- 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
- The paper provides reasonably comprehensive experiments across multiple EEG benchmarks, with competitive performance against both conventional EEG foundation models and EEG-LLM approaches.
- 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
- 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].
- 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
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
| Method | Dataset | Withheld F1 ↑ | Summary R-L ↑ | OWT ΔPPL ↓ |
|---|---|---|---|---|
| Core+residual (no proj.) | Mumtaz 2016 | 0.5980 ± 0.0168 | 0.8856 ± 0.0101 | — |
| Mental Arithmetic | 0.4745 ± 0.0186 | 0.8995 ± 0.0073 | — | |
| SHU-MI | 0.5027 ± 0.0147 | 0.8920 ± 0.0248 | — | |
| Dataset-macro | 0.5251 ± 0.0141 | 0.8924 ± 0.0131 | 0.1278 ± 0.0017 | |
| Mixed EEG–text | Mumtaz 2016 | 0.5925 ± 0.0174 | 0.8856 ± 0.0108 | — |
| Mental Arithmetic | 0.4685 ± 0.0191 | 0.8993 ± 0.0070 | — | |
| SHU-MI | 0.4991 ± 0.0159 | 0.8915 ± 0.0251 | — | |
| Dataset-macro | 0.5200 ± 0.0163 | 0.8921 ± 0.0135 | 0.1046 ± 0.0013 | |
| Replay FT | Mumtaz 2016 | 0.5809 ± 0.0190 | 0.8855 ± 0.0113 | — |
| Mental Arithmetic | 0.4618 ± 0.0202 | 0.8992 ± 0.0079 | — | |
| SHU-MI | 0.4883 ± 0.0166 | 0.8906 ± 0.0257 | — | |
| Dataset-macro | 0.5103 ± 0.0180 | 0.8918 ± 0.0130 | 0.0538 ± 0.0012 | |
| KL-regularized FT | Mumtaz 2016 | 0.5909 ± 0.0169 | 0.8856 ± 0.0103 | — |
| Mental Arithmetic | 0.4665 ± 0.0183 | 0.8993 ± 0.0072 | — | |
| SHU-MI | 0.4961 ± 0.0148 | 0.8911 ± 0.0245 | — | |
| Dataset-macro | 0.5178 ± 0.0139 | 0.8920 ± 0.0095 | 0.0494 ± 0.0011 | |
| Ewen | Mumtaz 2016 | 0.6137 ± 0.0155 | 0.8857 ± 0.0091 | — |
| Mental Arithmetic | 0.4861 ± 0.0172 | 0.8998 ± 0.0064 | — | |
| SHU-MI | 0.5163 ± 0.0135 | 0.8931 ± 0.0237 | — | |
| Dataset-macro | 0.5387 ± 0.0120 | 0.8929 ± 0.0113 | 0.0402 ± 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.
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
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.
↑ 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.