acceptodds
Under review as a conference paper at ICLR 2027

Ewen: A Covariate-Aware Foundation Model for EEG Understanding

14 views

Abstract

EEG foundation models have improved transferable neural representation learning, but existing EFM–LLM frameworks still largely use language models as closed-set text classifiers. We propose Ewen, a covariate-aware EEG foundation model that incorporates physiological context into EEG representations for decoding and natural-language description. Ewen converts EEG and available auxiliary physiological signals into descriptorized covariates, providing language-aligned context for interpreting neural signals. These covariates are integrated with EEG tokens through covariate-aware token geometry, connecting physiological context with the temporal and channel structure of EEG representations. The language backbone is then adapted through language-reference-guided gradient projection to control language drift during task learning. Experiments across six EEG benchmarks demonstrate improved closed-vocabulary decoding, while generation evaluations show that Ewen can describe spectral characteristics, physiological relationships, and temporal and spatial variations. Together, these results support the value of physiological context for EEG modeling and represent a step toward physiologically grounded EEG understanding.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG Foundation Model, EEG Understanding, Physiological Context

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_bUqaOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposed Ewen, a covariate-aware EEG Foundation model that integrate discrete EEG waveform tokens with handcrafted physiological descriptors through feature-wise modulation and hierarchical token geometry. It further introduces language-reference-guided gradient projection to mitigate language drift during adaptation. The authors report competitive classification results across multiple EEG benchmarks and evaluate physiological understanding through structured generation and waveform-dependent comparisons.

Critical strengths
  1. The paper provides reasonably comprehensive experiments across multiple EEG benchmarks, with competitive performance against both conventional EEG foundation models and EEG-LLM approaches.
  2. The ablation studies are informative and generally support the effectiveness of the proposed components. In particular, the gradient projection analysis provides empirical evidence that the language-reference-guided adaptation behaves as intended, while the language-retention experiments demonstrate a favorable trade-off compared with full fine-tuning.
Critical weaknesses
  1. Some benchmarks, particularly TUEV and TUAB, involve waveform morphology that may be interpretable through visual representations. Given that Ewen already employs a 4B-parameter LLM, it would be interesting to include a VLM-based baseline using rendered multichannel EEG waveforms [1].
  2. A critical baseline is missing to justify the proposed covariate-aware architecture. Since the physiological descriptors are essentially handcrafted numerical features, why not simply provide them to a pretrained LLM as structured JSON or key-value text? A comparison against such a baseline, ideally with the same EEG waveform tokens and training budget, would help establish whether the proposed covariate modulation and positional encoding offer benefits beyond explicitly providing physiological context.

[1] Pan, J., Wang, Y., Zhang, E., Lu, B., Zheng, W., & Li, D. (2026). Visualizing the Invisible: Generative Visual Grounding Empowers Universal EEG Understanding in MLLMs. ArXiv, abs/2605.18172.

Questions for authors

See weaknesses. I would like raise my score if Problem 2 is addressed. Besides, the necessity of the proposed language-reference-guided gradient projection remains somewhat unclear. Although the authors demonstrate better language retention than full fine-tuning, fine-tuning itself can substantially affect the general capabilities of relatively small LLMs [2].

[2] Kotha, S., Springer, J.M., & Raghunathan, A. (2023). Understanding Catastrophic Forgetting in Language Models via Implicit Inference. ArXiv, abs/2309.10105.

4 comments
AuthorsOpenReview ↗

Response to Reviewer bUqa [1/2]

Overview

Thank you for your comments and suggestions. Post 1: Weakness 2/Question (W2/Q, JSON input) and Weakness 1/Question (W1/Q, VLM comparisons); Post 2: the additional Question (Q, projection) and the general-language evaluation in Appendix R. Please refer to the updated PDF for detailed revisions, experiments, and analysis. Thank you also for recognizing the benchmark coverage and ablation analyses.

Response to Weakness 2/Question (W2/Q) (JSON Baseline)

Summary

Physiological descriptors supplied as structured text provide an important baseline for assessing the added value of covariate modulation and positional encoding.

Response

Thank you for suggesting this baseline. Following your suggestion, we added a Waveform + JSON experiment that provides the available physiological descriptors as JSON and compares it with Waveform-only and Full Ewen, examining the roles of physiological information and its integration.

MethodDatasetDirect F1 (higher)Withheld F1 (higher)Summary R-L (higher)
Waveform-onlyMumtaz 20160.5621 ± 0.02100.5009 ± 0.01050.8288 ± 0.0158
Mental Arithmetic0.4388 ± 0.01310.3969 ± 0.00790.7827 ± 0.0105
SHU-MI0.5079 ± 0.02270.4117 ± 0.00520.7826 ± 0.0140
Dataset-macro0.5026 ± 0.01400.4371 ± 0.00440.7981 ± 0.0092
Waveform + JSONMumtaz 20160.8774 ± 0.01580.5231 ± 0.01300.8238 ± 0.0125
Mental Arithmetic0.7177 ± 0.01750.3977 ± 0.01260.8280 ± 0.0085
SHU-MI0.8504 ± 0.02100.3989 ± 0.00940.8233 ± 0.0189
Dataset-macro0.8152 ± 0.01050.4399 ± 0.00680.8250 ± 0.0081
Full EwenMumtaz 20160.9414 ± 0.01050.6137 ± 0.01550.8857 ± 0.0091
Mental Arithmetic0.7780 ± 0.02190.4861 ± 0.01720.8998 ± 0.0064
SHU-MI0.8663 ± 0.01920.5163 ± 0.01350.8931 ± 0.0237
Dataset-macro0.8619 ± 0.01660.5387 ± 0.01200.8929 ± 0.0113

↑ higher is better, ↓ lower is better.

Adding JSON increases dataset-macro Direct F1 from 0.5026 to 0.8152, showing that the model can use the supplied values to generate corresponding descriptions. Explicit physiological values do supply part of the information needed for attribute description.

Full Ewen further increases Direct F1 to 0.8619, exceeding JSON by 0.0467. This improvement is consistent with jointly learning connections among physiological conditions, the current waveform, and description targets, supporting more accurate correspondence between physiological information and its verbal description.

For spatiotemporal relations that are not directly provided, JSON achieves a Withheld F1 of 0.4399, close to Waveform-only at 0.4371; Full Ewen reaches 0.5387, a gain of 0.0988 over JSON, with numerical improvements on all three datasets. These results support Ewen’s ability to recover unprovided relations by combining physiological context with waveform structure. For example, whole-window alpha power does not directly specify temporal changes or differences between local brain regions; these relations still require the temporal and channel structure of the waveform.

