acceptodds
Under review as a conference paper at ICLR 2027

BridgeRouter: Transfer-Aware Routing over Complementary Cross-Modal Bridges

Abstract

Lightweight representation bridges make pretrained, task-adapted EEG foundation models executable on other biosignal modalities without retraining their backbones. However, selecting a single bridge pathway globally assigns every instance to the same transfer route and ignores complementary errors across routes. We therefore formulate multi-bridge transfer as an instance-level algorithm-selection problem and introduce BridgeRouter, a gain-aware router over a portfolio of frozen cross-modal pathways. A validation-selected Single Best Solver (SBS) serves as a protected default route. During training, task labels are used to compute the reduction in negative log-likelihood provided by each alternative route relative to the SBS, and a lightweight MLP predicts these route-specific gains from a target-modality representation extracted by a frozen encoder. At inference, BridgeRouter observes only the target modality, requires neither EEG, task labels, nor predictions from all candidate routes, and executes exactly one bridge pathway: the route with the largest predicted gain when that gain exceeds a validation-calibrated threshold, and the SBS otherwise. We evaluate seven EEG foundation-model bridge pathways for EMG-based freezing-of-gait detection on FOG and ECG-based sleep classification on ISRUC. Across five seeds, BridgeRouter achieves balanced accuracy on FOG, improving over the BIOT SBS by percentage points, and on ISRUC, improving over the CBraMod SBS by points. Per-instance bridge oracles reach and , respectively, revealing substantial remaining portfolio headroom. These results show that anchor-relative gain prediction can convert complementary cross-modal bridge behavior into deployable instance-level routing improvements while retaining single-route inference.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
cross-modal knowledge transfer, model routing, algorithm selection, biosignal foundation models, multimodal representation learning, EEG

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_sCt2OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

BridgeRouter presents a clear and computationally attractive instance-level routing framework over cross-modal bridge portfolios. The paper is well written and provides strong diagnostics, with substantial gains over the validation-selected SBS on FOG and smaller gains on ISRUC. However, the practical motivation remains unclear because supervised routing requires target labels, while a simpler target-only MLP outperforms BridgeRouter on FOG. Concerns about gain supervision, cross-subject generalization, and collapsed routes further weaken the evidence. Overall, the method is promising but not yet sufficiently validated for acceptance.

Critical strengths
  1. The routing formulation is simple and well motivated at the algorithmic level.

2.The FOG improvement over the predefined SBS is substantial.

3.The paper is unusually transparent about negative and ambiguous results.

Critical weaknesses
  1. Practical motivation limited. BridgeRouter requires target-task labels for gain supervision, weakening the original label-scarce cross-modal transfer motivation. Once such labels are available, direct target-side supervision becomes a natural alternative; indeed, the target-only MLP outperforms BridgeRouter on FOG. The corresponding ISRUC baseline is missing.

  2. Training protocol and generalization. The same paired data are used for bridge fitting and router supervision, so routing targets reflect in-sample bridge behavior. Moreover, the strongest gain is obtained under the same-subject FOG split, while the subject-disjoint ISRUC setting shows smaller and less stable improvements.

  3. Role of collapsed routes. Several FOG routes are nearly single-class predictors, which can inflate oracle complementarity and potentially allow the supervised router to behave partly as a classifier. Although the compact-four oracle remains strong, supervised BridgeRouter performance using only non-collapsed routes is not reported.

Questions for authors

see limitations

3 comments

Weakness 1:

The central contribution of BridgeRouter is not to eliminate target-task supervision, but to investigate how that supervision can be used to adaptively select among pretrained cross-modal prediction pathways. Rather than training a new target-task predictor from scratch, the supervised Gain Router learns relative pathway performance and uses a threshold-protected decision rule to select one candidate pathway per instance.

This formulation allows us to isolate the value of instance-level selection while keeping the underlying cross-modal predictors fixed. As shown in our response to RA6J (Weakness 4), direct target-only classifiers remain strong alternatives and can outperform routing. We will distinguish these two approaches more explicitly and position BridgeRouter as a pathway-selection mechanism rather than a universally superior target-task classifier.

Weakness 2:

Our FOG experiments follow the original BioX-Bridge (ICLR 2026 Oral) sample-wise benchmark protocol to enable a controlled comparison between instance-level routing and the validation-selected fixed pathway using identical frozen predictors. This isolates the contribution of adaptive pathway selection within the established evaluation setting.

We acknowledge that the original FOG split contains overlapping windows from the same recordings across partitions. Consequently, these results demonstrate improvements under the original benchmark protocol but should not be interpreted as evidence of generalization to unseen subjects.

