acceptodds
Under review as a conference paper at ICLR 2027

SemLoop: Closing the Loop Between Semantic Relevance and EEG Foundation Model Utility

Abstract

EEG foundation models (EEG FMs) learn transferable representations from heterogeneous neural data, yet adapting them to task-specific patterns remains challenging. Language offers a natural interface for incorporating task knowledge, but task-relevant semantics can vary substantially in their utility to the target model. Existing approaches largely rely on predefined semantics, leaving a gap between semantic relevance and model utility. We propose SemLoop, a closed-loop semantic adaptation framework that bridges this gap through model feedback. SemLoop comprises three mechanisms: EEG-Grounded Semantic Alignment (EGSA) grounds language-derived neurophysiological concepts in EEG representations; Model-Validated Semantic Refinement (MVSR) uses LLM-based evaluator and planner agents to iteratively refine semantics with target-model feedback; and Experience-Internalized Semantic Planning (EISP) internalizes successful refinements for subsequent planning. Across nine EEG benchmarks and multiple EEG FMs, SemLoop consistently outperforms full fine-tuning and predefined semantic-guided approaches. Further analyses reveal heterogeneous utility among task-relevant semantics and show that model-validated refinement shifts them toward higher model utility. The code is available at https://anonymous.4open.science/r/SemLoop-75F0.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG Foundation Models

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

This paper proposes SemLoop, a closed-loop semantic adaptation framework for EEG foundation models that aims to bridge the gap between task-relevant semantic knowledge and its actual utility to downstream models. SemLoop consists of three components: (1) EEG-Grounded Semantic Alignment (EGSA), which aligns language-derived neurophysiological descriptions with EEG representations through contrastive learning and semantic interaction; (2) Model-Validated Semantic Refinement (MVSR), which employs LLM-based evaluator and planner agents to iteratively revise semantic descriptors based on descriptor-level utility and validation performance; and (3) Experience-Internalized Semantic Planning (EISP), which fine-tunes the planner using successful refinement experiences. The framework is evaluated on nine EEG benchmarks spanning six task categories, with additional experiments involving three EEG foundation model backbones. The authors report improvements over full fine-tuning and predefined semantic-guided approaches, together with ablation studies, refinement analyses, and cross-task experience-transfer experiments.

Critical strengths
  1. An interesting formulation of model-validated semantic adaptation for EEG foundation models. The paper addresses a meaningful limitation of existing semantic-guided EEG adaptation methods: task-relevant semantic descriptions do not necessarily provide useful information to a particular target model. Using descriptor-level utility and validation feedback to refine semantic guidance offers a practical approach to adapting task knowledge to model-specific requirements. The integration of EEG–text alignment with iterative semantic refinement provides a coherent framework for investigating this problem.
  2. Broad experimental coverage and reasonably comprehensive component analyses. The evaluation covers nine EEG datasets across six task categories and examines the method with multiple foundation model backbones. The authors also compare several refinement strategies, including unguided refinement, descriptor-utility feedback, performance feedback, and their combination. Additional experiments on experience internalization and cross-task transfer offer useful evidence regarding the behavior of the proposed framework. The generally positive results across diverse tasks suggest that the approach has practical potential.
Critical weaknesses

1.Limited methodological novelty and insufficient evidence for the contribution of EEG-specific semantic reasoning.

The methodological novelty of SemLoop appears relatively limited, as its three main components primarily integrate established techniques rather than introduce fundamentally new learning mechanisms. EGSA combines conventional EEG–text contrastive alignment with attention-based semantic fusion, MVSR employs LLM-based evaluator and planner agents for iterative optimization using model feedback, and EISP applies gain-weighted supervised fine-tuning to successful refinement trajectories. Although integrating these components into a closed-loop EEG adaptation framework is potentially useful, its distinction from existing feedback-driven LLM optimization paradigms remains insufficiently established.

More importantly, the experiments do not adequately isolate the benefits of EEG-specific semantic reasoning from those of generic validation-guided optimization. Although the authors include unguided refinement, utility filtering, and utility ranking baselines, stronger comparisons with budget-matched non-LLM search methods and semantically unrelated or randomly constructed descriptor banks are still missing. Consequently, it remains unclear to what extent the reported gains arise from meaningful neurophysiological knowledge rather than additional optimization over candidate semantic configurations.

Furthermore, the semantic grounding procedure is insufficiently specified. Section 3.2 states that EEG samples are associated with neurophysiological descriptions based on their characteristics. Although Appendix H provides examples of initial and refined descriptors, the paper does not clearly explain how sample-level positive EEG–text pairs are constructed or whether class labels, explicitly computed EEG features, or other supervision are involved. Since these associations directly determine the contrastive alignment objective, this omission affects the validity and reproducibility of the proposed semantic grounding mechanism.

2.Insufficient statistical evidence and robustness analysis for validation-driven semantic refinement.

Section 4.2 reports a two-sided Wilcoxon signed-rank test with p < 0.001 on paired results across nine datasets. However, if each dataset contributes one independent paired observation, the minimum attainable exact two-sided p-value is 0.00390625, even when all nine differences have the same sign. The reported significance level therefore requires clarification regarding the actual statistical units, aggregation procedure, and treatment of different metrics and random seeds.

In addition, SemLoop repeatedly evaluates candidate semantic configurations using validation feedback, selects the best-performing candidates across refinement rounds, and constructs successful experiences according to validation improvements. Although validation-based optimization is legitimate and the paper reports separate test sets, repeated adaptive selection may lead to overfitting to a particular validation partition, especially for datasets with relatively few validation subjects. The robustness of the resulting semantic configurations across alternative subject-level validation partitions is not adequately established.

Finally, the incremental improvements attributable to EISP are modest on several benchmarks. For example, Table 3 shows small gains over Full MVSR on BCIC-2A and FACED, while Table 5 reports only minor improvements over the Base Planner on some cross-task comparisons. Although these results are encouraging, the paper does not provide sufficient paired statistical evidence to establish the reliability of these incremental gains. This weakens the evidence that experience internalization consistently provides a substantial benefit beyond the underlying refinement procedure.

