acceptodds
Under review as a conference paper at ICLR 2027

From Representation to Competence: What Is Missing in EEG Foundation Models?

Abstract

Electroencephalography (EEG) foundation models are pretrained for broad transfer, yet their downstream performance remains uneven. Adding physiological descriptors improves prediction from frozen classifier features, and the same descriptor families are recoverable across the evaluated backbones. How much additional task performance can be recovered from these representations has remained unclear. At several internal locations, readout of the descriptor families is stronger than at the classifier input. A low rank residual adapter is therefore learned within the backbone from task labels, with the pretrained weights and classifier kept fixed. The preferred update depth varies across models and tasks, and task validation yields performance close to that of the better fixed placement. Across four backbones and four tasks, adaptation improves performance in most settings with only a few thousand trainable parameters. At the same budget, internal updates outperform updates at the structured final output in most paired comparisons. These results support placing a small residual adapter inside the backbone at a depth selected by task validation.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG foundation models, transfer learning, physiological readout, representation analysis, internal adaptation.

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_jNpLOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the representation-to-competence gap in EEG foundation models, where physiologically meaningful information encoded within pretrained representations may not be fully utilized by downstream classifiers. Motivated by the observation that certain descriptors are more accessible at internal layers than at the final classifier input, they investigate low-rank residual adaptation at different backbone depths while keeping the pretrained weights and classifier frozen. Experiments across 4 backbones and 4 tasks demonstrate that internal residual adaptation generally outperforms final-output adaptation under matched parameter budgets.

Critical strengths

1.Connecting physiological representation analysis with adaptation placement provides an interesting perspective

2.The paper compares internal and final-output residual adaptation under matched trainable parameter budgets

Critical weaknesses

1.Although SCOPE and NeuroTTT are cited, the proposed internal residual adaptation shares similarities with SCOPE's internal adapter strategy, while NeuroTTT also explores physiologically motivated adaptation objectives and LoRA-based adaptation. The authors should more clearly distinguish their contributions from these existing approaches.

2.The paper motivates internal adaptation using the observation that physiological descriptors are more accessible at certain internal layers. However, the proposed residual adapters are trained solely with downstream task labels, and their placement is selected through task validation rather than physiological readout scores.

3.For CBraMod and CSBrain, the authors show that final-output residual adaptation remains within the affine function class of the reference classifier, whereas internal adaptation benefits from additional nonlinear backbone transformations. Comparisons with nonlinear classification heads and established parameter-efficient fine-tuning strategies would be helpful.

4.Several reported improvements are relatively small, so statistical significance tests or confidence intervals would better show the significance of the performance.

Questions for authors

1.Could the authors further clarify the methodological differences with existing methods such as NeuroTTT and SCOPE?

2.Could the authors quantitatively examine whether physiological readout scores at different layers correlate with the downstream improvements obtained by adapting those layers?

6 comments
AuthorsOpenReview ↗

Response to Reviewer jNpL (1/5), Q1, W1 and Q2

We thank the reviewer. We first answer the two questions and then address the weaknesses in order. The answers refer to the revised manuscript (Section 4.6 and Appendix G). This response has five parts. Each table is reproduced below the answer that uses it, under its number in the revised manuscript.

Q1, W1. Methodological differences from SCOPE and NeuroTTT. The paper measures how much task information the frozen output leaves inside pretrained EEG backbones, and we use the residual as the instrument of this measurement. SCOPE addresses settings with few labels. It freezes the backbone and conditions modules in its deeper layers on prototypes. In this paper, we train a single linear residual from the task labels and study the effect of its site, from PRE to FINAL at a fixed budget of 3,200 parameters. The site has a large effect. For CBraMod on ADFTD, the gains are 3.1 to 4.4 points at blocks 1 to 6 and vanish at blocks 11 and 12 (Table 28, reproduced under Q2).

NeuroTTT trains the whole backbone with the supervised loss and self-supervised objectives and then adapts at test time, with LoRA across blocks as a baseline. In this paper, the backbone and the classifier stay frozen, the task loss is the only training signal, and the physiological descriptors are used as measurements. Used as auxiliary objectives, the descriptors stay below adaptation with the task loss alone in all 12 comparisons (Appendix F). We now compare with LoRA and bottleneck adapters in all blocks in Table 21, reproduced under W3.

Q2. Do readout scores at different layers correlate with the gains from adapting those layers? (Tables 27 and 28) We trained residuals at 13 sites of CBraMod on ADFTD, namely PRE and each of the 12 blocks (Table 28). Across the 13 sites, the Spearman correlation between the frozen readout and the gain is 0.57 for the smoothed spectral family and 0.82 for the morphology family, the two families with a positive D on this task. It is 0.49 and 0.02 for the two families with a negative D. We also trained residuals at PRE, MID and FINAL in 15 settings (Table 27). The best site is an internal one in all 15, and in 12 of them the readout of the family with the largest D on the development folds is also higher at an internal site than at FINAL. Both the readout and the gain are higher at internal sites.

Table 28. Residuals at 13 sites of CBraMod on ADFTD. Scores are pooled-window Macro-F1 (F1) and balanced accuracy (BAcc), reported as mean ± SD across three seeds (%). Gains are relative to the reference. Task validation selection chooses among the 13 sites within each fold and seed.

SiteF1BAccF1 gainBAcc gain
PRE45.90 ± 0.2446.60 ± 0.21+1.85+1.83
Block 147.18 ± 0.7047.50 ± 0.64+3.13+2.73
Block 247.24 ± 0.3247.56 ± 0.29+3.18+2.79
Block 347.84 ± 0.8248.07 ± 0.78+3.79+3.30
Block 447.55 ± 0.2047.84 ± 0.15+3.49+3.08
Block 548.46 ± 0.9848.73 ± 0.81+4.40+3.96
MID (block 6)47.51 ± 0.5647.67 ± 0.51+3.45+2.90
Block 746.27 ± 0.5646.62 ± 0.46+2.22+1.86
Block 846.79 ± 0.4747.16 ± 0.44+2.74+2.39
Block 946.14 ± 0.5046.51 ± 0.57+2.09+1.74
Block 1045.20 ± 0.4245.65 ± 0.39+1.14+0.88
Block 1143.99 ± 0.0944.65 ± 0.11-0.06-0.12
FINAL (block 12)44.02 ± 0.0344.74 ± 0.03-0.04-0.03
Validation selection47.92 ± 0.7648.14 ± 0.64+3.87+3.37

Table 27. Residuals at PRE, MID, and FINAL and task validation selection among the three sites. Each entry gives pooled-window Macro-F1 / balanced accuracy (%), averaged across three seeds. Selection is made within each fold and seed.