Modification

These results are now included in Table 4 (“Input ablations for physiological interpretation,” p. 8) in Section 4.4. The table reports input ablations and covariate permutation by dataset, with clean Ewen as the shared reference. The accompanying analysis (p. 8, starting at lines 406 and 425) discusses the information supplied by covariates, learned correspondence, and the complementary roles of waveforms and physiological context.

AuthorsOpenReview ↗

Response to Reviewer bUqa [2/2]

Response to the Additional Question (Q): Language-Reference Gradient Projection

Summary

Full fine-tuning can itself affect smaller LLMs, so the existing comparison does not sufficiently establish the necessity of language-reference gradient projection.

Response

Thank you for asking about the role of projection. To examine its contribution more directly, we added a same-architecture projection on/off control and comparisons with other adaptation strategies.

Table 2 compares Core+residual (no proj.), Mixed EEG–text, Replay FT, KL-regularized FT, and Ewen. Core+residual (no proj.) is Ewen with only gradient projection disabled; its core+residual adapter architecture and training settings are unchanged.

Table 2: Additional Adaptation Methods and Projection Controls

MethodDatasetWithheld F1 ↑Summary R-L ↑OWT ΔPPL ↓
Core+residual (no proj.)Mumtaz 20160.5980 ± 0.01680.8856 ± 0.0101—
Mental Arithmetic0.4745 ± 0.01860.8995 ± 0.0073—
SHU-MI0.5027 ± 0.01470.8920 ± 0.0248—
Dataset-macro0.5251 ± 0.01410.8924 ± 0.01310.1278 ± 0.0017
Mixed EEG–textMumtaz 20160.5925 ± 0.01740.8856 ± 0.0108—
Mental Arithmetic0.4685 ± 0.01910.8993 ± 0.0070—
SHU-MI0.4991 ± 0.01590.8915 ± 0.0251—
Dataset-macro0.5200 ± 0.01630.8921 ± 0.01350.1046 ± 0.0013
Replay FTMumtaz 20160.5809 ± 0.01900.8855 ± 0.0113—
Mental Arithmetic0.4618 ± 0.02020.8992 ± 0.0079—
SHU-MI0.4883 ± 0.01660.8906 ± 0.0257—
Dataset-macro0.5103 ± 0.01800.8918 ± 0.01300.0538 ± 0.0012
KL-regularized FTMumtaz 20160.5909 ± 0.01690.8856 ± 0.0103—
Mental Arithmetic0.4665 ± 0.01830.8993 ± 0.0072—
SHU-MI0.4961 ± 0.01480.8911 ± 0.0245—
Dataset-macro0.5178 ± 0.01390.8920 ± 0.00950.0494 ± 0.0011
EwenMumtaz 20160.6137 ± 0.01550.8857 ± 0.0091—
Mental Arithmetic0.4861 ± 0.01720.8998 ± 0.0064—
SHU-MI0.5163 ± 0.01350.8931 ± 0.0237—
Dataset-macro0.5387 ± 0.01200.8929 ± 0.01130.0402 ± 0.0011

↑ higher is better, ↓ lower is better.

In the same-architecture comparison, Ewen increases dataset-macro Withheld F1 from 0.5251 for Core+residual (no proj.) to 0.5387, with numerical gains on all three datasets; Summary R-L remains similar (0.8924 versus 0.8929). At the same time, OWT ΔPPL decreases from 0.1278 to 0.0402, a relative reduction of approximately 68.5%. This suggests that projection helps reduce language drift while improving EEG-relation recovery.

The other adaptation strategies provide additional context. Mixed EEG–text, Replay FT, and KL-regularized FT achieve OWT ΔPPL values of 0.1046, 0.0538, and 0.0494, with Withheld F1 scores of 0.5200, 0.5103, and 0.5178, respectively; Ewen achieves 0.0402 and 0.5387. In this comparison, Ewen achieves both lower language drift on the general reference distribution and higher relation-recovery performance.

These results are consistent with the design: the language-reference gradient constrains reference-sensitive core updates, while the residual retains capacity for task adaptation. Here, 68.5% denotes the relative reduction in ΔPPL, and the language-drift conclusion concerns the general-text reference distribution.

Appendix R further supports your concern about full fine-tuning: Full FT substantially affects general-language capabilities after EEG adaptation, including text prediction, reasoning, and instruction-following. Ewen performs better than Full FT throughout these evaluations. Together with its EEG classification and physiological-relation recovery, this demonstrates a trade-off between task adaptation and general-language retention. These behavioral evaluations complement the same-architecture projection control above. Full results are in Appendix R.

Modification

Section 4.5, Table 5 (p. 9; analysis from line 459) reports the adaptation strategies and same-architecture projection control; Appendix R (p. 48, Table 26; analysis from line 2556) adds general-language evaluations and their trade-off with EEG task performance.

I am glad to see the supplement VLM experiment on waveform-json baseline. While I still believe that a visual VLM baseline on TUEV/TUAB would be an interesting direction for further evaluation, I consider my primary concerns sufficiently addressed. Given the new evidence and the authors’ constructive responses, I have decided to raise my score from 2 to 3.

AuthorsOpenReview ↗

Appreciation for Reviewer bUqa‘s Follow-up

Dear Reviewer bUqa,

Thank you very much for your follow-up and positive reassessment. We are grateful for your constructive feedback and suggestions.

Following your suggestion, we will further explore VLM-based methods for EEG analysis in future work, including visual baselines using rendered multichannel EEG waveforms on TUEV and TUAB.

We are glad that our additional experiments and responses have addressed your main concerns. Thank you again for your time and for helping us improve the paper.

Best regards,
The Authors

Reviewer_Fb95OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Ewen, a covariate-aware EEG foundation model intended to move EEG–LLM systems beyond closed-vocabulary classification toward physiologically grounded natural-language interpretation. The framework first converts EEG-intrinsic statistics and available auxiliary biosignals (EOG, ECG, and EMG) into descriptorized covariates, then injects these covariates into EEG token representations through feature-wise modulation and hierarchical temporal/channel geometry. The language backbone is adapted with an online language-reference-guided gradient projection mechanism designed to reduce interference with pretrained language behavior while retaining task adaptation capacity. The method is evaluated on eight public EEG benchmarks, with six datasets used in the main closed-vocabulary comparison and additional analyses on generation. The paper reports strong classification results, ablations of hierarchical position encoding and covariate fusion, comparisons among LoRA, full fine-tuning, and the proposed projected adaptation, and expert evaluation of generated physiological summaries. The generation analysis distinguishes explicit descriptor facts from waveform-derived withheld relations and includes OpenWebText-based measurements of language drift.

