acceptodds
Under review as a conference paper at ICLR 2027

Data-Efficient Cross-Session Adaptation for Brain–Computer Interfaces via Structured Drift Modeling

Abstract

Long-term brain–computer interface (BCI) use requires neural decoders that can rapidly recover stable performance across recording sessions despite substantial neural nonstationarity. Existing cross-session adaptation methods often rely on batch-style recalibration and relatively high-dimensional parameter updates, limiting their effectiveness when only a small amount of new-session data is available. To address this limitation, we present the Distribution-based Drift Adaptation (DiDA) framework, which captures session-specific drift state through a compact distributional representation while preserving shared latent dynamics and a stable behavioral readout. By adapting this distribution rather than re-optimizing the full observation mapping, DiDA enables efficient cross-session adaptation as new trials become available sequentially. We further enable adaptation by leveraging lightweight feedback as a weak supervisory signal, rather than relying on purely unsupervised objectives. We evaluate the framework on motor cortical recordings from macaques performing center-out and random-target reaching tasks. Our experiments examine trial-by-trial adaptation under supervised, unsupervised, and preference-guided objectives, with offline batch fitting serving as a performance reference. Across these settings, DiDA enables rapid, data-efficient cross-session recalibration, highlighting its potential for long-term BCI use. Code is available at https://anonymous.4open.science/r/DiDA-3069/.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Brain–computer interface (BCI), cross-session adaptation, neural decoding, latent dynamics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_4pgyOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies few-trial calibration for cross-session neural decoding. The authors propose DiDA, which represents session-specific drift using a low-dimensional Gaussian distribution and generates alignment transformations from sampled drift states. The model is pretrained on Day 0 data and augmented versions of those data. During adaptation to a new session, only the mean and standard deviation of the drift distribution are updated, while the shared encoder, latent dynamics, and behavioral decoder remain fixed. The authors evaluate supervised, unsupervised, and preference-guided adaptation on macaque center-out and random-target reaching tasks. Under the evaluated few-trial, trial-by-trial adaptation settings, DiDA generally outperforms the compared methods.

Critical strengths
  1. The paper addresses the limited availability of calibration data at the beginning of a new session, with a clear motivation and supporting experiments. DiDA performs well in few-trial sequential adaptation, particularly in the unsupervised setting. These results support constraining adaptation through a pretrained structure and are relevant to reducing repeated calibration in BCIs.
  2. Preference-guided adaptation is a direction worth exploring. The drift distribution produces candidate decoded trajectories, and preferences between these trajectories are used to update the distribution. This offers an alternative when precise behavioral labels are unavailable. The authors clearly acknowledge that the current preferences are generated by an oracle using ground-truth trajectories. Although these experiments do not establish effectiveness with human feedback, they provide an initial proof of concept.
Critical weaknesses
  1. The respective contributions of the low-dimensional constraint and the probabilistic representation remain unclear. The appendix examines regularization terms, dynamics constraints, and drift dimensionality, but does not directly compare learning a drift distribution, optimizing a deterministic low-dimensional state, and optimizing an unconstrained transformation matrix under controlled conditions. The role of the distribution in candidate sampling and preference optimization is relatively clear. For supervised and unsupervised adaptation, however, it remains difficult to determine how much of the improvement comes from the low-dimensional constraint, probabilistic modeling, or other pretraining choices. This limits my assessment of the contribution of each design choice.

  2. Several aspects of the baseline comparisons and the consistency of the gains need clarification. In Table 2, some baselines have lower mean performance under supervised adaptation than under unsupervised adaptation, along with substantial variability across sessions. This does not by itself indicate inadequate tuning, but warrants an explanation in terms of hyperparameters, loss weights, and update budgets. The differences between the present setting and related methods such as FDA [1] and SPINT [2], and the reasons for not comparing them directly, also deserve discussion. Moreover, the aggregate means and standard deviations do not show whether DiDA improves performance in most sessions or how sensitive the results are to random seeds. These issues limit my confidence in the comparative conclusions.

[1] Wang, P., Qi, Y., Wang, Y., and Pan, G. Flow Matching for Few-Trial Neural Adaptation with Stable Latent Dynamics. In Proceedings of the 42nd International Conference on Machine Learning, Proceedings of Machine Learning Research, 267:64059–64083, 2025. [2] Le, T., Fang, H., Li, J., Nguyen, T., Mi, L., Orsborn, A. L., Sümbül, U., and Shlizerman, E. SPINT: Spatial Permutation-Invariant Neural Transformer for Consistent Intracortical Motor Decoding. In Advances in Neural Information Processing Systems, 38, 2025.