Questions for authors

Q1. How are sample-level EEG–semantic correspondences constructed for EGSA? Section 3.2 states that each EEG sample is associated with EEG-observable neurophysiological descriptions based on its characteristics. Could the authors specify the exact procedure for constructing these positive EEG–text pairs? In particular, are the descriptions derived from explicitly computed EEG features, class labels, expert annotations, or another mechanism? Although Appendix H provides examples of descriptor sets, it does not clarify the sample-level association procedure. Could the authors also confirm how training, validation, and test data are separated during semantic alignment?

Q2. Could the authors clarify the statistical testing procedure used to obtain p < 0.001? Section 4.2 states that a two-sided Wilcoxon signed-rank test was conducted on paired results across nine datasets. If each dataset contributes one independent paired observation, an exact two-sided p-value below 0.001 is not attainable. Could the authors specify the actual number of paired observations, how different metrics and random seeds were aggregated, which baselines were compared, and whether corrections for multiple comparisons were applied? Providing the individual paired differences and corresponding p-values would help verify the reported significance claims.

Q3. How is validation-set overfitting controlled during iterative semantic refinement? SemLoop repeatedly evaluates and selects semantic configurations using validation feedback. Could the authors clarify whether descriptor-level utility estimation, candidate selection, and successful experience construction all rely on the same validation partition? More importantly, do semantic configurations selected using one validation split retain their advantages under alternative subject-level splits or independently constructed validation partitions? Such evidence would help establish the robustness of the proposed feedback-driven refinement procedure.

Q4. What is the isolated contribution of LLM-based semantic reasoning and experience internalization? Although the authors evaluate utility filtering, utility ranking, and several refinement variants, could they provide stronger comparisons against budget-matched non-LLM search strategies and size-matched random or semantically unrelated descriptor banks? For EISP, the existing ablation and leave-one-task-out results are encouraging, but several improvements are small. Could the authors further isolate the contribution of gain-weighted planner fine-tuning, for example by comparing it with unweighted experience fine-tuning, and provide stronger paired statistical evidence for its benefits on held-out tasks?

7 comments
AuthorsOpenReview ↗

Response to Reviewer dvDp (1/5)

We sincerely thank the reviewer for their helpful suggestions.

W1: Methodological Novelty

Response:

Methodological novelty. Our contribution lies not in introducing entirely new individual components, but in establishing a new paradigm for semantic adaptation of EEG foundation models. Existing semantic-guided approaches primarily follow a one-way knowledge injection paradigm, incorporating predefined task-relevant semantics into EEG models without explicitly validating whether these semantics are useful to the target model. As illustrated in Fig. 1, this leaves a gap between semantic relevance and model utility. SemLoop addresses this gap through a closed-loop Ground–Validate–Refine–Internalize paradigm, in which model feedback is used to evaluate and iteratively refine task-relevant semantic knowledge, while successful refinement experiences are internalized for subsequent planning. This offers a new perspective on semantic-guided EEG foundation model adaptation.

Q1 & W1: Sample-level Descriptions

Response:

Training-time construction of SiS_i. For each training EEG sample XiX_i, we construct its associated semantic set SiS_i using features extracted directly from the EEG signal, without using class labels or external annotations. Specifically, we extract 73 EEG features, including band power, spatial asymmetry, and spectral entropy. Each feature is standardized using its mean and standard deviation computed exclusively from the training set. For each sample, we retain features with absolute z-scores of at least 0.35 and select up to eight with the largest absolute z-scores. If no feature meets this threshold, we retain the highest-ranked feature. Predefined feature-to-text templates convert the selected features into EEG-observable descriptions using the format [increased/decreased] [EEG feature], forming SiS_i. Each sample XiX_i is paired with all descriptions in its associated SiS_i, yielding positive EEG–text pairs for training the Semantic Aligner gϕg_\phi with a bidirectional multi-positive contrastive objective. Importantly, class labels are not used in feature selection, standardization, descriptor generation, or positive-pair construction.

Test-time usage. During training, the sample-level semantic sets SiS_i serve as auxiliary supervision for learning the alignment between EEG representations and EEG-derived neurophysiological descriptions. Once this alignment is learned, constructing sample-level semantic sets is no longer necessary at test time. Instead, the trained Semantic Aligner maps the task-level descriptor set TT, which is shared across samples within a target task, into a grounded semantic bank BB for downstream prediction. Importantly, TT contains no ground-truth class labels or label-derived information, and no test-sample labels are required throughout this process.

For completeness, the training protocol is summarized in our response to the Area Chair's Q4, clarifying the parameter updates across different stages of SemLoop.

AuthorsOpenReview ↗

Response to Reviewer dvDp (2/5)

Q2 & W2: Statistical Testing

Response:

Overall and Per-dataset Statistical Testing. In our original analysis, we treated the nine datasets and three evaluation metrics as 27 paired observations when comparing SemLoop with the best-performing baseline (underlined in Table 1) on each dataset. We acknowledge that this aggregation does not account for the dependence among metrics computed on the same dataset. In the revised analysis, we increase the number of random seeds from five to ten and conduct statistical comparisons against the baselines. Specifically, we use Balanced Accuracy as the primary metric. For each dataset, we conduct a two-sided exact Wilcoxon signed-rank test on the ten paired results obtained from the ten random seeds. For overall comparisons, we average the results over ten seeds for each dataset and perform two-sided exact Wilcoxon signed-rank tests on the nine paired dataset-level differences. Holm–Bonferroni correction is applied jointly to the nine per-dataset tests and the overall test within each baseline comparison. As shown in Table R4, the results demonstrate statistically significant improvements over most baselines on most individual datasets, as well as significant overall improvements across the nine datasets after correction.

Table R4:Per-dataset and overall statistical comparison of SemLoop against the representative baselines across nine EEG datasets.