Critical strengths
  1. A well-motivated attempt to connect EEG representation learning with physiologically grounded language generation. The paper identifies a meaningful limitation of many existing EEG–LLM systems: the language model is often used mainly as a text-formatted classifier rather than as an interface for describing signal structure and physiological context. Ewen addresses this with a coherent pipeline that combines handcrafted physiological descriptors, EEG token structure, and conditional generation. The distinction between direct descriptor facts and withheld waveform-derived relations is particularly useful because it attempts to test whether the model can recover information that is not explicitly supplied in the covariate vector.
  2. Broad evaluation and unusually careful analysis of generated physiological descriptions. The experimental section is extensive. The main benchmark covers multiple clinical, cognitive, motor-imagery, and affective EEG settings, while the appendices provide additional datasets, implementation details, and covariate-free comparisons. The expert audit is also a notable strength: outputs are evaluated on fixed test windows, claims are independently labeled by two EEG-trained experts, disagreements are reconciled, and subject-clustered uncertainty is reported. This goes beyond relying only on surface-form metrics such as ROUGE or BERTScore and materially strengthens the evidence for factual generation quality.
Critical weaknesses
  1. The central claim of “EEG understanding” remains difficult to disentangle from engineered feature extraction and structured target construction. A substantial part of Ewen's semantic grounding is built from deterministic physiological descriptors computed from the same input window, including band powers, instantaneous alpha frequency, aperiodic exponent, and cross-signal statistics. Descriptor text for tokenizer alignment is also generated from these structured metrics using deterministic templates and an off-the-shelf LLM. Consequently, part of the reported natural-language capability can be interpreted as learning to verbalize engineered signal features rather than acquiring a more general form of EEG understanding. The withheld targets are a useful countermeasure, but they are still predefined low-level spectral comparisons derived algorithmically from the waveform. I therefore find the paper's strongest conceptual claim somewhat broader than what the evaluation currently establishes. A stronger isolation study would help. In particular, the paper would benefit from a non-neural or lightweight feature-to-text baseline that receives exactly the same descriptor bank, a waveform-only generation model trained with the same targets, and shuffled/corrupted-covariate controls. These comparisons would clarify how much of the gain comes from the proposed EEG–language representation learning versus the availability of carefully designed physiological statistics and structured supervision.
  2. Evidence for “controlled language drift” is suggestive but does not yet demonstrate preservation of general language capability. The online gradient projection is technically interesting, but the evidence for language preservation is mainly based on OpenWebText perplexity/BERTScore and geometric diagnostics of gradient alignment. These measurements show reduced deviation on the chosen reference distribution, yet they do not establish that broader instruction-following, reasoning, or general language competence is preserved after EEG adaptation. This is especially relevant because only the core branch is projected, while the residual branch and EEG-side parameters still receive unprojected task gradients; the paper itself notes that optimizer dynamics and residual updates can still change language behavior. Standard post-adaptation evaluations on general language or reasoning benchmarks, or at least a diverse held-out instruction set, would make the “controlled language drift” claim substantially more convincing. A related fairness issue is that the main closed-vocabulary table compares Ewen against several EEG-only baselines even when Ewen can access auxiliary EOG/ECG/EMG channels. The appendix partially addresses this through a covariate-free Ewen comparison on three datasets, but an input-matched comparison across all datasets with auxiliary modalities would be preferable. This is important for attributing gains to the proposed modeling approach rather than to additional physiological inputs.
Questions for authors

Q1. How much of the generation improvement remains when the handcrafted descriptor pathway is removed or replaced by a deliberately weak/shuffled descriptor bank? The current input-source analysis is informative, but I would like to better isolate whether the proposed “EEG understanding” is primarily driven by the engineered physiological covariates. Could the authors report waveform-only generation under the same training target and adaptation scheme, and compare it with descriptor-only, full, shuffled-covariate, and possibly deterministic feature-to-text baselines? This would more directly establish that the model learns waveform-grounded information beyond structured feature narration. Q2. Can the authors provide broader evidence that the projected adaptation preserves general language ability? OWT ΔPPL and BERTScore are useful local diagnostics, but they cover a narrow reference distribution. Could the authors evaluate the pretrained and adapted backbones on a held-out set of general instructions or standard language/reasoning tasks? It would also be useful to report whether the unprojected residual branch alone causes measurable forgetting and how performance changes when projection is applied to a larger fraction of trainable parameters. Q3. How should the main classification gains be interpreted when auxiliary modalities are available to Ewen but not to the strongest EEG-only baselines? The covariate-free appendix comparison is helpful but is limited to three datasets. Could the authors provide input-matched results for every benchmark containing EOG/ECG/EMG, or alternatively compare against multimodal baselines that use the same auxiliary channels? This would allow the contribution of the architecture to be separated from the information advantage of additional biosignals. Q4. What is the intended scope of the term “EEG understanding”? The current evaluation focuses mainly on class labels, band structure, temporal changes, spatial contrasts, and cross-signal relationships. These are valuable forms of signal interpretation, but they remain constrained to a predefined family of measurable physiological attributes. Would the authors consider narrowing the claim to “physiologically grounded EEG description/interpretation,” or provide evidence for transfer to unseen concepts, unseen descriptive relations, or open-ended questions not represented in the training template bank?

4 comments
AuthorsOpenReview ↗

Response to Reviewer Fb95 [1/3]

Overview

Thank you for your comments and suggestions. Post 1: Weakness 1/Question 1 (W1/Q1, input controls); Post 2: Weakness 2/Question 2 (W2/Q2, projection and general-language evaluation); Post 3: Weakness 2/Question 3 (W2/Q3, input fairness) and Weakness 1/Question 4 (W1/Q4, claim scope). Please refer to the updated PDF for detailed revisions, experiments, and analysis. Thank you also for recognizing the physiological motivation and detailed generation evaluation.

Response to Weakness 1/Question 1 (W1/Q1): Physiological Information and Waveform Relations

Summary

The respective contributions of waveform learning, engineered descriptors, and structured supervision merit attention. Input-source and covariate-perturbation controls are suggested to help distinguish them.

Response

Thank you for suggesting these controls. Table 1 compares Waveform-only, Waveform + JSON, and Full Ewen.