ModelTaskFixed PREFixed MIDFixed FINALValidation selection
LaBraMTD-BRAIN63.78 / 62.5164.03 / 62.9363.42 / 62.1664.10 / 62.90
LaBraMADFTD47.86 / 48.3447.70 / 48.0146.09 / 46.8947.70 / 48.01
LaBraMCAUEEG51.09 / 52.0251.37 / 52.4551.33 / 52.3551.37 / 52.45
LaBraMTUSZ44.34 / 44.4545.49 / 45.7745.06 / 45.1545.57 / 45.83
CBraModTD-BRAIN62.46 / 61.6260.61 / 59.4760.01 / 58.7062.37 / 61.50
CBraModADFTD45.90 / 46.6047.51 / 47.6744.02 / 44.7446.93 / 47.34
CBraModCAUEEG51.98 / 53.1952.24 / 53.5050.11 / 51.4051.86 / 53.18
CBraModTUSZ42.64 / 42.7943.22 / 43.4842.69 / 43.1442.57 / 42.71
CSBrainTD-BRAIN53.35 / 53.1852.92 / 52.8452.85 / 52.7753.35 / 53.18
CSBrainADFTD48.97 / 48.8948.93 / 48.8548.73 / 48.6548.96 / 48.88
CSBrainCAUEEG44.04 / 44.2043.93 / 44.0744.01 / 44.1744.02 / 44.17
BrainOmniTD-BRAIN56.03 / 55.7856.06 / 55.8856.06 / 55.8056.08 / 55.85
BrainOmniADFTD52.76 / 52.6453.00 / 52.8752.73 / 52.6452.96 / 52.84
BrainOmniCAUEEG50.20 / 50.8450.88 / 51.5850.35 / 50.9150.88 / 51.58
BrainOmniTUSZ39.34 / 39.3439.27 / 39.3139.03 / 39.0539.48 / 39.48
AuthorsOpenReview ↗

Response to Reviewer jNpL (2/5), W2

W2. Training with task labels alone, and placement by task validation (Tables 18 and 19). The residual is indeed trained with the task loss alone. For placement, however, the two studies of the paper differ. In the location study of Table 1, we assigned the sites before training with a rule that reads D and the readout profile (Appendix A.7). In the depth comparison, the site is selected by task validation. The physiological analyses contribute as measurements at three points. Before adaptation, D names the family that the classifier at POST lacks, and E shows where this family can be read out. For LaBraM and CBraMod, the readout is higher at PRE or MID than at POST in 26 of 32 combinations. After adaptation, D of that family falls in the three settings where the residual raises Macro-F1 by more than 2 points (from 4.38 to 1.86 for CBraMod on ADFTD, from 2.57 to 1.15 for CBraMod on CAUEEG, and from 3.57 to 2.90 for LaBraM on ADFTD), and its readout at POST rises (Tables 18 and 19).

Table 18. Incremental task value D at POST before and after internal adaptation, for the family with the largest development value in Table 9. Each entry gives Macro-F1 / balanced accuracy differences (%). Gain is the assigned-site residual minus the reference. Adapted D is the mean ± SD across three seeds. Settings are ordered by Macro-F1 gain.

ModelTaskFamilyGainFrozen DAdapted D
CBraModADFTDA+3.45 / +2.90+4.38 / +3.88+1.86 ± 0.30 / +1.87 ± 0.29
CBraModCAUEEGR+2.55 / +2.37+2.57 / +2.46+1.15 ± 0.19 / +0.96 ± 0.22
LaBraMADFTDA+2.27 / +1.58+3.57 / +2.74+2.90 ± 0.25 / +2.38 ± 0.16
LaBraMTUSZR+0.91 / +1.21+2.94 / +2.64+2.75 ± 0.18 / +2.46 ± 0.15
CBraModTUSZR+0.78 / +0.57+5.66 / +4.98+4.65 ± 0.38 / +4.33 ± 0.43
BrainOmniCAUEEGR+0.73 / +0.70-0.08 / -0.10+0.21 ± 0.83 / +0.23 ± 0.82
CBraModTD-BRAINR+0.61 / +0.82+5.28 / +5.50+4.81 ± 0.22 / +4.97 ± 0.25
LaBraMTD-BRAINA+0.48 / +0.73+4.22 / +4.44+4.22 ± 0.21 / +4.43 ± 0.19
BrainOmniTUSZR+0.37 / +0.38+1.88 / +1.77+1.52 ± 0.04 / +1.43 ± 0.03
BrainOmniADFTDA+0.32 / +0.31-0.49 / -0.47-0.42 ± 0.06 / -0.38 ± 0.06
BrainOmniTD-BRAINR+0.27 / +0.18+2.37 / +2.05+2.45 ± 0.10 / +2.11 ± 0.11
LaBraMCAUEEGR-0.13 / -0.21+0.57 / +0.44+0.91 ± 0.13 / +0.71 ± 0.16

Table 19. Task scores, incremental task value D, and POST readout for the reference, the assigned-site residual, and LoRA in all attention blocks. Task scores and D are given as Macro-F1 (F1) and balanced accuracy (BAcc). Readout is the ridge readout at POST of the descriptors after subtraction of their class means on the training subjects. Params is the number of trainable parameters. Adapted entries are the mean ± SD across three seeds.

AdaptationParamsScore F1 (%)Score BAcc (%)D F1 (%)D BAcc (%)Readout
LaBraM ADFTD, family A
Frozen045.4346.44+3.57+2.740.1011
Single site residual3,20047.70 ± 0.3148.01 ± 0.24+2.90 ± 0.25+2.38 ± 0.160.1104 ± 0.0023
All block LoRA154,20351.13 ± 3.0151.24 ± 3.18-0.32 ± 0.70-0.31 ± 0.670.1184 ± 0.0069
CBraMod ADFTD, family A
Frozen044.0544.77+4.38+3.880.0907
Single site residual3,20047.51 ± 0.5647.67 ± 0.51+1.86 ± 0.30+1.87 ± 0.290.0968 ± 0.0001
All block LoRA154,20353.17 ± 0.4453.12 ± 0.51+0.45 ± 0.16+0.48 ± 0.170.1079 ± 0.0010
AuthorsOpenReview ↗

Response to Reviewer jNpL (3/5), W3 nonlinear heads

W3. Comparison with nonlinear classification heads (Tables 23 and 24). We trained an MLP classifier on the frozen POST features with the parameter count of the residual. The residual outperforms it in 7 of 8 settings for LaBraM and CBraMod (Table 24). We also added a FINAL residual with a GELU between its two matrices, which lies outside the affine class of Proposition A.8 and covers all 16 settings, including those of CSBrain (Table 23). It differs from the linear FINAL residual by at most 0.21 points in 15 of them, and the internal residual outperforms both in 12 of 14.

Table 24. MLP classifier on frozen POST features at the parameter count of the internal residual. Each entry gives pooled-window Macro-F1 / balanced accuracy (%), averaged across three seeds.

ModelTaskReferenceMLP classifierAssigned site
LaBraMTD-BRAIN63.54 / 62.2060.75 / 59.1164.03 / 62.93
LaBraMADFTD45.43 / 46.4445.57 / 46.3747.70 / 48.01
LaBraMCAUEEG51.50 / 52.6751.05 / 51.6351.37 / 52.45
LaBraMTUSZ44.58 / 44.5742.93 / 43.1245.49 / 45.77
CBraModTD-BRAIN60.00 / 58.6557.84 / 56.2960.61 / 59.47
CBraModADFTD44.05 / 44.7745.23 / 46.2347.51 / 47.67
CBraModCAUEEG49.69 / 51.1348.88 / 50.4152.24 / 53.50
CBraModTUSZ42.43 / 42.9242.51 / 43.6543.22 / 43.48

Table 23. FINAL controls and classification from frozen internal features. Each entry gives pooled-window Macro-F1 / balanced accuracy (%), averaged across three seeds. Reference, assigned site, and FINAL linear follow Table 1, and † marks an assigned site at FINAL. FINAL GELU places a GELU between the two matrices of the FINAL residual. The internal classifier is a logistic classifier on frozen features of one site among PRE, MID, and FINAL, selected by task validation.