DatasetEEG-CLIPUSEABLPMCBraMod
Gain (Bal.Acc.)pAdj. pGain (Bal.Acc.)pAdj. pGain (Bal.Acc.)pAdj. pGain (Bal.Acc.)pAdj. p
BCIC-2A+5.160.0019530.019531+2.350.0019530.019531+2.630.0039060.031250+2.700.0019530.019531
BCIC-2B+1.680.0019530.019531+2.010.0039060.019531+1.450.0058590.035156+1.190.0058590.031250
SHU-MI+1.560.0019530.019531+1.770.0195310.039062+1.170.0097660.039062+1.530.0039060.031250
FACED+4.590.0019530.019531+4.820.0019530.019531+1.430.0019530.019531+0.970.0019530.019531
SEED-V+0.730.0019530.019531+0.930.0273440.039062+0.790.0195310.039062+0.770.0273440.031250
Mumtaz2016+1.240.0019530.019531+0.650.0019530.019531+0.570.0136720.039062+0.610.0097660.031250
MentalArithmetic+5.200.0097660.019531+7.540.0019530.019531+6.010.0058590.035156+5.960.0097660.031250
BCIC2020-3+3.700.0019530.019531+2.960.0117190.035156+1.040.0117190.039062+0.810.0039060.031250
TUEV+2.420.0039060.019531+12.160.0019530.019531+2.520.0019530.019531+2.390.0039060.031250
Overall+2.920.0039060.019531+3.910.0039060.019531+1.960.0039060.031250+1.880.0039060.031250
DatasetEEGPTBrainOmniLaBraM
Gain (Bal.Acc.)pAdj. pGain (Bal.Acc.)pAdj. pGain (Bal.Acc.)pAdj. p
BCIC-2A+13.450.0019530.019531+10.750.0019530.019531+4.410.0039060.019531
BCIC-2B+4.160.0019530.019531+5.000.0019530.019531+3.070.0019530.019531
SHU-MI+9.400.0019530.019531+1.690.0488280.048828+1.250.0019530.019531
FACED+23.520.0019530.019531+35.040.0019530.019531+2.240.0019530.019531
SEED-V+15.580.0019530.019531+15.390.0019530.019531+0.490.0019530.019531
Mumtaz2016-0.370.6953120.695312+8.120.0019530.019531-0.180.6796880.679688
MentalArithmetic+12.600.0019530.019531+4.230.0097660.019531+2.820.0019530.019531
BCIC2020-3+11.160.0019530.019531+5.780.0019530.019531+7.200.0019530.019531
TUEV+1.560.0019530.019531+33.320.0019530.019531+3.840.0019530.019531
Overall+10.120.0078120.019531+13.260.0039060.019531+2.790.0078120.019531
AuthorsOpenReview ↗

Response to Reviewer dvDp (3/5)

Q3 & W2: Validation-set Overfitting

Response:

Alternative Subject-level Validation Splits. To assess the sensitivity of validation-driven refinement to validation-subject composition, we divide the original validation set into three subject-disjoint subsets and independently perform MVSR using each subset for feedback and candidate selection. The methods are evaluated on the same held-out test set. As shown in Table R3, we compare the performance across the three alternative validation splits with that obtained using the original validation set. The results show limited performance variations across splits for MVSR, demonstrating the robustness of validation-driven semantic refinement to different validation-subject compositions.

Table R3. Robustness of MVSR to alternative subject-level validation splits.

MethodValidation SettingSHU-MIFACEDTUEV
Bal.Acc.AUROCBal.Acc.KappaBal.Acc.Kappa
MVSROriginal61.23 ± 2.0567.24 ± 2.3457.79 ± 0.7252.14 ± 0.8261.98 ± 1.9660.25 ± 2.68
Split 161.14 ± 1.9167.19 ± 2.1657.68 ± 0.6751.98 ± 0.7962.12 ± 1.8760.31 ± 2.53
Split 261.27 ± 2.0267.26 ± 2.3157.85 ± 0.7652.19 ± 0.8461.84 ± 1.9860.07 ± 2.72
Split 361.21 ± 1.9867.23 ± 2.2957.74 ± 0.6252.12 ± 0.7162.01 ± 1.9460.29 ± 2.61

Q4 & W1 & W2: LLM-based Semantic Reasoning and Experience Internalization (1/3)

Response:

We thank the reviewer for this insightful suggestion.

1. Contribution of LLM-based semantic reasoning.

(1) Budget-matched Non-LLM Search Comparison. We compare MVSR with Evolutionary Search under the same search budget. Evolutionary Search generates candidates through descriptor mutation and selects them based on validation performance, without LLM-based semantic reasoning. As shown in Table R1, MVSR achieves better performance, demonstrating the benefit of LLM-based semantic reasoning beyond conventional search.

Table R1. Performance comparison of MVSR with budget-matched non-LLM search.

MethodSHU-MIFACEDTUEV
Bal.Acc.AUROCBal.Acc.KappaBal.Acc.Kappa
Evolutionary Search60.57 ± 1.5066.38 ± 1.9356.84 ± 0.6551.19 ± 0.7461.37 ± 1.8559.35 ± 1.43
MVSR61.23 ± 2.0567.24 ± 2.3457.79 ± 0.7252.14 ± 0.8261.98 ± 1.9660.25 ± 2.68

(2) Size-matched Random/Unrelated Descriptor Banks. We compare two additional baselines: Random EEG Search and Random Unrelated Search. Random EEG Search iteratively replaces descriptors with randomly sampled EEG-related descriptors, whereas Random Unrelated Search iteratively replaces descriptors with randomly sampled task-unrelated semantic descriptors from a predefined pool. Both baselines follow the same protocol as MVSR, with matched descriptor-bank sizes. As shown in Table R2, MVSR achieves the best performance, followed by Random EEG Search and Random Unrelated Search, demonstrating the importance of EEG-related semantic knowledge and task-aware semantic reasoning.