Questions for authors
  1. Could you compare three adaptation variants using the same pretrained DiDA model: (i) the original probabilistic drift distribution; (ii) a deterministic low-dimensional drift vector, retaining the fixed transformation-generating network; and (iii) an unconstrained session-specific transformation matrix optimized directly, without the low-dimensional parameterization? This would help distinguish the effects of probabilistic modeling and the low-dimensional constraint. If the existing Align-seqVAE experiment already provides the third control, please clarify whether the shared backbone, pretraining data, and adaptation settings are the same, and identify the corresponding results.
  2. The appendix reports DiDA’s adaptation steps and examines the update interval. Could you clarify how the hyperparameters, loss weights, and update budgets were selected for each main baseline? Is there any further diagnosis or explanation for the lower means and greater variability after adding behavioral supervision in some Table 2 results, such as CycleGAN with 25 adaptation trials? Please also discuss the comparability of FDA [1], CRRL [2], and SPINT [3] to the present setting and the reasons for not including direct comparisons. If comparable results are already available, please provide them.
  3. The reported means and standard deviations are computed across target sessions. Were the main comparisons also repeated with multiple random seeds? Could you provide per-session paired performance differences between DiDA and the main baselines, to show whether the average gains reflect improvements in most sessions or are driven primarily by a small number of sessions?
5 comments

We thank the reviewer for the insightful feedback and constructive recommendations. We will address these points in the revision to further improve the quality and completeness of our work.

W1 & Q1 Probabilistic vs. deterministic and unconstrained adaptation

The third variant is already covered by Align-seqVAE, which optimizes a full session-specific transformation matrix using the same backbone architecture, pretraining data, and adaptation protocol. We have additionally evaluated the second variant by replacing the probabilistic drift distribution with a deterministic low-dimensional vector while keeping the pretrained DiDA model fixed. The deterministic variant performs substantially worse, particularly under unsupervised adaptation. We attribute this difference to the ambiguity of drift estimation under limited supervision. While deterministic adaptation optimizes a single drift state that may be sensitive to noisy or weakly constrained alignment signals, probabilistic adaptation optimizes the expected loss over sampled drift states, introducing a smoothing effect on the optimization landscape. This can reduce sensitivity to sharp local minima and encourage solutions that perform consistently across neighboring drift states. Together, these comparisons highlight the complementary benefits of probabilistic modeling and structured low-dimensional parameterization.

Tab. 1 Comparison of probabilistic, deterministic, and unconstrained drift adaptation on CO.

Supervision ModeModel Variants102550
SupervisedProbabilistic Drift0.6961 ±\pm 0.10270.7646 ±\pm 0.06980.7968 ±\pm 0.0539
SupervisedDeterministic Drift0.4902 ±\pm 0.11030.6101 ±\pm 0.09650.6788 ±\pm 0.0897
SupervisedUnconstrained Transform0.6492 ±\pm 0.08640.7144 ±\pm 0.09740.7637 ±\pm 0.0588
UnsupervisedProbabilistic Drift0.5262 ±\pm 0.17400.6107 ±\pm 0.14140.6637 ±\pm 0.1002
UnsupervisedDeterministic Drift-0.2870 ±\pm 0.3938-0.1902 ±\pm 0.3522-0.0426 ±\pm 0.3202
UnsupervisedUnconstrained Transform-0.1624 ±\pm 0.2546-0.0651 ±\pm 0.2584-0.0144 ±\pm 0.2712

Tab. 2 Comparison of probabilistic, deterministic, and unconstrained drift adaptation on CO.

Supervision ModeModel Variants102550
SupervisedProbabilistic Drift0.4923 ±\pm 0.16420.5302 ±\pm 0.17340.5798 ±\pm 0.0970
SupervisedDeterministic Drift0.1845 ±\pm 0.40190.3743 ±\pm 0.19400.4230 ±\pm 0.2179
SupervisedUnconstrained Transform0.1068 ±\pm 0.94460.0266 ±\pm 0.96320.1807 ±\pm 0.7208
UnsupervisedProbabilistic Drift0.1909 ±\pm 0.54710.3331 ±\pm 0.31590.3625 ±\pm 0.3009
UnsupervisedDeterministic Drift-1.6615 ±\pm 1.3259-0.7563 ±\pm 1.4232-0.1291 ±\pm 0.6610
UnsupervisedUnconstrained Transform0.0606 ±\pm 0.3700-0.0612 ±\pm 0.60670.0588 ±\pm 0.4284

W2 & Q2 Additional baselines

We thank the reviewer for these suggestions. We have added FDA and SPINT as additional baselines under our sequential adaptation protocol.