ModelTaskReferenceAssigned siteFINAL linearFINAL GELUInternal classifier
LaBraMTD-BRAIN63.54 / 62.2064.03 / 62.9363.42 / 62.1663.51 / 62.2864.93 / 63.67
LaBraMADFTD45.43 / 46.4447.70 / 48.0146.09 / 46.8945.98 / 46.8244.33 / 45.61
LaBraMCAUEEG51.50 / 52.6751.37 / 52.4551.33 / 52.3551.47 / 52.4552.33 / 53.14
LaBraMTUSZ44.58 / 44.5745.49 / 45.7745.06 / 45.1544.86 / 44.9544.52 / 44.80
CBraModTD-BRAIN60.00 / 58.6560.61 / 59.4760.01 / 58.7059.95 / 58.6264.04 / 62.68
CBraModADFTD44.05 / 44.7747.51 / 47.6744.02 / 44.7444.02 / 44.7547.21 / 48.22
CBraModCAUEEG49.69 / 51.1352.24 / 53.5050.11 / 51.4049.98 / 51.3451.89 / 53.26
CBraModTUSZ42.43 / 42.9243.22 / 43.4842.69 / 43.1442.70 / 43.1838.10 / 39.35
CSBrainTD-BRAIN†52.82 / 52.7552.85 / 52.7752.85 / 52.7752.86 / 52.7664.03 / 62.64
CSBrainADFTD48.79 / 48.7148.93 / 48.8548.73 / 48.6548.79 / 48.7142.57 / 44.84
CSBrainCAUEEG†43.97 / 44.1144.01 / 44.1744.01 / 44.1743.93 / 44.0748.90 / 50.58
CSBrainTUSZ37.42 / 37.5036.41 / 37.1934.94 / 36.7335.72 / 37.5638.05 / 39.43
BrainOmniTD-BRAIN55.80 / 55.7056.06 / 55.8856.06 / 55.8055.97 / 55.7653.62 / 52.56
BrainOmniADFTD52.68 / 52.5653.00 / 52.8752.73 / 52.6452.73 / 52.6349.86 / 51.57
BrainOmniCAUEEG50.15 / 50.8850.88 / 51.5850.35 / 50.9150.17 / 50.8953.08 / 54.29
BrainOmniTUSZ38.90 / 38.9339.27 / 39.3139.03 / 39.0538.91 / 38.9441.94 / 42.54
Reviewer_YcWZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the gap between the information encoded in EEG foundation models and its usefulness for downstream classification. It evaluates the predictive value and decodability of handcrafted physiological descriptors, then explores low-rank residual adaptation at different backbone depths. Experiments across four EEG foundation models and four clinical tasks suggest that physiological descriptors can complement frozen output features and that internal adaptation generally outperforms final-layer adaptation.

Critical strengths

S1. Important research question

The paper investigates the gap between information encoded in EEG foundation models and its utilization in downstream classification. This is a relevant question for understanding the transfer capabilities of pretrained EEG representations.

S2. Broad experimental coverage with useful observations

The study evaluates four EEG backbones and four clinical tasks, covering physiological descriptors, layer-wise readout, and adaptation at different depths. The results suggest that handcrafted descriptors can complement frozen features, certain physiological properties are more accessible at intermediate layers, and internal adaptation generally performs better than final-layer adaptation.

Critical weaknesses

W1. Insufficient discussion of relevant prior work

Important studies such as REVE and Are Large Brainwave Foundation Models Capable Yet? Insights from Fine-Tuning receive insufficient attention. Their designs on EEG representation transfer and adaptation strategies are relevant to positioning this work.

W2. Limited methodological novelty and modest performance gains

The proposed update h′=h+BAhh'=h+BAh is a special case of the low-rank representation intervention introduced in ReFT: Representation Finetuning for Language Models. Although internal adaptation often outperforms final-layer adaptation, its improvements over the frozen baseline are relatively small in many settings.

W3. Limited interpretation of the D metric

The D metric measures the classification improvement from adding physiological descriptors Q to frozen POST features. However, it cannot distinguish genuinely missing information from information already encoded but less accessible to a linear classifier. A positive D therefore does not directly establish that the corresponding physiological information is absent from POST representations.

W4. Insufficient evidence linking performance gains to physiological information utilization

The paper does not sufficiently establish that adaptation gains arise from better utilization of the measured physiological information. In Table 3, substantial task improvements sometimes accompany nearly unchanged or reduced proximal physiological readout. Appendix F further shows that explicit physiological auxiliary supervision performs worse than task-only adaptation. The relationship between physiological information utilization and downstream gains remains unclear.

W5. Limited explanatory depth of the theoretical analysis

Appendix A.8 shows that FINAL adaptation preserves affine logits for CBraMod and CSBrain, whereas PRE/MID interventions undergo subsequent nonlinear backbone transformations. This establishes a function-class distinction but does not explain when or why internal adaptation should yield better downstream performance.

W6. Missing baseline using frozen intermediate representations

The paper identifies stronger physiological readout at certain intermediate layers but does not evaluate classifiers that directly use frozen intermediate features, as explored in Head2Toe: Utilizing Intermediate Representations for Better Transfer Learning. Such a comparison would help determine whether representation intervention is necessary to obtain the reported gains.

Questions for authors

Q1. Could the authors further clarify how the incremental predictive value (D) and physiological readout (E) jointly support the proposed representation-to-competence gap? In particular, how should the distinction between missing information and limited linear accessibility be interpreted?

Q2. Could the authors elaborate on how the physiological readout findings relate to the benefits of internal adaptation? The results in Table 3 and Appendix F may help clarify whether these analyses mainly motivate the intervention strategy or also explain its performance improvements.

4 comments
AuthorsOpenReview ↗

Response to Reviewer YcWZ (1/4), Q1 and W3

We thank the reviewer. We first answer the two questions and then address the weaknesses in order. The answers refer to the revised manuscript (Section 4.6 and Appendix G). This response has four parts. Each table is reproduced below the answer that uses it, under its number in the revised manuscript.

Q1, W3. How D and E jointly support the gap, and how to read a positive D (Table 24). The gap is task information that the backbone holds and that its frozen output does not deliver to the classifier. Three measurements support it together. D, the incremental task value, shows that the classifier at POST lacks task information that the descriptors carry. E, the readout, shows that the same descriptor families can be read out inside the backbone. For LaBraM and CBraMod, the readout is higher at PRE or MID than at POST in 26 of 32 combinations (Appendix C.2). The third measurement is new. A logistic classifier on the frozen features of one site, selected among PRE, MID and FINAL by validation, improves over the reference in 10 of 16 settings (Table 23, reproduced under W6). This shows that task information itself can be read linearly inside the backbone, beyond what the output delivers.

