/brainflow- dev/brainflow. 2024.url:https://brainflow.org(visited on 07/28/2026)

    2024

  55. [55]

    The PREP Pipeline: Standardized Preprocessing for Large-Scale EEG Analysis

    Nima Bigdely-Shamlo, Tim Mullen, Christian Kothe, Kyung-Min Su, and Kay A. Robbins. “The PREP Pipeline: Standardized Preprocessing for Large-Scale EEG Analysis”. In:Frontiers in Neuroinformatics9 (2015), p. 16

    2015

  56. [56]

    Autoreject: Automated Artifact Rejection for MEG and EEG Data

    Mainak Jas, Denis A. Engemann, Yousra Bekhti, Federico Raimondo, and Alexandre Gramfort. “Autoreject: Automated Artifact Rejection for MEG and EEG Data”. In:NeuroImage159 (2017), pp. 417–429

    2017

  57. [57]

    The Harvard Automated Processing Pipeline for Electroencephalography (HAPPE): Standardized Processing Software for Developmental and High-Artifact Data

    Laurel J. Gabard-Durnam, Adriana S. Mendez Leal, Carol L. Wilkinson, and April R. Levin. “The Harvard Automated Processing Pipeline for Electroencephalography (HAPPE): Standardized Processing Software for Developmental and High-Artifact Data”. In:Frontiers in Neuroscience 12 (2018), p. 97

    2018

  58. [58]

    HAPPILEE: HAPPE in Low Electrode Electroencephalography, a Standardized Pre-Processing Software for Lower Density Recordings

    Kelsie L. Lopez, Alexa D. Monachino, Santiago Morales, Stephanie C. Leach, Maureen E. Bowers, et al. “HAPPILEE: HAPPE in Low Electrode Electroencephalography, a Standardized Pre-Processing Software for Lower Density Recordings”. In:NeuroImage260 (2022), p. 119390

    2022

  59. [59]

    Introducing RELAX: An Automated Pre-Processing Pipeline for Cleaning EEG Data—Part 1: Algorithm and Application to Oscillations

    N. W. Bailey, Mana Biabani, A. T. Hill, Aleksandra Miljevic, Nigel C. Rogasch, et al. “Introducing RELAX: An Automated Pre-Processing Pipeline for Cleaning EEG Data—Part 1: Algorithm and Application to Oscillations”. In:Clinical Neurophysiology149 (2023), pp. 178–201

    2023

  60. [60]

    osl-ephys: A Python Toolbox for the Analysis of Electrophysiology Data

    Mats W. J. van Es, Chetan Gohil, Andrew J. Quinn, and Mark W. Woolrich. “osl-ephys: A Python Toolbox for the Analysis of Electrophysiology Data”. In:Frontiers in Neuroscience19 (2025), p. 1522675

    2025

  61. [61]

    Analysis Methods for Measuring Passive Auditory fNIRS Responses Generated by a Block- Design Paradigm

    Robert Luke, Eric Larson, Maureen J. Shader, Hamish Innes-Brown, Lindsey Van Yper, et al. “Analysis Methods for Measuring Passive Auditory fNIRS Responses Generated by a Block- Design Paradigm”. In:Neurophotonics8.2 (2021), p. 025008

    2021

  62. [62]

    iEEG-recon: A Fast and Scalable Pipeline for Accurate Reconstruction of Intracranial Electrodes and Implantable Devices

    Alfredo Lucas, Brittany H. Scheid, Akash R. Pattnaik, Ryan Gallagher, Marissa Mojena, et al. “iEEG-recon: A Fast and Scalable Pipeline for Accurate Reconstruction of Intracranial Electrodes and Implantable Devices”. In:Epilepsia65.3 (2024), pp. 817–829

    2024

  63. [63]

    iEEG-BIDS, Extending the Brain Imaging Data Structure Specification to Human In- tracranial Electrophysiology

    Christopher Holdgraf, Stefan Appelhoff, Stephan Bickel, Kristofer Bouchard, Sasha D’Ambrosio, et al. “iEEG-BIDS, Extending the Brain Imaging Data Structure Specification to Human In- tracranial Electrophysiology”. In:Scientific Data6.1 (2019), p. 102. 28

    2019

  64. [64]

    MEG-BIDS, the Brain Imaging Data Structure Extended to Magnetoencephalography

    Guiomar Niso, Krzysztof J. Gorgolewski, Elizabeth Bock, Teon L. Brooks, Guillaume Flandin, et al. “MEG-BIDS, the Brain Imaging Data Structure Extended to Magnetoencephalography”. In:Scientific Data5.1 (2018), p. 180110

    2018

  65. [65]

    NIRS- BIDS: Brain Imaging Data Structure Extended to Near-Infrared Spectroscopy

    Robert Luke, Robert Oostenveld, Helena Cockx, Guiomar Niso, Maureen J. Shader, et al. “NIRS- BIDS: Brain Imaging Data Structure Extended to Near-Infrared Spectroscopy”. In:Scientific Data12.1 (2025), p. 159

    2025

  66. [66]

    MNE-BIDS: Organizing Electrophysiological Data into the BIDS Format and Facilitating Their Analysis

    Stefan Appelhoff, Matthew Sanderson, Teon L. Brooks, Marijn van Vliet, Romain Quentin, et al. “MNE-BIDS: Organizing Electrophysiological Data into the BIDS Format and Facilitating Their Analysis”. In:Journal of Open Source Software4.44 (2019), p. 1896

    2019

  67. [67]

    The Neurodata Without Borders Ecosystem for Neurophysiological Data Science

    Oliver Rübel, Andrew Tritt, Ryan Ly, Benjamin K. Dichter, Satrajit Ghosh, et al. “The Neurodata Without Borders Ecosystem for Neurophysiological Data Science”. In:eLife11 (2022), e78362

    2022

  68. [68]

    Facilitating Analysis of Open Neurophysiology Data on the DANDI Archive Using Large Language Model Tools

    Jeremy F. Magland, Ryan Ly, Oliver Rübel, and Benjamin Dichter. “Facilitating Analysis of Open Neurophysiology Data on the DANDI Archive Using Large Language Model Tools”. In: Scientific Data12.1 (2025), pp. 1988–1988. 29 A Multi-modal Case Study Details This appendix presents the full pipeline specifications, QC metrics and selected reasoning excerpts fo...

    2025

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.