Table 1: Structured JSON Input Experiment

MethodDatasetDirect F1 ↑Withheld F1 ↑Summary R-L ↑
Waveform-onlyMumtaz 20160.5621 ± 0.02100.5009 ± 0.01050.8288 ± 0.0158
Mental Arithmetic0.4388 ± 0.01310.3969 ± 0.00790.7827 ± 0.0105
SHU-MI0.5079 ± 0.02270.4117 ± 0.00520.7826 ± 0.0140
Dataset-macro0.5026 ± 0.01400.4371 ± 0.00440.7981 ± 0.0092
Waveform + JSONMumtaz 20160.8774 ± 0.01580.5231 ± 0.01300.8238 ± 0.0125
Mental Arithmetic0.7177 ± 0.01750.3977 ± 0.01260.8280 ± 0.0085
SHU-MI0.8504 ± 0.02100.3989 ± 0.00940.8233 ± 0.0189
Dataset-macro0.8152 ± 0.01050.4399 ± 0.00680.8250 ± 0.0081
Full EwenMumtaz 20160.9414 ± 0.01050.6137 ± 0.01550.8857 ± 0.0091
Mental Arithmetic0.7780 ± 0.02190.4861 ± 0.01720.8998 ± 0.0064
SHU-MI0.8663 ± 0.01920.5163 ± 0.01350.8931 ± 0.0237
Dataset-macro0.8619 ± 0.01660.5387 ± 0.01200.8929 ± 0.0113

JSON raises Direct F1 from 0.5026 to 0.8152, showing that explicit covariates supply part of the attribute information. Full Ewen reaches 0.8619, consistent with learning correspondence among physiological conditions, waveforms, and descriptions. For Withheld F1, JSON's 0.4399 remains close to Waveform-only's 0.4371, whereas Ewen reaches 0.5387. Whole-window and channel summaries do not directly specify local spatiotemporal differences. The gains support recovering unprovided relations by combining physiological context with waveform structure.

For Table 2, we retain the receiving window's waveform and targets while permuting the physiological context and its associated descriptor text.

Table 2: Test-Time Covariate Permutation Experiment

MethodDatasetDirect F1 ↑Withheld F1 ↑Summary R-L ↑
Full Ewen (permuted)Mumtaz 20160.5197 ± 0.00180.4854 ± 0.00470.7832 ± 0.0029
Mental Arithmetic0.4278 ± 0.00660.3730 ± 0.01400.7472 ± 0.0136
SHU-MI0.5121 ± 0.00170.3616 ± 0.00290.7386 ± 0.0084
Dataset-macro0.4865 ± 0.00200.4067 ± 0.00400.7563 ± 0.0053
Full Ewen (clean)Mumtaz 20160.9414 ± 0.01050.6137 ± 0.01550.8857 ± 0.0091
Mental Arithmetic0.7780 ± 0.02190.4861 ± 0.01720.8998 ± 0.0064
SHU-MI0.8663 ± 0.01920.5163 ± 0.01350.8931 ± 0.0237
Dataset-macro0.8619 ± 0.01660.5387 ± 0.01200.8929 ± 0.0113

Permutation reduces Direct F1 from 0.8619 to 0.4865, Withheld F1 from 0.5387 to 0.4067, and Summary R-L from 0.8929 to 0.7563. Mismatched covariates interfere with attribute expression and relation assessment, indicating that some descriptive information comes from context. Permuted Direct/Withheld F1 is slightly below Waveform-only's 0.5026/0.4371, suggesting interference with waveform-based judgments.

With matched context, Ewen exceeds JSON on both Direct and Withheld F1. Together with Waveform-only's recovery of some attributes and relations, this supports learned correspondence and the complementary roles of waveform information and physiological context.

These controls complement the original descriptor-only analysis and cover input removal, structured-text input, and test-time permutation.

Modification

Section 4.4, Table 4 (p. 8; “Input ablations for physiological interpretation”), presents these controls by dataset with clean Ewen as the shared reference, and analyzes information sources (starting at lines 406 and 425), learned correspondence, and complementarity.

AuthorsOpenReview ↗

Response to Reviewer Fb95 [2/3]

Response to Weakness 2/Question 2 (W2/Q2): Language Drift and Projection Scope

Summary

Language retention beyond OWT and gradient diagnostics, and the effects of unprojected residual updates and projection scope on language behavior, warrant attention.

Response

Thank you for your suggestions on the language-evaluation scope and the projected parameter branches.

To examine the practical effect of projection, we added the comparison of five adaptation strategies in Table 3, reporting physiological-relation recovery, summary quality, and language drift together.

Core+residual (no proj.) is Ewen with only gradient projection disabled. Its core+residual adapter architecture and training settings are unchanged.

Table 3: Additional Adaptation Methods and Projection Controls

MethodDatasetWithheld F1 ↑Summary R-L ↑OWT ΔPPL ↓
Core+residual (no proj.)Mumtaz 20160.5980 ± 0.01680.8856 ± 0.0101—
Mental Arithmetic0.4745 ± 0.01860.8995 ± 0.0073—
SHU-MI0.5027 ± 0.01470.8920 ± 0.0248—
Dataset-macro0.5251 ± 0.01410.8924 ± 0.01310.1278 ± 0.0017
Mixed EEG–textMumtaz 20160.5925 ± 0.01740.8856 ± 0.0108—
Mental Arithmetic0.4685 ± 0.01910.8993 ± 0.0070—
SHU-MI0.4991 ± 0.01590.8915 ± 0.0251—
Dataset-macro0.5200 ± 0.01630.8921 ± 0.01350.1046 ± 0.0013
Replay FTMumtaz 20160.5809 ± 0.01900.8855 ± 0.0113—
Mental Arithmetic0.4618 ± 0.02020.8992 ± 0.0079—
SHU-MI0.4883 ± 0.01660.8906 ± 0.0257—
Dataset-macro0.5103 ± 0.01800.8918 ± 0.01300.0538 ± 0.0012
KL-regularized FTMumtaz 20160.5909 ± 0.01690.8856 ± 0.0103—
Mental Arithmetic0.4665 ± 0.01830.8993 ± 0.0072—
SHU-MI0.4961 ± 0.01480.8911 ± 0.0245—
Dataset-macro0.5178 ± 0.01390.8920 ± 0.00950.0494 ± 0.0011
EwenMumtaz 20160.6137 ± 0.01550.8857 ± 0.0091—
Mental Arithmetic0.4861 ± 0.01720.8998 ± 0.0064—
SHU-MI0.5163 ± 0.01350.8931 ± 0.0237—
Dataset-macro0.5387 ± 0.01200.8929 ± 0.01130.0402 ± 0.0011
↑ higher is better, ↓ lower is better.

In the same-architecture comparison, Core+residual (no proj.) and Ewen achieve Withheld F1 scores of 0.5251 and 0.5387 and OWT ΔPPL values of 0.1278 and 0.0402, respectively, with similar Summary R-L (0.8924 versus 0.8929). These results suggest that projection helps reduce language drift while retaining EEG-relation recovery.

The other language-retention strategies provide additional points of comparison. Mixed EEG–text, Replay FT, and KL-regularized FT achieve ΔPPL values of 0.1046, 0.0538, and 0.0494 and Withheld F1 scores of 0.5200, 0.5103, and 0.5178, respectively. In this comparison, Ewen achieves both lower ΔPPL and higher Withheld F1.

Post-adaptation general-language evaluation

Following your suggestion, we added post-adaptation evaluations on WikiText, C4, MMLU, ARC-C, and IFEval, covering language modeling, reasoning, and instruction-following. These evaluations directly assess the final model’s language behavior after the complete adaptation procedure.