Table R2. Performance comparison of MVSR with size-matched random EEG-related and task-unrelated descriptor search.

MethodSHU-MIFACEDTUEV
Bal.Acc.AUROCBal.Acc.KappaBal.Acc.Kappa
Random Unrelated Search60.35 ± 1.6465.89 ± 1.8156.62 ± 0.7650.93 ± 0.8061.05 ± 2.1258.92 ± 1.24
Random EEG Search60.53 ± 1.5966.10 ± 1.6856.75 ± 0.6951.08 ± 0.8461.23 ± 1.9359.14 ± 1.36
MVSR61.23 ± 2.0567.24 ± 2.3457.79 ± 0.7252.14 ± 0.8261.98 ± 1.9660.25 ± 2.68
Reviewer_xZxcOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SemLoop, a closed-loop semantic adaptation framework for EEG foundation models. The key idea is that task-relevant semantics are not necessarily useful to a target model. SemLoop therefore grounds language-derived neurophysiological semantics in EEG representations, evaluates their utility using model feedback, iteratively refines them with LLM-based evaluator/planner agents, and internalizes successful refinement experiences for future planning. The method is evaluated on nine datasets and multiple EEG models. Paper reports improvements over standard full fine-tuning and several semantic-guided baselines, and provides ablations for semantic grounding, interaction, model feedback, and experience internalization.

Critical strengths
  1. The paper presents a clear and well-motivated distinction between task-relevant and model-useful semantics.

  2. The experimental evaluation is broad, covering nine datasets and multiple EEG foundation models.

  3. The ablation study is relatively comprehensive and helps isolate the contributions of semantic grounding, model feedback, and experience internalization.

Critical weaknesses
  1. The semantic refinement procedure has a substantially larger validation-search budget than the fixed-semantic baselines. MVSR generates multiple candidate semantic revisions per round, evaluates them on the target model, selects the best candidate based on validation performance, and repeats this process for 10 refinement rounds. This creates a potential confound that part of the improvement may come from repeated validation-guided search rather than specifically from model-aware semantic reasoning.

  2. Model utility should not be interpreted as neurophysiological validity. The proposed descriptor utility measures whether a textual semantic representation improves prediction for a specific model and dataset. It does not establish that the corresponding neurophysiological claim is biologically valid or actually reflected in the EEG signal. The qualitative observation that refined semantics is interesting, but stronger physiological validation would be required before interpreting these descriptors as improved neuroscientific explanations.

  3. The gains on many individual benchmarks are relatively modest compared with the additional complexity introduced by SemLoop. Although the paper reports positive aggregate improvements, several individual gains are relatively small. For example, on some BCIC-2B metrics, the improvement over the strongest baseline is very small. Given the added LLM-based evaluator/planner, repeated semantic refinement, and validation-time model evaluations, the practical benefit is not yet fully convincing.

  4. Validation feedback drives the refinement loop and final selection, raising a risk of overfitting to the validation set with repeated agent proposals; no safeguard (e.g., nested validation, small held-out tuning split) or sensitivity analysis is reported.

Questions for authors
  1. The paper reports p<0.001p<0.001 using a Wilcoxon signed-rank test across datasets. What exactly is the statistical unit used in this test, and are the improvements over the strongest baseline significant on individual datasets?

  2. Since SemLoop repeatedly evaluates candidate semantic sets using validation performance, how much of the gain is due to the proposed semantic reasoning rather than the larger validation-search budget?

  3. What is the computational overhead of SemLoop compared with standard downstream fine-tuning?

  4. It is unclear whether the EEG backbone is frozen throughout downstream adaptation or fine-tuned during EGSA/MVSR; different sections refer to “frozen” features but comparisons include “full fine-tuning” baselines, creating ambiguity on fairness.

6 comments
AuthorsOpenReview ↗

Response to Reviewer xZxc (1/4)

We sincerely thank the reviewer for their helpful suggestions.

Q1 & W3:

Response:

Overall and Per-dataset Statistical Testing. In our original analysis, we treated the nine datasets and three evaluation metrics as 27 paired observations when comparing SemLoop with the best-performing baseline (underlined in Table 1) on each dataset. We acknowledge that this aggregation does not account for the dependence among metrics computed on the same dataset. In the revised analysis, we increase the number of random seeds from five to ten and conduct statistical comparisons against the baselines. Specifically, we use Balanced Accuracy as the primary metric. For each dataset, we conduct a two-sided exact Wilcoxon signed-rank test on the ten paired results obtained from the ten random seeds. For overall comparisons, we average the results over ten seeds for each dataset and perform two-sided exact Wilcoxon signed-rank tests on the nine paired dataset-level differences. Holm–Bonferroni correction is applied jointly to the nine per-dataset tests and the overall test within each baseline comparison. As shown in Table R4, the results demonstrate statistically significant improvements over most baselines on most individual datasets, as well as significant overall improvements across the nine datasets after correction.

Table R4:Per-dataset and overall statistical comparison of SemLoop against the representative baselines across nine EEG datasets.