Importantly, ISRUC provides a complementary subject-disjoint evaluation, where the corrected router outperforms the validation-selected SBS by 3.83 percentage points on the test set, with positive gains across all five seeds. We will explicitly clarify these evaluation differences and their implications in the revised manuscript.

Weakness 3:

Thank you for identifying this missing experiment. We have now evaluated the supervised Gain Router on a compact four-pathway portfolio. Specifically, we excluded pathways whose dominant predicted class accounted for at least 95% of their predictions on the training split. This criterion uses no ground-truth labels and removes LaBraM, EEGPT, and REVE, retaining BIOT, CBraMod, CSBrain, and BENDR.

We retained the original FOG data split, training configuration, and five seeds (42-46). BIOT remains the validation-selected single best solver (SBS); router checkpoints and switching thresholds are selected exclusively on the internal development split.

MethodTest balanced accuracyImprovement over SBS
BIOT SBS69.89%—
Compact-4 majority voting¹71.58%+1.70 pp
Compact-4 Gain Router75.77 ± 0.53%+5.88 pp
Full seven-pathway Gain Router81.38 ± 1.11%+11.49 pp

¹ Ties are resolved in favor of BIOT. Router results are reported as mean ± sample standard deviation across five training seeds.

The Compact-4 router outperforms BIOT in all five seeds and exceeds majority voting by 4.19 percentage points on average. Thus, the routing improvement is not entirely attributable to near-single-class pathways.

Nevertheless, removing these pathways reduces mean balanced accuracy by 5.61 percentage points, demonstrating that they contribute substantially to the full-portfolio result. We will explicitly acknowledge this dependence and include the compact-portfolio results in the revision. We also agree that oracle performance alone is insufficient evidence of useful complementarity since two complementary constant predictors already yield a perfect oracle.

This experiment addresses the contribution of near-single-class pathways under the original same-subject FOG protocol. It does not establish cross-subject generalization or rule out the possibility that supervised routing partly learns a target-task classifier.

Thank you for the response and additional experiments. I appreciate the Compact-4 results, which provide useful evidence that the routing gains are not entirely driven by collapsed pathways.

Regarding the FOG evaluation protocol, I accept the argument that the sample-wise split follows the established BioX-Bridge benchmark for now. However, given the overlapping windows across partitions, I still encourage evaluation under a recording-disjoint or subject-disjoint split to better assess generalization.

My main concern remains the necessity and practical value of bridge-based routing when target-task supervision is already available. Since the router uses target-task labels to learn pathway-specific gains, why not directly train a lightweight supervised MLP on the same target-modality representations? Such a classifier could be integrated into the existing architecture without requiring multiple cross-modal bridges. Indeed, the target-only MLP outperforms BridgeRouter on FOG (83.40% vs. 81.38% BAcc). The authors' reframing of the contribution as supervised pathway selection does not fully explain what unique advantage the bridge-based approach provides.

The ISRUC results further raise concerns about generalization. While I agree that the SBS should be selected using validation data, BridgeRouter (62.92%) still underperforms three fixed pathways: EEGPT (65.17%), BENDR (64.45%), and BIOT (64.05%). This makes the practical benefit of instance-level routing less convincing under the subject-disjoint setting.

Finally, I appreciate the authors' acknowledgment that supervised routing may partly learn a target-task classifier. This reinforces my concern about the distinction between the proposed routing mechanism and direct supervised classification.

Overall, the rebuttal addresses some concerns, but the core motivation and practical advantages of bridge-based routing remain insufficiently justified. I maintain my original rating for now, but remain open to further clarification or evidence addressing these concerns.

We thank the reviewer for the thoughtful follow-up. We would like to clarify that our goal is not to outperform direct supervised classifiers, but to investigate a fundamental limitation of fixed-pathway cross-modal transfer in BioX-Bridge.

The target-only MLP serves as a reference rather than the primary baseline our method seeks to replace. Its strong performance is valuable, but it addresses a different question: whether direct target-task supervision can achieve higher classification accuracy. Our work instead asks whether relying on a single fixed bridge is sufficient when different foundation-model pathways exhibit complementary strengths across instances.

BioX-Bridge establishes the feasibility of cross-modal transfer, but its performance depends heavily on the underlying foundation models and bridge compatibility. Since no individual pathway is necessarily optimal for every instance, we investigate instance-level routing as a natural extension of BioX-Bridge, rather than treating fixed-pathway selection as the final solution.

We deliberately retain the original pathways to isolate the contribution of routing. Fine-tuning bridges, replacing foundation models, or optimizing classifiers could improve absolute performance, but these are separate directions beyond our present scope.