Ewen achieves a task–language trade-off: strong EEG classification and physiological-relation recovery together with substantially better general-language retention than Full FT. Appendix R shows that Full FT substantially affects language modeling, knowledge, and instruction-following. Ewen outperforms Full FT throughout this evaluation and LoRA on IFEval, while LoRA has lower perplexity and higher MMLU/ARC-C scores. These results show Ewen’s balance between effective EEG adaptation and retained general-language capabilities.

Together with the same-architecture projection control, these evaluations provide behavioral evidence beyond OWT and gradient diagnostics. Residual-only and wider-projection controls remain future directions due to time limitations. Detailed results and analysis are reported in Appendix R: Post-Adaptation General Language Evaluation.

Modification

The post-adaptation evaluations of language modeling, reasoning, and instruction-following are reported in a dedicated Appendix R (p. 48, Table 26; analysis from line 2556). Section 4.5, Table 5 (p. 9; analysis from line 459) retains the complementary same-architecture and adaptation-strategy comparisons.

AuthorsOpenReview ↗

Response to Reviewer Fb95 [3/3]

Response to Weakness 1/Question 4 (W1/Q4): Scope of the Claim

Summary

The claim could be aligned more closely with the predefined attributes and relations evaluated, for example by using “physiological interpretation.”

Response

Thank you for your suggestion on the scope of the claim. Following this suggestion, we have revised the paper’s title to:

Ewen: A Covariate-Aware EEG Foundation Model for Physiological Interpretation

This wording more closely reflects the evaluations in the paper: physiological attribute expression, recovery of predefined spatiotemporal relations from waveforms, and improved descriptions using matched physiological context. The revised title identifies the research goal as physiological interpretation across a broad range of BCI settings.

Modification

The title and relevant claims in the abstract (p. 1), Section 1 (pp. 1–2), the method (pp. 3–5), and the experiments (pp. 5–10) have been updated accordingly. Section 4.4 (pp. 7–8; protocol from line 351, analysis from line 406)** presents the evaluation and analysis of Physiological Interpretation.

Response to Weakness 2/Question 3 (W2/Q3): Auxiliary Modalities and Input Fairness

Summary

The review highlights input fairness when Ewen can access auxiliary EOG/ECG/EMG signals, and the coverage of input-matched comparisons.

Response

Thank you for your suggestion regarding input fairness.

Taken together, Table 1 and Appendix K.2 provide EEG-only results on four main benchmarks: SHU-MI, Mumtaz 2016, SEED-V, and Mental Arithmetic. These tasks span motor imagery, depression detection, emotion recognition, and cognitive workload.

SHU-MI, Mumtaz 2016, and SEED-V use only EEG in the main comparison, where Ewen-L achieves the highest means on all nine corresponding metrics. For Mental Arithmetic, Appendix K.2 separately reports a covariate-free EEG-only control: Ewen-B exceeds NeuroLM and ECHO on balanced accuracy and PR AUC. For example, its balanced accuracy is 0.7184, versus 0.6851 for ECHO. The main-table Arithmetic configuration includes ECG. The EEG-only evidence here comes from this separate control.

The covariate-free comparison in Appendix K.2 itself includes Mental Arithmetic, SHU-MI, and Mumtaz 2016, with Ewen-B leading seven of nine metrics. Combined with the main-table EEG-only SEED-V results, the evidence covers four distinct main benchmarks. These results support Ewen’s competitiveness across tasks and show that its performance does not depend entirely on additional physiological modalities. In the framework, EEG-intrinsic descriptors provide the basic conditioning, with auxiliary modalities as optional extensions. The same framework applies when only EEG is available.

Modification

Section 4.2 (p. 7, line 324) points to EEG-only comparisons. Appendix K.2 (p. 40, starting at line 2139) now identifies the four main benchmarks covered jointly by the main results and covariate-free controls, and distinguishes the two Arithmetic input settings.

Reviewer_Fb95OpenReview ↗

Thanks

I would like to thank the authors for their detailed feedback. I am positive about the work and maintain my current positive score.

Reviewer_wpSZOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This article presents Ewen, which combines EEG representations with physiological descriptors extracted from EEG, EOG, ECG, and EMG recordings. Covariate modulation and temporal–channel encoding support classification and natural-language description. The language backbone is adapted through language-reference-guided gradient projection to control language drift during task learning. The model is evaluated on eight EEG classification benchmarks. Its generated descriptions are assessed through physiological attribute recovery and expert review, while general-text perplexity is used to measure language drift.

Critical strengths
  1. Integration of physiological information with EEG representations. Descriptors derived from EEG and other physiological signals provide complementary information. The ablation studies support their value for EEG analysis.

  2. Support for both classification and natural-language description. The framework supports classification and physiological description, expressing spectral, spatial, and cross-signal relationships in natural language and broadening its potential applications.

  3. Relatively comprehensive evaluation of generation. Beyond text similarity metrics, the evaluation includes recovery of physiological attributes not explicitly provided as inputs, expert review, and general-text perplexity to assess physiological consistency and language drift.

Critical weaknesses
  1. The distinction from existing EEG description methods remains unclear. NeuroNarrator, CELM, and CerebraGloss have already explored EEG report generation or fine-grained interpretation. Incorporating physiological covariates is valuable, but the paper does not sufficiently establish which additional descriptive capabilities this design enables or how much these capabilities depend on the covariates.

  2. The independent effect of gradient projection requires a more direct ablation. The LoRA comparison does not use the same adapter structure, making it difficult to attribute changes in task performance and language drift specifically to gradient projection. A comparison with and without projection under the same Ewen adapter structure would help isolate its contribution.

  3. Limited comparison with recent EEG foundation models. The experiments do not include newer EEG foundation models, limiting the assessment of the model’s classification competitiveness.

Questions for authors
  1. What additional physiological description capabilities does Ewen offer compared with NeuroNarrator, CELM, and CerebraGloss? Could direct comparisons or targeted analyses clarify the gains from physiological covariates?

  2. Could the authors compare adaptation with and without language-reference gradient projection while keeping Ewen’s adapter structure and training settings unchanged, reporting both task performance and language drift?

  3. Could the authors add comparisons with more recent EEG foundation models?

4 comments
AuthorsOpenReview ↗

Response to Reviewer wpSZ [1/3]

Overview

Thank you for your comments and suggestions. Post 1: Weakness 1/Question 1 (W1/Q1, related work and input controls); Post 2: Weakness 2/Question 2 (W2/Q2, projection); Post 3: Weakness 3/Question 3 (W3/Q3, recent baselines). Please refer to the updated PDF for detailed revisions, experiments, and analysis. Thank you also for recognizing the physiological integration and evaluation.