We agree that a positive D alone leaves the distinction between missing information and limited linear accessibility open, and Section 3.1 now states this. It is compatible with information that is absent from the backbone, with information that is present inside and does not reach POST, and with information that is present at POST beyond the reach of a linear classifier. Two results help to separate these cases. First, an MLP classifier on POST with the parameter count of the residual stays below the linear classifier on POST with the descriptor family in all 8 settings (Table 24), so a nonlinear classifier of this size does not recover what the descriptors add. Second, D falls after adaptation where the residual helps (Table 18, reproduced under Q2). The residual is a linear function of one frozen hidden state, and the backbone and the classifier stay frozen, so this part of the information was present inside the backbone and had not reached POST.

Table 24. MLP classifier on frozen POST features at the parameter count of the internal residual. Each entry gives pooled-window Macro-F1 / balanced accuracy (%), averaged across three seeds.

ModelTaskReferenceMLP classifierAssigned site
LaBraMTD-BRAIN63.54 / 62.2060.75 / 59.1164.03 / 62.93
LaBraMADFTD45.43 / 46.4445.57 / 46.3747.70 / 48.01
LaBraMCAUEEG51.50 / 52.6751.05 / 51.6351.37 / 52.45
LaBraMTUSZ44.58 / 44.5742.93 / 43.1245.49 / 45.77
CBraModTD-BRAIN60.00 / 58.6557.84 / 56.2960.61 / 59.47
CBraModADFTD44.05 / 44.7745.23 / 46.2347.51 / 47.67
CBraModCAUEEG49.69 / 51.1348.88 / 50.4152.24 / 53.50
CBraModTUSZ42.43 / 42.9242.51 / 43.6543.22 / 43.48
AuthorsOpenReview ↗

Response to Reviewer YcWZ (2/4), Q2 and W4

Q2, W4. Do the readout findings motivate the intervention, or do they also explain the improvement? (Tables 18 and 19) The readout findings motivate placing the update inside the backbone, and the new analyses of Section 4.6 explain a part of the improvement. In the three settings where the residual raises Macro-F1 by more than 2 points, D falls in every seed (from 4.38 to 1.86 for CBraMod on ADFTD, from 2.57 to 1.15 for CBraMod on CAUEEG, and from 3.57 to 2.90 for LaBraM on ADFTD). Among the 10 settings with a positive frozen D, the Spearman correlation between the fall of D and the gain is 0.88 (Table 18). The update thus delivers to the classifier a part of the information that the descriptors add at the frozen POST. On ADFTD, the readout of the smoothed spectral family at POST rises in every seed under the residual and under LoRA in all blocks (Table 19). Across 13 sites of CBraMod on ADFTD, the Spearman correlation between readout and gain is 0.57 and 0.82 for the two families with a positive D (Table 28, reproduced under W5).

In Table 3, the proximal ridge readers were refit on an invertible linear map of the frozen features, so they recover the same fit up to the ridge penalty. This explains the unchanged proximal readout, and we have removed the column. Appendix F compares adaptation with an auxiliary descriptor loss against adaptation with the task loss alone, which scores higher. Relative to the frozen reference, supervision toward the target family still improves Macro-F1 for CBraMod (61.17 against 60.00 on TD-BRAIN and 46.04 against 44.05 on ADFTD), and it scores above supervision toward a shuffled target. This is why we use the descriptors to measure the information that is not delivered and train with the task loss alone.

Table 18. Incremental task value D at POST before and after internal adaptation, for the family with the largest development value in Table 9. Each entry gives Macro-F1 / balanced accuracy differences (%). Gain is the assigned-site residual minus the reference. Adapted D is the mean ± SD across three seeds. Settings are ordered by Macro-F1 gain.

ModelTaskFamilyGainFrozen DAdapted D
CBraModADFTDA+3.45 / +2.90+4.38 / +3.88+1.86 ± 0.30 / +1.87 ± 0.29
CBraModCAUEEGR+2.55 / +2.37+2.57 / +2.46+1.15 ± 0.19 / +0.96 ± 0.22
LaBraMADFTDA+2.27 / +1.58+3.57 / +2.74+2.90 ± 0.25 / +2.38 ± 0.16
LaBraMTUSZR+0.91 / +1.21+2.94 / +2.64+2.75 ± 0.18 / +2.46 ± 0.15
CBraModTUSZR+0.78 / +0.57+5.66 / +4.98+4.65 ± 0.38 / +4.33 ± 0.43
BrainOmniCAUEEGR+0.73 / +0.70-0.08 / -0.10+0.21 ± 0.83 / +0.23 ± 0.82
CBraModTD-BRAINR+0.61 / +0.82+5.28 / +5.50+4.81 ± 0.22 / +4.97 ± 0.25
LaBraMTD-BRAINA+0.48 / +0.73+4.22 / +4.44+4.22 ± 0.21 / +4.43 ± 0.19
BrainOmniTUSZR+0.37 / +0.38+1.88 / +1.77+1.52 ± 0.04 / +1.43 ± 0.03
BrainOmniADFTDA+0.32 / +0.31-0.49 / -0.47-0.42 ± 0.06 / -0.38 ± 0.06
BrainOmniTD-BRAINR+0.27 / +0.18+2.37 / +2.05+2.45 ± 0.10 / +2.11 ± 0.11
LaBraMCAUEEGR-0.13 / -0.21+0.57 / +0.44+0.91 ± 0.13 / +0.71 ± 0.16

Table 19. Task scores, incremental task value D, and POST readout for the reference, the assigned-site residual, and LoRA in all attention blocks. Task scores and D are given as Macro-F1 (F1) and balanced accuracy (BAcc). Readout is the ridge readout at POST of the descriptors after subtraction of their class means on the training subjects. Params is the number of trainable parameters. Adapted entries are the mean ± SD across three seeds.

AdaptationParamsScore F1 (%)Score BAcc (%)D F1 (%)D BAcc (%)Readout
LaBraM ADFTD, family A
Frozen045.4346.44+3.57+2.740.1011
Single site residual3,20047.70 ± 0.3148.01 ± 0.24+2.90 ± 0.25+2.38 ± 0.160.1104 ± 0.0023
All block LoRA154,20351.13 ± 3.0151.24 ± 3.18-0.32 ± 0.70-0.31 ± 0.670.1184 ± 0.0069
CBraMod ADFTD, family A
Frozen044.0544.77+4.38+3.880.0907
Single site residual3,20047.51 ± 0.5647.67 ± 0.51+1.86 ± 0.30+1.87 ± 0.290.0968 ± 0.0001
All block LoRA154,20353.17 ± 0.4453.12 ± 0.51+0.45 ± 0.16+0.48 ± 0.170.1079 ± 0.0010
AuthorsOpenReview ↗

Response to Reviewer YcWZ (3/4), W1, W2 and W5

W1. Discussion of REVE and Lee et al. (2025). We now discuss both papers in the related work. Lee et al. (2025) find that LoRA helps when several components are adapted together. This corresponds to the rows of Table 21 with updates in all blocks, where LoRA outperforms the single site residual by 3.4 and 5.7 points on ADFTD with 48 times as many parameters.