We acknowledge that the ISRUC results demonstrate limitations in routing generalization. Nevertheless, our contribution is not a claim of universal superiority, but the introduction and evaluation of instance-adaptive pathway selection within the BioX-Bridge framework. We believe this opens a meaningful direction for improving cross-modal transfer beyond fixed-pathway designs.

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

The paper presents BridgeRouter, an instance-level routing framework for cross-modal transfer from pretrained EEG foundation models. Rather than relying on a single globally selected bridge, the method learns to predict the relative task-loss gain of candidate bridges and selects an alternative route only when the predicted gain is sufficiently large. This enables instance-adaptive bridge selection while executing only one bridge pathway at inference time.

Critical strengths

The gain-aware routing method is well-motivated and practically efficient. The empirical gains are meaningful, particularly on FOG (+11.49 balanced accuracy improvement over the SBS). The the oracle/headroom analysis also shows the portfolio complementarity and how much the router recovers.

Critical weaknesses
  1. The method is just a model selection scheme over a few frozen pretrained models, without involving FM model training and providing new insigths in the cross-modal transfer field.

  2. Cross-dataset generalization seems limited, with only 3.8% improvement on ISRUC.

Questions for authors
  1. The paper should clearly define terms before use, for example 'bridge' and 'pathway'. A proforlio seems to include one pathway for each model, would multiple pathways per EEG model be possible?

  2. The Router always selects exactly one pathway. It might be worth discussing whether selecting top-K pathways could recover additional headroom, although it may increase more complexity.

  3. The model architectures of both the bridge B and the router G are not fully described. It would be better to summarize the key details and motivation of the choices in Section 3.

2 comments

Weakness 1:

BioX-Bridge provides a mechanism for constructing cross-modal prediction pathways; BridgeRouter addresses the subsequent decision of which pathway to use for a particular target input. A backbone's performance on its native EEG task does not tell us which bridged pathway will work best on a given target input. We therefore hold the backbones and bridges fixed when comparing adaptive routing with the validation-selected single best pathway, so that the measured gain is attributable to the selection policy rather than to changes in the underlying predictors. On FOG, after excluding three near-single-class pathways, the router achieved 75.77±0.53%75.77 \pm 0.53\% test balanced accuracy across five seeds, compared with 69.89%69.89\% for the validation-selected best fixed pathway in the same portfolio. In ISRUC, the corrected seven-pathway router achieved 62.92±2.17%62.92 \pm 2.17\%, compared with 59.09%59.09\% for the validation-selected CBraMod SBS. Both execute one pathway per sample. This shows that, even with the available cross-modal pathways held fixed, choosing a pathway per input can improve on committing to one pathway for the entire dataset.

Weakness 2:

Although the improvement on ISRUC is smaller than on FOG, the two benchmarks evaluate distinct downstream tasks and target modalities. On ISRUC, the corrected Gain Router achieves 62.92±2.17%62.92 \pm 2.17\% test balanced accuracy, compared with 59.09%59.09\% for the validation-selected CBraMod SBS, a gain of 3.833.83 percentage points. Importantly, the improvement is positive across all five router seeds (42--46), and both methods use identical frozen pathway predictions and the same subject-disjoint test population.

These results provide complementary evidence that instance-level routing can improve upon a fixed, validation-selected pathway across different cross-modal transfer settings. The magnitude of the gain depends on the available pathway complementarity and the reliability of gain prediction, rather than being expected to remain identical across datasets.

Question 1:

We distinguish a bridge, which performs the learned cross-modal mapping, from a pathway, which denotes the complete prediction route associated with a foundation model and its bridge. In our experiments, each EEG foundation model is paired with one bridge, resulting in a one-to-one correspondence between bridges and pathways. This is why the two terms may appear interchangeably in the experimental discussion, although they refer to different levels of the architecture.

Importantly, BridgeRouter does not require this one-to-one correspondence. Multiple bridges can be associated with the same foundation model, allowing different pathways to specialize while sharing the pretrained backbone. We will clarify this distinction in the revised manuscript.

Question 2:

We evaluated Top-K execution (K=1,2,3K=1,2,3) on FOG and ISRUC across five seeds (42--46), using frozen router checkpoints and unchanged switching thresholds. When switching is triggered, we uniformly average the class probabilities of the KK highest-ranked pathways; otherwise, the anchor prediction is used. Top-1 exactly reproduces the original router.