DatasetEEG-CLIPUSEABLPMCBraMod
Gain (Bal.Acc.)pAdj. pGain (Bal.Acc.)pAdj. pGain (Bal.Acc.)pAdj. pGain (Bal.Acc.)pAdj. p
BCIC-2A+5.160.0019530.019531+2.350.0019530.019531+2.630.0039060.031250+2.700.0019530.019531
BCIC-2B+1.680.0019530.019531+2.010.0039060.019531+1.450.0058590.035156+1.190.0058590.031250
SHU-MI+1.560.0019530.019531+1.770.0195310.039062+1.170.0097660.039062+1.530.0039060.031250
FACED+4.590.0019530.019531+4.820.0019530.019531+1.430.0019530.019531+0.970.0019530.019531
SEED-V+0.730.0019530.019531+0.930.0273440.039062+0.790.0195310.039062+0.770.0273440.031250
Mumtaz2016+1.240.0019530.019531+0.650.0019530.019531+0.570.0136720.039062+0.610.0097660.031250
MentalArithmetic+5.200.0097660.019531+7.540.0019530.019531+6.010.0058590.035156+5.960.0097660.031250
BCIC2020-3+3.700.0019530.019531+2.960.0117190.035156+1.040.0117190.039062+0.810.0039060.031250
TUEV+2.420.0039060.019531+12.160.0019530.019531+2.520.0019530.019531+2.390.0039060.031250
Overall+2.920.0039060.019531+3.910.0039060.019531+1.960.0039060.031250+1.880.0039060.031250
DatasetEEGPTBrainOmniLaBraM
Gain (Bal.Acc.)pAdj. pGain (Bal.Acc.)pAdj. pGain (Bal.Acc.)pAdj. p
BCIC-2A+13.450.0019530.019531+10.750.0019530.019531+4.410.0039060.019531
BCIC-2B+4.160.0019530.019531+5.000.0019530.019531+3.070.0019530.019531
SHU-MI+9.400.0019530.019531+1.690.0488280.048828+1.250.0019530.019531
FACED+23.520.0019530.019531+35.040.0019530.019531+2.240.0019530.019531
SEED-V+15.580.0019530.019531+15.390.0019530.019531+0.490.0019530.019531
Mumtaz2016-0.370.6953120.695312+8.120.0019530.019531-0.180.6796880.679688
MentalArithmetic+12.600.0019530.019531+4.230.0097660.019531+2.820.0019530.019531
BCIC2020-3+11.160.0019530.019531+5.780.0019530.019531+7.200.0019530.019531
TUEV+1.560.0019530.019531+33.320.0019530.019531+3.840.0019530.019531
Overall+10.120.0078120.019531+13.260.0039060.019531+2.790.0078120.019531
AuthorsOpenReview ↗

Response to Reviewer xZxc (2/4)

Q2/W1: Validation-search Budget

Response:

To disentangle the contribution of semantic reasoning from validation-search effects, we conduct two complementary experiments.

(1) Budget-matched Non-LLM Search Comparison. We compare MVSR with Evolutionary Search under the same search budget. Evolutionary Search generates candidates through descriptor mutation and selects them based on validation performance, without LLM-based semantic reasoning. As shown in Table R1, MVSR achieves better performance, demonstrating the benefit of LLM-based semantic reasoning beyond conventional search.

Table R1. Performance comparison of MVSR with budget-matched non-LLM search.

MethodSHU-MIFACEDTUEV
Bal.Acc.AUROCBal.Acc.KappaBal.Acc.Kappa
Evolutionary Search60.57 ± 1.5066.38 ± 1.9356.84 ± 0.6551.19 ± 0.7461.37 ± 1.8559.35 ± 1.43
MVSR61.23 ± 2.0567.24 ± 2.3457.79 ± 0.7252.14 ± 0.8261.98 ± 1.9660.25 ± 2.68

(2) Size-matched Random/Unrelated Descriptor Banks. We compare two additional baselines: Random EEG Search and Random Unrelated Search. Random EEG Search iteratively replaces descriptors with randomly sampled EEG-related descriptors, whereas Random Unrelated Search iteratively replaces descriptors with randomly sampled task-unrelated semantic descriptors from a predefined pool. Both baselines follow the same protocol as MVSR, with matched descriptor-bank sizes. As shown in Table R2, MVSR achieves the best performance, followed by Random EEG Search and Random Unrelated Search, demonstrating the importance of EEG-related semantic knowledge and task-aware semantic reasoning.

Table R2. Performance comparison of MVSR with size-matched random EEG-related and task-unrelated descriptor search.

MethodSHU-MIFACEDTUEV
Bal.Acc.AUROCBal.Acc.KappaBal.Acc.Kappa
Random Unrelated Search60.35 ± 1.6465.89 ± 1.8156.62 ± 0.7650.93 ± 0.8061.05 ± 2.1258.92 ± 1.24
Random EEG Search60.53 ± 1.5966.10 ± 1.6856.75 ± 0.6951.08 ± 0.8461.23 ± 1.9359.14 ± 1.36
MVSR61.23 ± 2.0567.24 ± 2.3457.79 ± 0.7252.14 ± 0.8261.98 ± 1.9660.25 ± 2.68
AuthorsOpenReview ↗

Response to Reviewer xZxc (3/4)

Q3 & Q4 & W3: Fairness and Computational Cost (1/2)

Response:

We thank the reviewers for raising concerns about backbone fairness and computational overhead. We clarify the training protocol and provide additional computational cost analyses.

Training protocol. SemLoop follows a four-stage pipeline: SemLoop = Full Fine-tuning + EGSA + MVSR + EISP. As summarized in Table R7, the EEG backbone is fine-tuned during initialization and remains frozen throughout all subsequent stages, including EGSA, MVSR, and EISP. During MVSR, candidate semantic banks are evaluated using fixed model weights without retraining the backbone. EISP only updates the Planner Agent through LoRA.

Computational overhead. We provide three complementary analyses. Table R8 breaks down the computational cost of individual SemLoop components, including initial full fine-tuning. Table R9 compares the total adaptation cost of SemLoop with standard full fine-tuning and static semantic baselines, using standard full fine-tuning as the reference. Table R10 compares their inference costs in terms of latency, GPU memory, and FLOPs.

Results and analysis. The computational results highlight three practical advantages of SemLoop.

(1) Controlled Adaptation Cost. As shown in Tables R8 and R9, SemLoop incurs a total adaptation cost of 2.60× and 2.35× relative to standard full fine-tuning on the two datasets, respectively. EGSA accounts for the largest share of this additional cost, reflecting the cost of learning task-specific semantic representations and interactions. Importantly, the EEG backbone is updated only during initialization and remains frozen throughout subsequent adaptation stages.