Response to Weakness 1/Question 1 (W1/Q1): Related Work and Physiological Conditioning

Summary

Ewen’s distinction from EEG description methods and the capabilities and contributions of physiological covariates need further explanation.

Response

Thank you for your suggestion. We added related-work comparisons, JSON-input controls, and covariate permutation.

NeuroNarrator focuses on alignment between EEG and scalp topographic maps, together with cross-segment context across EEG domains, including BCI-related tasks. CELM focuses on sequence aggregation and multisection report generation for long-duration clinical EEG. CerebraGloss focuses on fine-grained instruction understanding for EEG visualizations. Ewen conditions on intrinsic EEG descriptors and available auxiliary physiological descriptors, incorporating them into waveform representations through covariate modulation and temporal/channel structure. Its evaluation distinguishes directly provided attributes from waveform relationships that are not directly provided.

Ewen studies physiological interpretation across broad BCI settings, including cognitive workload, motor imagery, emotion, and imagined speech. Its scope extends beyond long-duration clinical reporting. CELM freezes its language backbone; Ewen instead studies task performance and language drift on a general reference distribution when language-side adaptation is permitted.

Table 1: Structured JSON Input Experiment

MethodDatasetDirect F1 ↑Withheld F1 ↑Summary R-L ↑
Waveform-onlyMumtaz 20160.5621 ± 0.02100.5009 ± 0.01050.8288 ± 0.0158
Mental Arithmetic0.4388 ± 0.01310.3969 ± 0.00790.7827 ± 0.0105
SHU-MI0.5079 ± 0.02270.4117 ± 0.00520.7826 ± 0.0140
Dataset-macro0.5026 ± 0.01400.4371 ± 0.00440.7981 ± 0.0092
Waveform + JSONMumtaz 20160.8774 ± 0.01580.5231 ± 0.01300.8238 ± 0.0125
Mental Arithmetic0.7177 ± 0.01750.3977 ± 0.01260.8280 ± 0.0085
SHU-MI0.8504 ± 0.02100.3989 ± 0.00940.8233 ± 0.0189
Dataset-macro0.8152 ± 0.01050.4399 ± 0.00680.8250 ± 0.0081
Full EwenMumtaz 20160.9414 ± 0.01050.6137 ± 0.01550.8857 ± 0.0091
Mental Arithmetic0.7780 ± 0.02190.4861 ± 0.01720.8998 ± 0.0064
SHU-MI0.8663 ± 0.01920.5163 ± 0.01350.8931 ± 0.0237
Dataset-macro0.8619 ± 0.01660.5387 ± 0.01200.8929 ± 0.0113

JSON raises macro Direct F1 from 0.5026 to 0.8152; Full Ewen reaches 0.8619. Their Withheld F1 scores are 0.4399 and 0.5387. Ewen improves both attribute expression and recovery of unprovided relations, consistent with learning correspondence among physiological conditions, waveforms, and descriptions.

Table 2: Test-Time Covariate Permutation Experiment

MethodDatasetDirect F1 ↑Withheld F1 ↑Summary R-L ↑
Full Ewen (permuted)Mumtaz 20160.5197 ± 0.00180.4854 ± 0.00470.7832 ± 0.0029
Mental Arithmetic0.4278 ± 0.00660.3730 ± 0.01400.7472 ± 0.0136
SHU-MI0.5121 ± 0.00170.3616 ± 0.00290.7386 ± 0.0084
Dataset-macro0.4865 ± 0.00200.4067 ± 0.00400.7563 ± 0.0053
Full Ewen (clean)Mumtaz 20160.9414 ± 0.01050.6137 ± 0.01550.8857 ± 0.0091
Mental Arithmetic0.7780 ± 0.02190.4861 ± 0.01720.8998 ± 0.0064
SHU-MI0.8663 ± 0.01920.5163 ± 0.01350.8931 ± 0.0237
Dataset-macro0.8619 ± 0.01660.5387 ± 0.01200.8929 ± 0.0113

In both tables, bold marks the best mean for each dataset and metric; ↑ higher is better.

Permuted Direct/Withheld F1 (0.4865/0.4067) falls below matched inputs (0.8619/0.5387) and slightly below Waveform-only (0.5026/0.4371). Covariates supply part of the descriptive information, and mismatched context can interfere with attribute and relation judgments. Together with Waveform-only's retained recovery, these controls support complementary roles for waveform information and matched physiological context.

Modification

Section 2 (p. 3, line 136) explains the scientific and technical distinctions. Section 4.4, Table 4 (p. 8; analysis from lines 406 and 425) reports both input controls against clean Ewen and analyzes information sources, learned correspondence, and complementarity.

AuthorsOpenReview ↗

Response to Reviewer wpSZ [2/3]

Response to Weakness 2/Question 2 (W2/Q2): Projection Controls with the Same Architecture and the Task–Language Trade-off

Summary

The LoRA comparison uses a different adapter structure. Keeping Ewen’s architecture and training settings unchanged would allow a more direct assessment of projection’s independent effect on task performance and language drift.

Response

Thank you for this suggestion. Following it, we added a projection on/off comparison that keeps the adapter architecture and training settings unchanged. Table 3 also includes Core+residual (no proj.), Mixed EEG–text, Replay FT, KL-regularized FT, and Ewen to compare the adaptation strategies.

Core+residual (no proj.) is Ewen with only gradient projection disabled. Its core+residual adapter architecture and training settings are unchanged.

Table 3: Additional Adaptation Methods and Projection Controls

MethodDatasetWithheld F1 ↑Summary R-L ↑OWT ΔPPL ↓
Core+residual (no proj.)Mumtaz 20160.5980 ± 0.01680.8856 ± 0.0101—
Mental Arithmetic0.4745 ± 0.01860.8995 ± 0.0073—
SHU-MI0.5027 ± 0.01470.8920 ± 0.0248—
Dataset-macro0.5251 ± 0.01410.8924 ± 0.01310.1278 ± 0.0017
Mixed EEG–textMumtaz 20160.5925 ± 0.01740.8856 ± 0.0108—
Mental Arithmetic0.4685 ± 0.01910.8993 ± 0.0070—
SHU-MI0.4991 ± 0.01590.8915 ± 0.0251—
Dataset-macro0.5200 ± 0.01630.8921 ± 0.01350.1046 ± 0.0013
Replay FTMumtaz 20160.5809 ± 0.01900.8855 ± 0.0113—
Mental Arithmetic0.4618 ± 0.02020.8992 ± 0.0079—
SHU-MI0.4883 ± 0.01660.8906 ± 0.0257—
Dataset-macro0.5103 ± 0.01800.8918 ± 0.01300.0538 ± 0.0012
KL-regularized FTMumtaz 20160.5909 ± 0.01690.8856 ± 0.0103—
Mental Arithmetic0.4665 ± 0.01830.8993 ± 0.0072—
SHU-MI0.4961 ± 0.01480.8911 ± 0.0245—
Dataset-macro0.5178 ± 0.01390.8920 ± 0.00950.0494 ± 0.0011
EwenMumtaz 20160.6137 ± 0.01550.8857 ± 0.0091—
Mental Arithmetic0.4861 ± 0.01720.8998 ± 0.0064—
SHU-MI0.5163 ± 0.01350.8931 ± 0.0237—
Dataset-macro0.5387 ± 0.01200.8929 ± 0.01130.0402 ± 0.0011

↑ higher is better, ↓ lower is better.