Dataset / PortfolioKTest BAcc (mean ± SD)Δ vs. Top-1Mean routes executed
FOG / Full181.38 ± 1.11%—1.00
FOG / Full281.48 ± 1.28%+0.10 pp1.70
FOG / Full381.96 ± 1.30%+0.58 pp2.40
FOG / Compact-4175.77 ± 0.53%—1.00
FOG / Compact-4274.53 ± 0.39%−1.24 pp1.39
FOG / Compact-4372.37 ± 0.12%−3.40 pp1.77
ISRUC / Full162.92 ± 2.17%—1.00
ISRUC / Full263.01 ± 2.19%+0.08 pp1.13
ISRUC / Full363.03 ± 2.24%+0.11 pp1.26

Overall, Top-K fusion provides no consistent advantage over Top-1 while requiring additional pathway executions. We therefore retain Top-1 as the default under a single-pathway execution budget. These results are specific to frozen routers with uniform probability averaging and do not rule out improvements from other multi-pathway fusion strategies.

Question 3:

BridgeRouter consists of two components: the frozen cross-modal prediction pathways inherited from BioX-Bridge and a lightweight gain-prediction router. The router uses target-side representations to estimate the relative task-loss improvement of alternative pathways over a validation-selected anchor. A threshold-based decision rule then selects either the anchor or the predicted best alternative, executing only one pathway per instance.

The routing mechanism is shared across FOG and ISRUC, while the target-side representations and input dimensions follow their respective modality encoders. The router is implemented as a small MLP rather than an additional foundation model. We will clarify the architecture and training configuration in the revised manuscript to make the distinction between the existing bridges and our routing contribution explicit.

Thank you for the response and the additional experiments. However, the primarily empirical treatment, together with the lack of new insights into learning and cross-modal transfer, leaves my score unchanged.

Reviewer_RA6JOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes BridgeRouter, a framework for selecting cross-modal pathways at the instance level. Building upon the BioX-Bridge paper, which proposed a pathway to align representations from two different modalities, this paper argues that different pathways can fail on different instances; a pathway that is weaker overall can still perform better on specific instances. Therefore, they propose a router to dynamically decide which pathway should be chosen for a given instance. A threshold is also introduced to determine whether to select a new pathway or default to the anchor pathway. The experimental results show that BridgeRouter performs better than the Single Best Solver model on both the EEG-to-EMG and EEG-to-ECG tasks.

Critical strengths
  • The study points out that a globally selected pathway, even the Single Best Solver (SBS), may fail to generalize across unseen subjects. BridgeRouter leverages the complementary errors among different pathways to exploit the substantial headroom that a single model cannot achieve.
  • BridgeRouter performs substantially better than the SBS model.
  • The paper is well organized and easy to follow.
Critical weaknesses
  • Relying on multiple different EEG models seems both incremental and practically inefficient for real-life scenarios. It is not entirely clear why such a large portfolio of models is required, or if some could simply be dropped without performance loss. Ultimately, if a single model can generalize effectively across all tasks, the proposed routing framework becomes unnecessary. On the other hand, if the initial portfolio of pathways is poor or lacks complementarity, the routing process will not yield significant value.
  • The system depends on two crucial components: the default pathway and the protection threshold. Both of these components are selected exclusively based on the validation dataset. This implies that if the validation set is noisy or fails to accurately reflect the true distribution of unseen subjects, the system might select an incorrect SBS model or set an ineffective threshold, leading to inaccurate routing.
  • The experiments are conducted on only one EMG-EEG and one ECG-EEG dataset. This limited scope may introduce dataset bias, making the advantages of the method less convincing.
  • In Table 10 (Appendix A.6), a simple Target-only MLP outperforms BridgeRouter. This crucial finding undermines the central claim that cross-modal transfer from EEG foundation models is necessary or beneficial for target modality classification.
Questions for authors
  • Although BridgeRouter demonstrates improvement over the SBS, the results show that there is still substantial headroom compared to the ideal performance of the Oracle. Do the authors have any specific proposals or future directions for the Router's architecture to narrow this gap?
  • If we develop a foundation model in the future that is powerful enough and generalizes well across almost all data instances, would this complex routing framework with multiple pathways still be necessary?
  • It is somewhat out of the scope of the paper, but which approach do the authors think makes more sense: applying different bridges to the same model, or applying one bridge for each model?
3 comments

Weakness 1:

BridgeRouter is not tied to a fixed portfolio size. To examine pathway redundancy, we evaluated nested subsets selected using validation router performance and fixed before test evaluation. On the original FOG split, the three-pathway router achieved 82.35±0.69%82.35 \pm 0.69\% test balanced accuracy across five matched seeds, compared with 81.03±0.79%81.03 \pm 0.79\% for the full seven-pathway router and 69.89%69.89\% for the validation-selected BIOT baseline.