(2) Limited Inference Overhead. Neither the evaluator nor the planner is invoked during inference. Predictions are generated directly using the refined semantic bank and adapted EEG model. As shown in Table R10, SemLoop introduces only 0.51 ms and 0.26 ms of additional latency per sample over standard full fine-tuning on the two datasets, respectively.

(3) Reusable Refinement Experience. EISP internalizes successful refinement trajectories into the Planner Agent, enabling their reuse for subsequent cross-task adaptation. As demonstrated in Table 5, the internalized experience improves semantic planning on unseen tasks, highlighting the transferability of learned refinement strategies.

Together with the performance improvements reported in Table 1, these analyses demonstrate that SemLoop achieves a favorable trade-off between adaptation performance, computational cost, and inference efficiency.

Notes: The reported metrics are defined as follows: Time Share=Component TimeTotal SemLoop Time\mathrm{Time\ Share} = \frac{\mathrm{Component\ Time}}{\mathrm{Total\ SemLoop\ Time}}.

Relative Cost(×)=Method Adaptation TimeFull Fine-tuning Time\mathrm{Relative\ Cost}(\times) = \frac{\mathrm{Method\ Adaptation\ Time}}{\mathrm{Full\ Fine\text{-}tuning\ Time}}.

Additional Adaptation Overhead=Relative Adaptation Cost−1.\mathrm{Additional\ Adaptation\ Overhead} = \mathrm{Relative\ Adaptation\ Cost}-1.

Table R7. Training protocol of SemLoop. ✓ indicates that the module is trainable and updated during the corresponding stage, whereas ✗ indicates that its parameters remain frozen.

StageEEG FMClassifierSemantic AlignerSemantic Interaction ModulePlanner Agent (LoRA)
Full Fine-tuning✓✓✗✗✗
EGSA (Semantic Grounding)✗✗✓✗✗
EGSA (Semantic Interaction)✗✓✗✓✗
MVSR✗✗✗✗✗
EISP✗✗✗✗✓

Table R8:Computational cost breakdown of individual SemLoop components across two representative EEG datasets.

ComponentSEED-VTUEV
Time (min)Peak GPU (GB)Time Share (%)Time (min)Peak GPU (GB)Time Share (%)
Full Fine-tuning33.111.0838.45%42.621.6242.61%
EGSA (Semantic Grounding)24.581.2328.54%33.321.9433.31%
EGSA (Semantic Interaction)8.161.149.48%7.892.377.89%
MVSR9.771.9711.34%7.692.077.68%
EISP10.4917.0212.19%8.5117.228.51%
SemLoop Total86.1017.02100%100.0317.22100%

Table R9:Adaptation cost comparison between SemLoop, standard full fine-tuning, and static semantic baselines. Relative adaptation costs are normalized by standard full fine-tuning.

MethodSEED-VTUEV
Time (min)Relative Cost (×)Peak GPU (GB)Time (min)Relative Cost (×)Peak GPU (GB)
Standard Full Fine-tuning33.111.00×1.0842.621.00×1.62
EEG-CLIP85.672.59×3.0882.001.92×4.47
USEA83.792.53×3.0898.362.31×4.51
BLPM83.052.51×2.2192.322.17×2.74
SemLoop86.102.60×17.02100.032.35×17.22
Reviewer_AcHtOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SemLoop, a closed-loop semantic adaptation framework for EEG Foundation Models (EEG FMs) that bridges the gap between task-relevant neurophysiological semantics and actual target-model utility. SemLoop integrates three modules: (1) EGSA to align and fuse CLIP-encoded neurophysiological concepts with EEG representations via contrastive learning and gated cross-attention; (2) MVSR to iteratively refine descriptors using leave-one-out utility attribution and LLM-based Evaluator/Planner agents; and (3) EISP to internalize successful refinement trajectories into the planner via gain-weighted SFT. Experiments across 9 EEG benchmarks and 3 EEG FM backbones show consistent gains over full fine-tuning and static semantic baselines.

Critical strengths
  1. Identifying that human-defined task-relevant semantics vary substantially in model utility is insightful, and shifting from static one-way knowledge injection to closed-loop, utility-driven semantic refinement is well-motivated.

  2. Combining leave-one-out descriptor utility with LLM-based diagnosis and planning effectively discovers high-utility neurophysiological concepts, demonstrably outperforming naive utility filtering and ranking.

  3. The framework is thoroughly validated across 9 diverse EEG benchmarks and 3 backbones, with comprehensive ablations confirming each module's contribution and cross-task transferability.

Critical weaknesses
  1. How are the sample-level descriptions S_i generated for each EEG sample X_i during EGSA, and how does a frozen Semantic Aligner reliably project out-of-distribution concepts (e.g., connectivity) introduced in later MVSR rounds?

  2. During MVSR, are model weights re-trained for each candidate descriptor set, and what is the exact wall-clock runtime overhead of SemLoop compared to standard fine-tuning and static semantic baselines?

  3. The related work overlooks several important recent advances in EEG foundation models, such as CSBrain and BIOT.

Questions for authors

Please refer to the above Weaknesses.

6 comments
AuthorsOpenReview ↗

Response to Reviewer AcHt (1/3)

We sincerely thank the reviewer for their helpful suggestions.

W1: Sample-level Descriptions and Out-of-distribution Semantic Grounding

Response:

1. Sample-level description construction.

(1) Training-time construction of SiS_i. For each training EEG sample XiX_i, we construct its associated semantic set SiS_i using features extracted directly from the EEG signal, without using class labels or external annotations. Specifically, we extract 73 EEG features, including band power, spatial asymmetry, and spectral entropy. Each feature is standardized using its mean and standard deviation computed exclusively from the training set. For each sample, we retain features with absolute z-scores of at least 0.35 and select up to eight with the largest absolute z-scores. If no feature meets this threshold, we retain the highest-ranked feature. Predefined feature-to-text templates convert the selected features into EEG-observable descriptions using the format [increased/decreased] [EEG feature], forming SiS_i. Each sample XiX_i is paired with all descriptions in its associated SiS_i, yielding positive EEG–text pairs for training the Semantic Aligner gϕg_\phi with a bidirectional multi-positive contrastive objective. Importantly, class labels are not used in feature selection, standardization, descriptor generation, or positive-pair construction.