In the same-architecture comparison, Ewen increases dataset-macro Withheld F1 from 0.5251 without projection to 0.5387, with numerical gains on all three datasets; Summary R-L remains similar (0.8924 versus 0.8929). OWT ΔPPL decreases from 0.1278 to 0.0402, a relative reduction of approximately 68.5%. This comparison provides more direct evidence for the independent contribution of projection to reducing language drift while maintaining relation recovery.

The other adaptation strategies also help reduce drift. Mixed EEG–text, Replay FT, and KL-regularized FT achieve ΔPPL values of 0.1046, 0.0538, and 0.0494 and Withheld F1 scores of 0.5200, 0.5103, and 0.5178, respectively. Ewen achieves 0.0402 and 0.5387, with similar Summary R-L across methods (0.8918–0.8929). Across these comparisons, Ewen achieves lower language drift on the general reference distribution together with higher relation-recovery performance. This observation is consistent with constraining reference-sensitive core updates while retaining task-adaptation capacity through the residual.

Modification

These comparisons are now included in Table 5 (“Physiological interpretation and language drift across adaptation strategies,” p. 9) in Section 4.5, with analysis (starting at line 459) of relation recovery, summary quality, and language drift for no projection, mixed EEG–text, replay, KL regularization, and Ewen.

AuthorsOpenReview ↗

Response to Reviewer wpSZ [3/3]

Response to Weakness 3/Question 3 (W3/Q3): Comparisons with Recent EEG Foundation Models

Summary

The review notes limited coverage of recent EEG foundation models and suggests including them to assess classification competitiveness.

Response

Thank you for this suggestion. We have added CodeBrain, together with BrainPro and LEAF, both recently accepted at NeurIPS 2026, following each method’s official settings. Their final results across six datasets have been incorporated into manuscript Table 1, and the table below provides the complete comparison of these three recent baselines with Ewen-L.

Table 4: Comparison of CodeBrain, BrainPro, LEAF, and Ewen-L

DatasetTUABMental Arithmetic
MethodACC-B ↑ROC AUC ↑PR AUC ↑ACC-B ↑ROC AUC ↑PR AUC ↑
CodeBrain0.7726 ± 0.00510.8481 ± 0.00470.8477 ± 0.00730.6125 ± 0.10960.8045 ± 0.07220.6334 ± 0.0738
BrainPro0.8052 ± 0.00430.8946 ± 0.00110.8920 ± 0.00070.5139 ± 0.01200.8313 ± 0.01710.7312 ± 0.0271
LEAF0.8106 ± 0.00760.8856 ± 0.00900.8869 ± 0.01300.6574 ± 0.03990.7402 ± 0.00370.5186 ± 0.0422
Ewen-L0.8149 ± 0.01860.8990 ± 0.00500.8867 ± 0.00460.8790 ± 0.01250.9317 ± 0.00800.9635 ± 0.0095
DatasetSHU-MIMumtaz 2016
MethodACC-B ↑ROC AUC ↑PR AUC ↑ACC-B ↑ROC AUC ↑PR AUC ↑
CodeBrain0.6116 ± 0.01270.6764 ± 0.02430.6867 ± 0.01730.9107 ± 0.01170.9785 ± 0.00830.9823 ± 0.0061
BrainPro0.6126 ± 0.00520.6791 ± 0.00650.6793 ± 0.00620.8944 ± 0.00670.9629 ± 0.00300.9694 ± 0.0032
LEAF0.6144 ± 0.00660.6646 ± 0.01460.6806 ± 0.01240.9409 ± 0.01200.9844 ± 0.00050.9853 ± 0.0004
Ewen-L0.6362 ± 0.01670.6929 ± 0.00570.6869 ± 0.00480.9415 ± 0.01240.9903 ± 0.00160.9893 ± 0.0024
DatasetBCIC IV 2aSEED-V
MethodACC-B ↑Cohen’s Kappa ↑Weighted F1 ↑ACC-B ↑Cohen’s Kappa ↑Weighted F1 ↑
CodeBrain0.3658 ± 0.03420.1544 ± 0.04570.2843 ± 0.05300.2550 ± 0.00480.0709 ± 0.00460.2556 ± 0.0042
BrainPro0.5292 ± 0.06650.3723 ± 0.08870.5233 ± 0.07150.2526 ± 0.00640.0694 ± 0.00440.2468 ± 0.0086
LEAF0.6082 ± 0.01130.4776 ± 0.01510.6068 ± 0.01330.2542 ± 0.00930.0702 ± 0.00940.2552 ± 0.0020
Ewen-L0.5014 ± 0.01760.3312 ± 0.01730.4971 ± 0.01560.2558 ± 0.01550.0711 ± 0.01830.2562 ± 0.0004

↑ higher is better, ↓ lower is better.

These results show how the models’ relative strengths vary by task and metric. In the four-method comparison above, Ewen-L achieves the highest means on all three metrics for Mental Arithmetic, SHU-MI, Mumtaz 2016, and SEED-V, as well as TUAB ACC-B and ROC AUC.

BrainPro achieves the highest mean on TUAB PR AUC (0.8920); LEAF leads all three BCIC IV 2a metrics.

In manuscript Table 1, which includes all original baselines, Ewen-L achieves the highest means on 13 of the 18 metrics, while LaBraM leads TUAB ROC AUC and PR AUC. Taken together, these comparisons support Ewen’s competitive performance across cognitive workload, motor imagery, and emotion recognition. The physiological-description and relation-recovery results provide an accompanying assessment of interpretation beyond classification.

Modification

We have incorporated the final CodeBrain, BrainPro, and LEAF results into manuscript Table 1 (p. 6) and updated the result analysis in Section 4.2 (pp. 6–7, starting at line 322).

AuthorsOpenReview ↗

Rebuttal and updated manuscript submitted

Dear Reviewer wpSZ,

As the discussion period closes on Nov 18, we would like to briefly summarize how our responses and updated manuscript address your three main concerns. Thank you again for your thoughtful feedback which has helped us further strengthen the analysis of Ewen.