W2. Novelty relative to ReFT and size of the gains (Tables 21 and 22). The update has the form of a ReFT intervention at one site, and the related work now states this. We compared it with LoReFT at the same site and a matched budget. LoReFT is better than the residual for CBraMod and worse for LaBraM (Table 21). The contribution of the paper is the measurement of task information that the frozen output leaves inside the backbone, with the residual as its instrument. On the size of the gains, the residual has 3,200 parameters and keeps the classifier frozen. It recovers 64, 82 and 83 percent of the Macro-F1 gain of full fine-tuning for LaBraM on ADFTD, CBraMod on ADFTD and LaBraM on TUSZ, 35 percent for CBraMod on CAUEEG, and none for LaBraM on CAUEEG (Table 22).

Table 21. Comparison of the internal residual with other updates on ADFTD. Scores are pooled-window Macro-F1 (F1) and balanced accuracy (BAcc), reported as mean ± SD across three seeds (%). * marks updates that train the classifier. Full fine-tuning has 5,820,539 trainable parameters for LaBraM, where the three seeds coincide, and 4,884,203 for CBraMod.

UpdateTrainable parametersLaBraM F1LaBraM BAccCBraMod F1CBraMod BAcc
Reference045.4346.4444.0544.77
Classifier adaptation*60345.3846.4244.0644.76
Internal residual3,20047.70 ± 0.3148.01 ± 0.2447.51 ± 0.5647.67 ± 0.51
LoRA, one block3,20048.97 ± 1.4249.06 ± 1.3245.65 ± 0.4646.36 ± 0.52
LoReFT at MID3,20845.87 ± 1.9346.88 ± 1.5348.28 ± 1.9848.43 ± 1.86
Bottleneck adapters, all blocks*149,76349.43 ± 0.6949.47 ± 0.6147.06 ± 1.9747.24 ± 2.24
LoRA, all blocks*154,20351.13 ± 3.0151.24 ± 3.1853.17 ± 0.4453.12 ± 0.51
Full fine-tuning*5.8M, 4.9M49.0048.8448.25 ± 1.6548.52 ± 1.74

Table 22. Internal residual and full fine-tuning. Each entry gives pooled-window Macro-F1 / balanced accuracy (%), with the SD across three seeds for runs that vary across seeds.

ModelTaskReferenceAssigned siteFull fine-tuning
LaBraMADFTD45.43 / 46.4447.70 ± 0.31 / 48.01 ± 0.2449.00 / 48.84
LaBraMCAUEEG51.50 / 52.6751.37 ± 0.21 / 52.45 ± 0.1754.00 / 54.06
LaBraMTUSZ44.58 / 44.5745.49 ± 0.32 / 45.77 ± 0.3145.68 / 46.42
CBraModADFTD44.05 / 44.7747.51 ± 0.56 / 47.67 ± 0.5148.25 ± 1.65 / 48.52 ± 1.74
CBraModCAUEEG49.69 / 51.1352.24 ± 0.15 / 53.50 ± 0.2156.97 ± 1.79 / 56.52 ± 1.27

W5. When and why internal adaptation helps (Table 28). An update at an internal site is processed by the pretrained layers that come after it, and the internal sites hold task information that the output does not deliver. Updates on the output side stay at the level of the linear FINAL residual. A FINAL residual with a GELU between its two matrices, which lies outside the affine class of Proposition A.8, differs from the linear one by at most 0.21 points in 15 of 16 settings (Table 23, reproduced under W6). The sweep over depth agrees with this account. For CBraMod on ADFTD, the gains are 3.1 to 4.4 points at blocks 1 to 6 and vanish at blocks 11 and 12 (Table 28).

Table 28. Residuals at 13 sites of CBraMod on ADFTD. Scores are pooled-window Macro-F1 (F1) and balanced accuracy (BAcc), reported as mean ± SD across three seeds (%). Gains are relative to the reference. Task validation selection chooses among the 13 sites within each fold and seed.

SiteF1BAccF1 gainBAcc gain
PRE45.90 ± 0.2446.60 ± 0.21+1.85+1.83
Block 147.18 ± 0.7047.50 ± 0.64+3.13+2.73
Block 247.24 ± 0.3247.56 ± 0.29+3.18+2.79
Block 347.84 ± 0.8248.07 ± 0.78+3.79+3.30
Block 447.55 ± 0.2047.84 ± 0.15+3.49+3.08
Block 548.46 ± 0.9848.73 ± 0.81+4.40+3.96
MID (block 6)47.51 ± 0.5647.67 ± 0.51+3.45+2.90
Block 746.27 ± 0.5646.62 ± 0.46+2.22+1.86
Block 846.79 ± 0.4747.16 ± 0.44+2.74+2.39
Block 946.14 ± 0.5046.51 ± 0.57+2.09+1.74
Block 1045.20 ± 0.4245.65 ± 0.39+1.14+0.88
Block 1143.99 ± 0.0944.65 ± 0.11-0.06-0.12
FINAL (block 12)44.02 ± 0.0344.74 ± 0.03-0.04-0.03
Validation selection47.92 ± 0.7648.14 ± 0.64+3.87+3.37
AuthorsOpenReview ↗

Response to Reviewer YcWZ (4/4), W6

W6. Classifier on frozen intermediate features (Table 23). A logistic classifier on the frozen features of a site selected by validation improves over the reference in 10 of 16 settings and outperforms the residual in 8. In those settings the gain is available without an intervention. In 7 settings, however, the residual outperforms this classifier. It keeps the pretrained classifier, and its largest loss to the reference is 1.0 point, against 6.2 for the classifier on frozen features. Both results indicate that the backbone holds task information beyond its frozen output.

Table 23. FINAL controls and classification from frozen internal features. Each entry gives pooled-window Macro-F1 / balanced accuracy (%), averaged across three seeds. Reference, assigned site, and FINAL linear follow Table 1, and † marks an assigned site at FINAL. FINAL GELU places a GELU between the two matrices of the FINAL residual. The internal classifier is a logistic classifier on frozen features of one site among PRE, MID, and FINAL, selected by task validation.

ModelTaskReferenceAssigned siteFINAL linearFINAL GELUInternal classifier
LaBraMTD-BRAIN63.54 / 62.2064.03 / 62.9363.42 / 62.1663.51 / 62.2864.93 / 63.67
LaBraMADFTD45.43 / 46.4447.70 / 48.0146.09 / 46.8945.98 / 46.8244.33 / 45.61
LaBraMCAUEEG51.50 / 52.6751.37 / 52.4551.33 / 52.3551.47 / 52.4552.33 / 53.14
LaBraMTUSZ44.58 / 44.5745.49 / 45.7745.06 / 45.1544.86 / 44.9544.52 / 44.80
CBraModTD-BRAIN60.00 / 58.6560.61 / 59.4760.01 / 58.7059.95 / 58.6264.04 / 62.68
CBraModADFTD44.05 / 44.7747.51 / 47.6744.02 / 44.7444.02 / 44.7547.21 / 48.22
CBraModCAUEEG49.69 / 51.1352.24 / 53.5050.11 / 51.4049.98 / 51.3451.89 / 53.26
CBraModTUSZ42.43 / 42.9243.22 / 43.4842.69 / 43.1442.70 / 43.1838.10 / 39.35
CSBrainTD-BRAIN†52.82 / 52.7552.85 / 52.7752.85 / 52.7752.86 / 52.7664.03 / 62.64
CSBrainADFTD48.79 / 48.7148.93 / 48.8548.73 / 48.6548.79 / 48.7142.57 / 44.84
CSBrainCAUEEG†43.97 / 44.1144.01 / 44.1744.01 / 44.1743.93 / 44.0748.90 / 50.58
CSBrainTUSZ37.42 / 37.5036.41 / 37.1934.94 / 36.7335.72 / 37.5638.05 / 39.43
BrainOmniTD-BRAIN55.80 / 55.7056.06 / 55.8856.06 / 55.8055.97 / 55.7653.62 / 52.56
BrainOmniADFTD52.68 / 52.5653.00 / 52.8752.73 / 52.6452.73 / 52.6349.86 / 51.57
BrainOmniCAUEEG50.15 / 50.8850.88 / 51.5850.35 / 50.9150.17 / 50.8953.08 / 54.29
BrainOmniTUSZ38.90 / 38.9339.27 / 39.3139.03 / 39.0538.91 / 38.9441.94 / 42.54
Reviewer_KZUzOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies frozen EEG foundation models (LaBraM, CBraMod, CSBrain, BrainOmni) on four clinical three-class tasks. It reports three findings:

  • Adding handcrafted physiological descriptors (spectral residuals, smoothed spectrum, channel-correlation geometry, temporal morphology) to frozen POST features improves a logistic classifier.
  • These descriptors are partially recoverable by ridge and MLP probes at internal sites, sometimes more so than at POST.
  • A single rank-8 residual inserted at an internal layer and trained on the task outperforms an equal-budget residual at the final output.

The authors frame this as closing a "representation-to-competence gap."

Critical strengths
  • The question of how much usable task structure frozen EFM outputs leave unexploited is relevant.
  • The descriptor analysis includes a matched-noise control, and the paired readout includes shuffled-target controls.
  • Experimental details are thorough.
Critical weaknesses
  1. The descriptor analyses are disconnected from the adapter. The descriptor results show that some information is complementary to POST features and partially readable internally. Nothing shows that the adapter exploits this information. The paper's own evidence points the other way:

    • The adapter is trained with task loss only, and the paired readout barely changes. Proximal ridge E is unchanged to four decimals on ADFTD (0.2056→0.2056, 0.1707→0.1707), and POST ΔE is at most about 0.01. On TD-BRAIN, a 2.5 pp task gain occurs with essentially no readout change.
    • Explicitly supervising the adapter toward the descriptors hurts relative to task-only adaptation in all 12 comparisons (Table 17).
    • The direct test is missing: recompute the incremental value D on adapted POST features. If the adapter recovers the descriptor information, D should shrink.
    • A descriptor-only classifier baseline is also missing. It would show whether the frozen models simply underperform handcrafted features, particularly the smoothed spectrum on dementia tasks.
  2. The readout evidence is weak. E values are mostly 0.1–0.25. Comparisons across sites are confounded by different input dimensionalities: averaged 200-d tokens internally versus POST features of up to 38,000 dimensions for CSBrain with a fixed ridge λ=1. The negative POST scores for CSBrain likely reflect this mismatch rather than missing information. Probe-capacity controls (Hewitt & Liang, which the paper cites) are not applied to the main readout analysis.

  3. There are no comparisons to standard PEFT. The method is essentially a single linear low-rank residual on hidden states, i.e., a minimal LoRA or adapter variant. The only baselines are the frozen linear probe, classifier re-fitting, and a FINAL residual. For CBraMod and CSBrain, the FINAL residual is, by the authors' own proposition (A.8), within the affine class of classifier adaptation, so "internal beats FINAL" is largely expected by construction. Missing comparisons include:

    • LoRA across all blocks (used in NeuroTTT, which is cited)
    • standard adapters
    • prompt or prefix tuning
    • surgical or partial fine-tuning (Lee et al., cited)
    • LP-FT
    • full fine-tuning, as an upper reference

    Without these, it is unclear whether single-site placement offers anything beyond what existing methods already provide.

  4. Gains are small, and their significance is unclear. Most internal-site gains over the reference are under about 1.5 pp. A few exceptions reach 2.5–3.5 pp (CBraMod ADFTD/CAUEEG, LaBraM ADFTD). In two settings, the internal update is worse than the reference or classifier adaptation (CSBrain TUSZ, LaBraM CAUEEG). The reported SDs reflect only adapter initialization (reference SD = 0), not subject or fold variability, and no paired significance tests across subjects are reported.

  5. The depth analysis is too limited to support the practical claims. It covers only PRE versus MID, on 2 backbones × 2 tasks, with no layer sweep.

    • Validation selection never meaningfully beats the better fixed depth, and it is below fixed MID for CBraMod on ADFTD.
    • Selected depths are inconsistent across folds (Table 12).
    • "Close to the better of two options" is a weak basis for the recommendation to choose depth by validation.
    • The main location study uses a fixed "assigned site" (MID, or FINAL for two CSBrain settings) without explaining how it was chosen.
  6. Framing overclaims. "Competence" and the "representation-to-competence gap" are not operationally defined beyond frozen-probe accuracy. The conclusion that physiological structure "locates" where to adapt is not supported, given points 1 and 5.

Questions for authors
  1. Does the descriptor increment D decrease when computed on adapted POST features?
  2. How does a descriptor-only classifier perform on each task?
  3. How does the method compare with multi-layer LoRA, adapters, and partial or full fine-tuning at matched and unmatched budgets?
  4. How were the assigned sites chosen, and what does a full layer sweep look like?
  5. Are the gains significant under subject-level resampling?
6 comments
AuthorsOpenReview ↗

Response to Reviewer KZUz (1/5), W1 and Q1

We thank the reviewer for the detailed reading and for the specific tests proposed in the review. The results are in Section 4.6 and Appendix G of the revised manuscript. The paper measures how much task information the frozen output leaves inside pretrained EEG backbones, and we use the residual as the instrument of this measurement. This response has five parts. Each table is reproduced below the answer that uses it, under its number in the revised manuscript.

W1. Readout after adaptation (Table 3) and descriptor supervision (Table 17). The proximal ridge readout is unchanged to four decimals by construction. The adapted proximal features are an invertible linear map of the frozen features, so a ridge reader refit on them recovers the same fit up to the ridge penalty. We have removed this column from Table 3. At POST, the ridge readout of the smoothed spectral family rises in every seed on ADFTD, and it also rises under the residual and under LoRA in all blocks when the class means of the descriptors are subtracted before the readout (Table 19). These changes are small, and the test on D proposed in the review is more informative. It is reported under Q1.

In Table 17, adaptation with an auxiliary descriptor loss scores lower than adaptation with the task loss alone in all 12 comparisons. Relative to the frozen reference, however, supervision toward the target family improves Macro-F1 for CBraMod (61.17 against 60.00 on TD-BRAIN and 46.04 against 44.05 on ADFTD), and it scores above supervision toward a shuffled target. This is why we use the descriptors as measurements and train the residual with the task loss alone.