(2) Test-time usage. During training, the sample-level semantic sets SiS_i serve as auxiliary supervision for learning the alignment between EEG representations and EEG-derived neurophysiological descriptions. Once this alignment is learned, constructing sample-level semantic sets is no longer necessary at test time. Instead, the trained Semantic Aligner maps the task-level descriptor set TT, which is shared across samples within a target task, into a grounded semantic bank BB for downstream prediction. Importantly, TT contains no ground-truth class labels or label-derived information, and no test-sample labels are required throughout this process.

2. Generalization to unseen semantic expressions. The Semantic Grounding stage in EGSA consists of a frozen pretrained CLIP text encoder, a Semantic Aligner (a trainable MLP projection), and a frozen EEG backbone. The frozen CLIP text encoder retains rich semantic knowledge, while the Semantic Aligner learns to map text embeddings into the EEG feature space through EEG–text contrastive supervision. This allows unseen textual expressions to leverage CLIP's pretrained semantic knowledge and the cross-modal mapping learned during EGSA. During subsequent MVSR, all three components remain frozen.

To investigate this capability, we compare restricted and unrestricted descriptor generation under matched search budgets. The restricted setting uses only descriptions observed during EGSA training, whereas the unrestricted setting allows the LLM to generate new textual expressions. We analyze the proportion of newly generated expressions selected by SemLoop and evaluate downstream performance. As shown in Table R11, the results show that CLIP's pretrained semantic knowledge and learned cross-modal alignment enable previously unseen textual expressions to be incorporated into SemLoop, improving downstream performance.

Table R11. Performance comparison of restricted and unrestricted semantic descriptor generation, including the proportion of selected novel expressions and downstream performance.

SettingBCIC-2AFACEDMentalArithmetic
Selected Novel Expression RatioBal.Acc.KappaSelected Novel Expression RatioBal.Acc.KappaSelected Novel Expression RatioBal.Acc.AUROC
Restricted Generation0.0%48.21 ± 1.3430.92 ± 1.490.0%57.52 ± 0.6351.91 ± 0.730.0%63.18 ± 4.6772.62 ± 1.84
Unrestricted Generation (Ours)15.9%49.08 ± 1.1732.11 ± 1.569.3%57.96 ± 0.4852.34 ± 0.5710.9%65.56 ± 4.4474.51 ± 1.60
Gain—+0.87+1.19—+0.44+0.43—+2.38+1.89
AuthorsOpenReview ↗

Response to Reviewer AcHt (2/3)

W2: Computational Cost and Runtime Overhead (1/2)

Response:

We thank the reviewers for raising concerns about backbone fairness and computational overhead. We clarify the training protocol and provide additional computational cost analyses.

Training protocol. SemLoop follows a four-stage pipeline: SemLoop = Full Fine-tuning + EGSA + MVSR + EISP. As summarized in Table R7, the EEG backbone is fine-tuned during initialization and remains frozen throughout all subsequent stages, including EGSA, MVSR, and EISP. During MVSR, candidate semantic banks are evaluated using fixed model weights without retraining the backbone. EISP only updates the Planner Agent through LoRA.

Computational overhead. We provide three complementary analyses. Table R8 breaks down the computational cost of individual SemLoop components, including initial full fine-tuning. Table R9 compares the total adaptation cost of SemLoop with standard full fine-tuning and static semantic baselines, using standard full fine-tuning as the reference. Table R10 compares their inference costs in terms of latency, GPU memory, and FLOPs.

Results and analysis. The computational results highlight three practical advantages of SemLoop.

(1) Controlled Adaptation Cost. As shown in Tables R8 and R9, SemLoop incurs a total adaptation cost of 2.60× and 2.35× relative to standard full fine-tuning on the two datasets, respectively. EGSA accounts for the largest share of this additional cost, reflecting the cost of learning task-specific semantic representations and interactions. Importantly, the EEG backbone is updated only during initialization and remains frozen throughout subsequent adaptation stages.

(2) Limited Inference Overhead. Neither the evaluator nor the planner is invoked during inference. Predictions are generated directly using the refined semantic bank and adapted EEG model. As shown in Table R10, SemLoop introduces only 0.51 ms and 0.26 ms of additional latency per sample over standard full fine-tuning on the two datasets, respectively.

(3) Reusable Refinement Experience. EISP internalizes successful refinement trajectories into the Planner Agent, enabling their reuse for subsequent cross-task adaptation. As demonstrated in Table 5, the internalized experience improves semantic planning on unseen tasks, highlighting the transferability of learned refinement strategies.

Together with the performance improvements reported in Table 1, these analyses demonstrate that SemLoop achieves a favorable trade-off between adaptation performance, computational cost, and inference efficiency.

Notes: The reported metrics are defined as follows: Time Share=Component TimeTotal SemLoop Time\mathrm{Time\ Share} = \frac{\mathrm{Component\ Time}}{\mathrm{Total\ SemLoop\ Time}}.

Relative Cost(×)=Method Adaptation TimeFull Fine-tuning Time\mathrm{Relative\ Cost}(\times) = \frac{\mathrm{Method\ Adaptation\ Time}}{\mathrm{Full\ Fine\text{-}tuning\ Time}}.

Additional Adaptation Overhead=Relative Adaptation Cost−1.\mathrm{Additional\ Adaptation\ Overhead} = \mathrm{Relative\ Adaptation\ Cost}-1.

Table R7. Training protocol of SemLoop. ✓ indicates that the module is trainable and updated during the corresponding stage, whereas ✗ indicates that its parameters remain frozen.