  1. Distinction from existing EEG description methods.

NeuroNarrator focuses on alignment between EEG and scalp topographic maps. CELM focuses on sequence aggregation and multisection report generation for long-duration clinical EEG. CerebraGloss focuses on fine-grained instruction understanding for EEG visualizations. Ewen studies physiological interpretation across broad BCI settings, including cognitive workload, motor imagery, emotion, and imagined speech, with a scope extending beyond clinical reporting. It also examines the trade-off between EEG task performance and language drift on a general reference distribution when the language backbone is adapted.

  1. Independent effect of gradient projection.

We introduced Core+residual (no proj.), Ewen’s no-projection version, keeping the adapter structure and training settings unchanged. This direct comparison shows that projection reduces general-reference language drift while improving Withheld F1. Table 5 also compares alternative adaptation strategies, providing further evidence for the balance between physiological task performance and language drift.

  1. Comparisons with recent EEG foundation models.

We have added CodeBrain, together with BrainPro and LEAF, both recently accepted at NeurIPS 2026. With both the original and new baselines included, Ewen-L achieves the highest mean on 13 of the 18 reported metrics.

We hope these additional results and explanations address your concerns. Given the approaching deadline, we would greatly appreciate your feedback on any remaining questions or points that need further clarification.

Thank you again for your time and consideration.

Best regards,
The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response and Revision of the Paper

We thank all reviewers for their careful reading and constructive comments. We appreciate their recognition of Ewen's connection between EEG representation learning and physiological interpretation (Fb95, wpSZ), broad benchmark coverage, competitive performance (bUqa), informative ablations (bUqa), and physiological evaluation(Fb95, wpSZ).

The additional experiments and analyses further substantiate Ewen's approach. We broadened comparisons and tested the roles of physiological context and projected adaptation. We summarize the updates in three areas: manuscript's presentation, expanded comparisons, and targeted controls. All locations below refer to the updated PDF.

A. We refined the manuscript's presentation in the following areas

  1. Claim scope. We aligned the title, abstract, introduction, and contributions around physiological interpretation, matching the evaluated attributes and relations while clarifying physiological semantic context. (p. 1, lines 0–2, 11–25, and 45–53; p. 2, lines 84–102.)

  2. Related work. We expanded the discussion of NeuroNarrator's topographic-map alignment, CELM's long-duration clinical reporting, and CerebraGloss's visual instruction understanding. We explained Ewen's physiological conditioning, relation recovery, and adaptation control across broad BCI settings. (Section 2, p. 3, lines 136–144.)

  3. Visual EEG research direction. We discussed rendered multichannel EEG and VLM comparisons as a complementary direction for future evaluation. (Appendix P, p. 43, lines 2294–2295.)

B. We expanded the comparisons, and Ewen retains clear advantages in most of them

  1. Recent foundation models. We added CodeBrain, BrainPro, and LEAF across six main benchmarks, following their official settings. Including original and added baselines, Ewen-L achieves the highest mean on 13 of 18 metrics. (Table 1, p. 6, lines 270–306; analysis, pp. 6–7, lines 322–325.)

  2. Structured JSON input and physiological conditioning. We added Waveform + JSON alongside Waveform-only and Full Ewen on three datasets, reporting Direct F1, Withheld F1, and Summary R-L. JSON improves direct attribute description, while Full Ewen further improves both Direct and Withheld F1. These results support learned correspondence among context, waveforms, and language, including temporal and spatial relations not directly supplied. (Section 4.4, Table 4, p. 8, lines 406–424.)

  3. Alternative adaptation strategies. We added Mixed EEG–text, Replay FT, and KL-regularized FT. Ewen achieves the lowest OWT drift and highest Withheld F1 among these strategies, with similar summary quality. (Table 5, p. 9, lines 444–464.)

C. Targeted controls and analyses further support the design choices

  1. Covariate correspondence. We added a permutation control. Its effects on attribute expression and relation recovery, together with waveform-only results, support complementary roles for waveforms and correctly matched physiological context. (Table 4, p. 8, lines 413–429.)

  2. Gradient projection and complementary evidence. We added Core+residual (no proj.), Ewen's no-projection version, retaining the adapter and training settings. This control shows that projection reduces general-reference language drift while improving Withheld F1. We also explain how Appendix R's post-adaptation evaluations provide complementary evidence on the model's general-language behavior. (Section 4.5, p. 9, lines 435–443 and 459–461; Appendix R, p. 48, lines 2570–2575.)

  3. Post-adaptation evaluation and EEG–language trade-off. We added a dedicated appendix comparing the original backbone, LoRA, Full FT, and Ewen on WikiText, C4, MMLU, ARC-C, and IFEval, beyond OWT and gradient diagnostics. Ewen outperforms LoRA in EEG task-field accuracy and physiological-relation recovery, while Ewen and LoRA show different balances across instruction-following, language modeling, and reasoning. Ewen outperforms Full FT across all general-language metrics. These results demonstrate a trade-off between effective EEG adaptation and retained general-language capabilities. (Table 3, p. 8, lines 393–403; Figure 5, p. 10, lines 503–506; Appendix R, Table 26, p. 48, lines 2540–2568.)

  4. EEG-only evidence and input fairness. We clarified EEG-only evidence on SHU-MI, Mumtaz 2016, SEED-V, and the separate Mental Arithmetic control, supporting effectiveness without auxiliary signals. This control differs from the main-table ECG-inclusive setting. (Appendix K.2, p. 40, lines 2121–2142.)

We hope that Ewen offers useful insights for physiological interpretation across EEG and BCI applications and encourages further research in this direction. We sincerely welcome the reviewers' feedback and constructive suggestions to further strengthen our work.

Best regards,

The authors

AuthorsOpenReview ↗

Rebuttal and updated manuscript submitted — we welcome your further feedback

Dear Reviewers,

Thank you again for your careful review and constructive feedback on Ewen. We have submitted our detailed responses and updated manuscript, including additional experiments and analyses that further support our approach. The main updates include:

  • Research scope and related work: Refined Ewen’s scope and detailed distinctions from related work.
  • Physiological conditioning: Examined the contributions of physiological context and its correspondence with EEG waveforms through the Waveform + JSON comparison and covariate-permutation control.
  • Gradient projection: Assessed the role of projection using Ewen’s no-projection version , alongside alternative adaptation strategies.
  • Recent foundation models: Included CodeBrain, BrainPro, and LEAF across six main benchmarks.
  • EEG–language trade-off: Evaluated general-language capabilities on WikiText, C4, MMLU, ARC-C, and IFEval, alongside EEG performance.
  • Input fairness: Clarified EEG-only settings and supporting comparisons.

With only one day remaining in the rebuttal period, we would greatly appreciate your feedback on whether our responses and updated manuscript address your concerns. Your timely response is crucial.

Thank you for your time and consideration.

Best regards,
The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.