Table 19. Task scores, incremental task value D, and POST readout for the reference, the assigned-site residual, and LoRA in all attention blocks. Task scores and D are given as Macro-F1 (F1) and balanced accuracy (BAcc). Readout is the ridge readout at POST of the descriptors after subtraction of their class means on the training subjects. Params is the number of trainable parameters. Adapted entries are the mean ± SD across three seeds.

AdaptationParamsScore F1 (%)Score BAcc (%)D F1 (%)D BAcc (%)Readout
LaBraM ADFTD, family A
Frozen045.4346.44+3.57+2.740.1011
Single site residual3,20047.70 ± 0.3148.01 ± 0.24+2.90 ± 0.25+2.38 ± 0.160.1104 ± 0.0023
All block LoRA154,20351.13 ± 3.0151.24 ± 3.18-0.32 ± 0.70-0.31 ± 0.670.1184 ± 0.0069
CBraMod ADFTD, family A
Frozen044.0544.77+4.38+3.880.0907
Single site residual3,20047.51 ± 0.5647.67 ± 0.51+1.86 ± 0.30+1.87 ± 0.290.0968 ± 0.0001
All block LoRA154,20353.17 ± 0.4453.12 ± 0.51+0.45 ± 0.16+0.48 ± 0.170.1079 ± 0.0010

Q1, W1. Does D decrease on adapted POST features? (Table 18) D decreases in the settings where the residual helps. We recomputed it on the frozen and on the adapted POST features in 12 settings. In the three settings where the residual raises Macro-F1 by more than 2 points, D falls in every seed. It falls from 4.38 to 1.86 for CBraMod on ADFTD, from 2.57 to 1.15 for CBraMod on CAUEEG, and from 3.57 to 2.90 for LaBraM on ADFTD. For LaBraM on CAUEEG, where the residual generally lowers the scores, D rises from 0.57 to 0.91. Among the 10 settings with a positive frozen D, the scores rise in 9, and D falls in 7 of these 9. The Spearman correlation between the fall of D and the Macro-F1 gain is 0.88. This is the outcome that the test predicts when the adapter recovers descriptor information.

Table 18. Incremental task value D at POST before and after internal adaptation, for the family with the largest development value in Table 9. Each entry gives Macro-F1 / balanced accuracy differences (%). Gain is the assigned-site residual minus the reference. Adapted D is the mean ± SD across three seeds. Settings are ordered by Macro-F1 gain.

ModelTaskFamilyGainFrozen DAdapted D
CBraModADFTDA+3.45 / +2.90+4.38 / +3.88+1.86 ± 0.30 / +1.87 ± 0.29
CBraModCAUEEGR+2.55 / +2.37+2.57 / +2.46+1.15 ± 0.19 / +0.96 ± 0.22
LaBraMADFTDA+2.27 / +1.58+3.57 / +2.74+2.90 ± 0.25 / +2.38 ± 0.16
LaBraMTUSZR+0.91 / +1.21+2.94 / +2.64+2.75 ± 0.18 / +2.46 ± 0.15
CBraModTUSZR+0.78 / +0.57+5.66 / +4.98+4.65 ± 0.38 / +4.33 ± 0.43
BrainOmniCAUEEGR+0.73 / +0.70-0.08 / -0.10+0.21 ± 0.83 / +0.23 ± 0.82
CBraModTD-BRAINR+0.61 / +0.82+5.28 / +5.50+4.81 ± 0.22 / +4.97 ± 0.25
LaBraMTD-BRAINA+0.48 / +0.73+4.22 / +4.44+4.22 ± 0.21 / +4.43 ± 0.19
BrainOmniTUSZR+0.37 / +0.38+1.88 / +1.77+1.52 ± 0.04 / +1.43 ± 0.03
BrainOmniADFTDA+0.32 / +0.31-0.49 / -0.47-0.42 ± 0.06 / -0.38 ± 0.06
BrainOmniTD-BRAINR+0.27 / +0.18+2.37 / +2.05+2.45 ± 0.10 / +2.11 ± 0.11
LaBraMCAUEEGR-0.13 / -0.21+0.57 / +0.44+0.91 ± 0.13 / +0.71 ± 0.16
AuthorsOpenReview ↗

Response to Reviewer KZUz (2/5), Q2, W1, Q3 and W3

Q2, W1. How does a descriptor only classifier perform on each task? (Table 20) We fitted the same logistic classifier as the reference on one descriptor family alone. In Macro-F1, the best single family reaches 56.50 on TD-BRAIN (relational), 49.12 on ADFTD (smoothed spectral), 52.23 on CAUEEG (morphology) and 41.75 on TUSZ (relational). A single family outperforms the frozen reference in 10 of the 16 settings, including the smoothed spectral family on ADFTD for LaBraM, CBraMod and CSBrain. In these settings the frozen output gives the classifier less task information than the handcrafted features. The backbones still hold more than the descriptors. POST features together with one family outperform that family alone in 54 of 64 combinations, and with LoRA in all blocks LaBraM and CBraMod reach 51.1 and 53.2 on ADFTD.

Table 20. Classification from one descriptor family alone, with the reference of each backbone from Table 1. Each entry gives pooled-window Macro-F1 / balanced accuracy (%).

Classifier inputTD-BRAINADFTDCAUEEGTUSZ
P alone38.07 / 41.5631.34 / 36.8332.84 / 39.4724.41 / 33.10
A alone46.33 / 47.3249.12 / 49.2245.25 / 48.2630.12 / 35.65
R alone56.50 / 55.2541.91 / 42.2342.86 / 44.6641.75 / 41.74
M alone54.07 / 52.7641.13 / 44.2452.23 / 53.1339.24 / 40.29
LaBraM reference63.54 / 62.2045.43 / 46.4451.50 / 52.6744.58 / 44.57
CBraMod reference60.00 / 58.6544.05 / 44.7749.69 / 51.1342.43 / 42.92
CSBrain reference52.82 / 52.7548.79 / 48.7143.97 / 44.1137.42 / 37.50
BrainOmni reference55.80 / 55.7052.68 / 52.5650.15 / 50.8838.90 / 38.93

Q3, W3. Comparison with standard updates at matched and unmatched budgets (Tables 21 and 22). We added LoRA in one block, LoReFT, bottleneck adapters in all blocks, LoRA in all blocks and full fine-tuning. Table 21 gives the results on ADFTD for LaBraM and CBraMod. At a matched budget with a frozen classifier, LoRA in one block and LoReFT are each better than the residual on one backbone and worse on the other. At larger budgets with a trained classifier, LoRA in all blocks has 48 times as many parameters as the residual and outperforms it by 3.4 and 5.7 points on ADFTD. It matches the residual for CBraMod on CAUEEG. Bottleneck adapters in all blocks are better than the residual for LaBraM and worse for CBraMod. Against full fine-tuning (Table 22), the residual recovers 64, 82 and 83 percent of the Macro-F1 gain for LaBraM on ADFTD, CBraMod on ADFTD and LaBraM on TUSZ, 35 percent for CBraMod on CAUEEG, and none for LaBraM on CAUEEG.

The residual trains one site with the backbone and the classifier frozen, so each of its gains uses information that is already present at that site. The larger gain of LoRA in all blocks shows how much more the frozen output leaves inside the backbone.