Because the three-pathway subset includes two near-single-class pathways, we additionally evaluated a Compact-4 portfolio excluding such pathways. As detailed in our response to sCt2 (Weakness 3), the Compact-4 router still outperformed BIOT by 5.885.88 percentage points. Applying the same TRAIN-only 95% dominant-prediction criterion to ISRUC excluded none of its seven pathways.

Together, these results show that routing gains do not require a fixed seven-pathway portfolio, while the opportunities for pathway pruning depend on dataset-specific prediction behavior.

Weakness 2:

BridgeRouter separates global pathway selection from instance-level adaptation. The SBS is selected using validation performance and serves as the anchor, while the router switches only when the predicted gain exceeds a development-selected threshold. This preserves a strong fixed-pathway reference and avoids unrestricted switching.

On ISRUC, the corrected router improves over the validation-selected CBraMod anchor by 3.833.83 percentage points, with positive gains across all five seeds. This provides evidence that the learned switching rule can improve upon the validation-selected pathway in the evaluated setting. We will clarify the selection procedure and its dependence on representative development data.

Weakness 3:

We agree that two datasets do not rule out dataset-specific effects. We selected FOG (EEG-to-EMG transfer for freezing-of-gait detection) and ISRUC (EEG-to-ECG transfer for sleep staging) because they are the two EEG-related benchmarks in BioX-Bridge, which is an accepted ICLR conference paper in 2026 and provide a common benchmark basis for studying what instance-level routing adds to bridge-based transfer. BioX-Bridge also evaluates WESAD, but that benchmark concerns ECG–PPG transfer and falls outside our focus on transfer from EEG foundation models. Our results therefore cover two tasks and two target modalities, with one dataset per setting.

Weakness 4:

Direct target-side learning and cross-modal pathway routing address different prediction strategies. A target-only classifier learns a task-specific decision function directly from labeled target representations, whereas BridgeRouter selects among existing frozen cross-modal predictors based on estimated instance-level gains.

Direct target-side learning provides an important alternative to pathway routing. On FOG, the target-only NormWear MLP achieved 83.65±0.41%83.65 \pm 0.41\% BAcc; on ISRUC, the H4 linear and MLP controls achieved 69.01%69.01\% and 67.13±1.61%67.13 \pm 1.61\%, respectively. These results demonstrate that direct target supervision is competitive and can outperform routing.

BridgeRouter addresses a complementary question: whether target supervision can be used to adaptively select among existing frozen cross-modal predictors rather than train a new target-task prediction pathway. Our controlled comparisons demonstrate improvements over validation-selected fixed pathways on both benchmarks. We will clarify this distinction in the revised manuscript.

Question 1:

The oracle measures the potential headroom from perfect instance-level pathway selection, rather than the performance attainable by a learned router. The remaining gap highlights the challenge of translating oracle-level pathway complementarity into reliable instance-level routing decisions. Future improvements could focus on more reliable gain estimation, better-calibrated switching decisions, and richer representations of pathway-specific uncertainty.

Importantly, oracle headroom alone does not guarantee learnable complementarity. As shown in our response to sCt2 (Weakness 3), the Compact-4 experiment demonstrates that routing retains a measurable improvement over the SBS even after excluding near-single-class pathways. This provides evidence that the observed routing gains are not solely explained by degenerate pathway predictions.

Question 2:

A strong foundation model is a natural candidate for the default, but the relevant question is whether its bridged pathway is the best choice for every target input. Our FOG and ISRUC results show that, in the evaluated portfolios, a strong fixed pathway and an adaptive router are complementary: the router retains the default and switches when it predicts an advantage for another pathway. If a future pathway is consistently the best choice, it can be used alone. In the settings we study, however, alternative pathways provide measurable gains without requiring multiple pathway executions per prediction.

Question 3:

We see these as two complementary ways to build a pathway portfolio. In this paper, we use one bridge per EEG foundation model to study whether instance-level routing can exploit differences across backbones. Multiple bridges attached to the same model offer a different opportunity: the bridges can specialize on different subsets of inputs while sharing a backbone. We have begun exploring this direction on both FOG and ISRUC dataset by training a second bridge for the same SBS model correspondingly, with more weight on instances that the first bridge aligns poorly.

Thank you for the responses and additional experiments. I appreciate the clarification regarding portfolio size. While the rebuttal addresses some of my concerns, the incremental nature and generalization remain insufficiently justified. I maintain my original score for now, though I remain open to further clarification or evidence.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.