StageEEG FMClassifierSemantic AlignerSemantic Interaction ModulePlanner Agent (LoRA)
Full Fine-tuning✓✓✗✗✗
EGSA (Semantic Grounding)✗✗✓✗✗
EGSA (Semantic Interaction)✗✓✗✓✗
MVSR✗✗✗✗✗
EISP✗✗✗✗✓

Table R8:Computational cost breakdown of individual SemLoop components across two representative EEG datasets.

ComponentSEED-VTUEV
Time (min)Peak GPU (GB)Time Share (%)Time (min)Peak GPU (GB)Time Share (%)
Full Fine-tuning33.111.0838.45%42.621.6242.61%
EGSA (Semantic Grounding)24.581.2328.54%33.321.9433.31%
EGSA (Semantic Interaction)8.161.149.48%7.892.377.89%
MVSR9.771.9711.34%7.692.077.68%
EISP10.4917.0212.19%8.5117.228.51%
SemLoop Total86.1017.02100%100.0317.22100%

Table R9:Adaptation cost comparison between SemLoop, standard full fine-tuning, and static semantic baselines. Relative adaptation costs are normalized by standard full fine-tuning.

MethodSEED-VTUEV
Time (min)Relative Cost (×)Peak GPU (GB)Time (min)Relative Cost (×)Peak GPU (GB)
Standard Full Fine-tuning33.111.00×1.0842.621.00×1.62
EEG-CLIP85.672.59×3.0882.001.92×4.47
USEA83.792.53×3.0898.362.31×4.51
BLPM83.052.51×2.2192.322.17×2.74
SemLoop86.102.60×17.02100.032.35×17.22
AuthorsOpenReview ↗

Response to Reviewer AcHt (3/3)

W2: Computational Cost and Runtime Overhead (2/2)

Response:

Table R10: Inference cost comparison between standard full fine-tuning, static semantic baselines, and SemLoop across two representative EEG datasets.

MethodSEED-VTUEV
Latency (ms/sample)Peak GPU (GB)FLOPs (G/sample)Latency (ms/sample)Peak GPU (GB)FLOPs (G/sample)
Standard Full Fine-tuning4.11 ± 0.150.0400.6304.17 ± 0.060.0960.823
EEG-CLIP4.35 ± 0.510.0400.6304.25 ± 0.340.0960.823
USEA4.23 ± 0.270.0400.6304.22 ± 0.140.0960.823
BLPM8.27 ± 0.230.0911.4928.24 ± 0.340.1481.736
SemLoop4.62 ± 0.360.0410.6494.43 ± 0.150.0970.843

W3: Related Work and Generalization across EEG Foundation Models

Response:

We thank the reviewer for pointing out these relevant works. We will expand the Related Work section to include recent EEG foundation models, particularly CSBrain and BIOT. Specifically, we discuss their advances in cross-scale spatiotemporal modeling and unified biosignal tokenization, respectively. While these works focus on improving EEG representation learning through architectural and pretraining innovations, SemLoop addresses the gap between semantic relevance and model utility in downstream EEG foundation model adaptation by grounding task-relevant semantics in EEG representations and iteratively validating and refining them through target-model feedback.

Our original experiments already evaluate SemLoop across diverse EEG foundation model backbones, including LaBraM and EEGPT, comparing it with standard full fine-tuning and semantic-guided baselines (USEA, BLPM, and EEG-CLIP). To further strengthen this evaluation, we extend our experiments to two additional backbones, BIOT and CSBrain. Following the same evaluation protocol as Table 2, we compare SemLoop with standard full fine-tuning and all three semantic-guided baselines using the same backbone for each comparison. The results are presented in Table R12 and demonstrate the applicability of SemLoop across diverse EEG foundation model architectures and its advantages over existing semantic-guided adaptation methods.

Table R12. Performance comparison across additional EEG foundation model backbones (CSBrain and BIOT).

MethodBCIC-2BSEED-VMentalArithmetic
Bal.Acc.AUROCBal.Acc.KappaBal.Acc.AUROC
BIOT (Full FT)67.26 ± 1.7976.41 ± 1.6935.97 ± 1.2019.90 ± 1.4262.29 ± 8.8471.92 ± 6.66
USEA67.84 ± 1.6676.82 ± 1.8537.03 ± 1.2620.61 ± 2.3161.74 ± 5.9671.54 ± 5.43
BLPM67.23 ± 1.4576.47 ± 1.8037.58 ± 0.9421.18 ± 1.2262.47 ± 5.5272.01 ± 4.06
EEG-CLIP67.38 ± 0.9176.63 ± 0.5336.74 ± 0.9120.87 ± 0.8362.83 ± 2.2172.68 ± 2.87
Ours69.42 ± 1.4378.09 ± 1.6138.18 ± 0.8921.67 ± 1.1465.61 ± 3.7274.72 ± 2.93
ΔIMP ↑+2.33%+1.65%+1.60%+2.31%+4.42%+2.81%
MethodBCIC-2BSEED-VMentalArithmetic
Bal.Acc.AUROCBal.Acc.KappaBal.Acc.AUROC
CSBrain (Full FT)66.18 ± 1.9472.28 ± 2.3541.62 ± 0.5827.43 ± 0.7174.93 ± 1.7884.21 ± 2.49
USEA66.74 ± 1.9272.91 ± 1.7242.16 ± 1.5928.07 ± 2.7275.14 ± 4.8984.63 ± 2.61
BLPM66.25 ± 1.3872.58 ± 1.9342.58 ± 1.0628.81 ± 1.1575.88 ± 4.4185.06 ± 3.36
EEG-CLIP66.37 ± 0.8172.73 ± 0.7642.36 ± 0.5128.61 ± 0.7376.26 ± 1.5885.31 ± 1.73
Ours68.34 ± 1.3674.06 ± 1.5843.48 ± 0.3929.36 ± 0.4779.48 ± 1.2787.97 ± 2.26
ΔIMP ↑+2.40%+1.58%+2.11%+1.91%+4.22%+3.12%

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.