Table 21. Comparison of the internal residual with other updates on ADFTD. Scores are pooled-window Macro-F1 (F1) and balanced accuracy (BAcc), reported as mean ± SD across three seeds (%). * marks updates that train the classifier. Full fine-tuning has 5,820,539 trainable parameters for LaBraM, where the three seeds coincide, and 4,884,203 for CBraMod.

UpdateTrainable parametersLaBraM F1LaBraM BAccCBraMod F1CBraMod BAcc
Reference045.4346.4444.0544.77
Classifier adaptation*60345.3846.4244.0644.76
Internal residual3,20047.70 ± 0.3148.01 ± 0.2447.51 ± 0.5647.67 ± 0.51
LoRA, one block3,20048.97 ± 1.4249.06 ± 1.3245.65 ± 0.4646.36 ± 0.52
LoReFT at MID3,20845.87 ± 1.9346.88 ± 1.5348.28 ± 1.9848.43 ± 1.86
Bottleneck adapters, all blocks*149,76349.43 ± 0.6949.47 ± 0.6147.06 ± 1.9747.24 ± 2.24
LoRA, all blocks*154,20351.13 ± 3.0151.24 ± 3.1853.17 ± 0.4453.12 ± 0.51
Full fine-tuning*5.8M, 4.9M49.0048.8448.25 ± 1.6548.52 ± 1.74

Table 22. Internal residual and full fine-tuning. Each entry gives pooled-window Macro-F1 / balanced accuracy (%), with the SD across three seeds for runs that vary across seeds.

ModelTaskReferenceAssigned siteFull fine-tuning
LaBraMADFTD45.43 / 46.4447.70 ± 0.31 / 48.01 ± 0.2449.00 / 48.84
LaBraMCAUEEG51.50 / 52.6751.37 ± 0.21 / 52.45 ± 0.1754.00 / 54.06
LaBraMTUSZ44.58 / 44.5745.49 ± 0.32 / 45.77 ± 0.3145.68 / 46.42
CBraModADFTD44.05 / 44.7747.51 ± 0.56 / 47.67 ± 0.5148.25 ± 1.65 / 48.52 ± 1.74
CBraModCAUEEG49.69 / 51.1352.24 ± 0.15 / 53.50 ± 0.2156.97 ± 1.79 / 56.52 ± 1.27
AuthorsOpenReview ↗

Response to Reviewer KZUz (3/5), Q4 and W5

Q4, W5. How were the assigned sites chosen, and what does a full layer sweep look like? (Tables 27 and 28) We fixed the sites of Table 1 before any adapter was trained, with a rule on the development folds that is now stated in Appendix A.7. The rule names a descriptor family when that family has the largest D consistently across folds and seeds, and it places the site before the largest drop of the ridge readout of this family along PRE, MID, FINAL and POST. The default is MID. For CSBrain on TD-BRAIN and on CAUEEG the rule returns FINAL, because the readout of the relational family falls most between FINAL and POST in these two settings.

We then trained residuals at PRE, MID and FINAL in 15 settings (Table 27). The best of the three fixed sites is an internal one in all 15, whichever site the rule had assigned. Relative to the best fixed site of each setting, validation selection loses 0.11 Macro-F1 points on average. A fixed site loses more in the settings where it is the wrong one, with largest losses of 1.61 for PRE, 1.85 for MID and 3.49 for FINAL.

For the layer sweep, we trained residuals at 13 sites of CBraMod on ADFTD, namely PRE and each of the 12 blocks (Table 28). The gains are 3.1 to 4.4 points at blocks 1 to 6 and vanish at blocks 11 and 12. Validation selection among the 13 sites reaches 47.92, between fixed MID (47.51) and the best fixed site (48.46). The gain profile is broad over the first half of the backbone, and this is where the selected depths vary across folds in Table 12.

Table 27. Residuals at PRE, MID, and FINAL and task validation selection among the three sites. Each entry gives pooled-window Macro-F1 / balanced accuracy (%), averaged across three seeds. Selection is made within each fold and seed.

ModelTaskFixed PREFixed MIDFixed FINALValidation selection
LaBraMTD-BRAIN63.78 / 62.5164.03 / 62.9363.42 / 62.1664.10 / 62.90
LaBraMADFTD47.86 / 48.3447.70 / 48.0146.09 / 46.8947.70 / 48.01
LaBraMCAUEEG51.09 / 52.0251.37 / 52.4551.33 / 52.3551.37 / 52.45
LaBraMTUSZ44.34 / 44.4545.49 / 45.7745.06 / 45.1545.57 / 45.83
CBraModTD-BRAIN62.46 / 61.6260.61 / 59.4760.01 / 58.7062.37 / 61.50
CBraModADFTD45.90 / 46.6047.51 / 47.6744.02 / 44.7446.93 / 47.34
CBraModCAUEEG51.98 / 53.1952.24 / 53.5050.11 / 51.4051.86 / 53.18
CBraModTUSZ42.64 / 42.7943.22 / 43.4842.69 / 43.1442.57 / 42.71
CSBrainTD-BRAIN53.35 / 53.1852.92 / 52.8452.85 / 52.7753.35 / 53.18
CSBrainADFTD48.97 / 48.8948.93 / 48.8548.73 / 48.6548.96 / 48.88
CSBrainCAUEEG44.04 / 44.2043.93 / 44.0744.01 / 44.1744.02 / 44.17
BrainOmniTD-BRAIN56.03 / 55.7856.06 / 55.8856.06 / 55.8056.08 / 55.85
BrainOmniADFTD52.76 / 52.6453.00 / 52.8752.73 / 52.6452.96 / 52.84
BrainOmniCAUEEG50.20 / 50.8450.88 / 51.5850.35 / 50.9150.88 / 51.58
BrainOmniTUSZ39.34 / 39.3439.27 / 39.3139.03 / 39.0539.48 / 39.48

Table 28. Residuals at 13 sites of CBraMod on ADFTD. Scores are pooled-window Macro-F1 (F1) and balanced accuracy (BAcc), reported as mean ± SD across three seeds (%). Gains are relative to the reference. Task validation selection chooses among the 13 sites within each fold and seed.

SiteF1BAccF1 gainBAcc gain
PRE45.90 ± 0.2446.60 ± 0.21+1.85+1.83
Block 147.18 ± 0.7047.50 ± 0.64+3.13+2.73
Block 247.24 ± 0.3247.56 ± 0.29+3.18+2.79
Block 347.84 ± 0.8248.07 ± 0.78+3.79+3.30
Block 447.55 ± 0.2047.84 ± 0.15+3.49+3.08
Block 548.46 ± 0.9848.73 ± 0.81+4.40+3.96
MID (block 6)47.51 ± 0.5647.67 ± 0.51+3.45+2.90
Block 746.27 ± 0.5646.62 ± 0.46+2.22+1.86
Block 846.79 ± 0.4747.16 ± 0.44+2.74+2.39
Block 946.14 ± 0.5046.51 ± 0.57+2.09+1.74
Block 1045.20 ± 0.4245.65 ± 0.39+1.14+0.88
Block 1143.99 ± 0.0944.65 ± 0.11-0.06-0.12
FINAL (block 12)44.02 ± 0.0344.74 ± 0.03-0.04-0.03
Validation selection47.92 ± 0.7648.14 ± 0.64+3.87+3.37

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.