For SPINT, its gradient-free adaptation avoids test-time parameter updates, but also limits its ability to explicitly correct session-specific drift using incoming trials. In contrast, DiDA progressively refines its drift distribution as more observations become available.

For FDA, we selected its MMD variant as the baseline because it demonstrated better performance in the original paper. However, we observed substantial sensitivity to hyperparameter selection and occasional representation collapse. A key issue is that both source and target distributions are produced by the same trainable encoder during adaptation. Consequently, minimizing MMD allows both distributions to shift together toward degenerate solutions without necessarily preserving behaviorally relevant representations. We therefore consider FDA a less reliable baseline for evaluating strictly unsupervised adaptation, as its reported performance may benefit from information that would not be available in practical label-free settings.

We will include the additional baseline results and implementation details in the revised manuscript.

Tab. 3 Comparison of DiDA with FDA and SPINT on the CO task.

Supervision ModeMethod102550Offline
SupervisedDiDA0.6961 ±\pm 0.10270.7646 ±\pm 0.06980.7968 ±\pm 0.05390.8263 ±\pm 0.0414
SupervisedFDA0.5400 ±\pm 0.10630.6736 ±\pm 0.08770.7720 ±\pm 0.06290.8939 ±\pm 0.0197
UnsupervisedDiDA0.5262 ±\pm 0.17400.6107 ±\pm 0.14140.6637 ±\pm 0.10020.6660 ±\pm 0.0973
UnsupervisedFDA0.2052 ±\pm 0.14570.0952 ±\pm 0.06800.0271 ±\pm 0.02340.3139 ±\pm 0.2024
UnsupervisedSPINT0.2508 ±\pm 0.22900.2648 ±\pm 0.20790.2852 ±\pm 0.22550.3390 ±\pm 0.2151

Tab. 4 Comparison of DiDA with FDA and SPINT on the RT task.

Supervision ModeMethod102550Offline
SupervisedDiDA0.4923 ±\pm 0.16420.5302 ±\pm 0.17340.5798 ±\pm 0.09700.6067 ±\pm 0.0803
SupervisedFDA0.2470 ±\pm 0.12490.2888 ±\pm 0.14410.2774 ±\pm 0.16890.5901 ±\pm 0.0727
UnsupervisedDiDA0.1909 ±\pm 0.54710.3331 ±\pm 0.31590.3625 ±\pm 0.30090.4537 ±\pm 0.1295
UnsupervisedFDA0.1959 ±\pm 0.09130.1038 ±\pm 0.06110.0111 ±\pm 0.03080.2922 ±\pm 0.1395
UnsupervisedSPINT0.0543 ±\pm 0.15480.1556 ±\pm 0.12780.2072 ±\pm 0.14130.2511 ±\pm 0.1428

W2 & Q4 Multi-seed evaluation, paired per-session differences, and performance variability.

We thank the reviewer for this suggestion. We have repeated the main experiments with multiple random seeds and will update the reported results by averaging across seeds. We have also added paired per-session performance differences between DiDA and the main baselines to assess whether the improvements are consistent across sessions rather than driven by a few favorable cases. The per-session analysis helps explain the large variability and occasional performance degradation observed in Table 2, particularly for CycleGAN under supervised adaptation with 25 trials. We found that a small number of challenging target sessions exhibit unusually poor decoding performance, substantially lowering the mean and increasing the standard deviation. Thus, these fluctuations are largely driven by extreme session-specific outcomes rather than a consistent degradation across all sessions. We will include the multi-seed results and per-session analyses in the revised manuscript to make these effects more transparent.

Importantly, our comparison focuses on adaptation efficiency under limited sequential observations, rather than maximum performance with unrestricted optimization. The strong offline performance of the baselines, together with the relatively comparable supervised results, indicates that they are capable of learning effective transformations. However, their performance degrades more substantially under unsupervised sequential adaptation, where limited observations provide weaker and noisier gradient signals. This interpretation is further supported by Appendix F.1, where Align-seqVAE improves substantially with larger update intervals, while DiDA remains relatively stable. These findings suggest that the performance differences are closely related to the methods' sensitivity to sparse sequential updates, rather than simply inadequate hyperparameter tuning. We will clarify these details in the appendix.

Tab. 5 Per-session supervised adaptation performance (nadapt=25n_{adapt}=25) on the CO task.

Day offsetDiDAAlign-SeqVAEERDiffCycleGANFDA
10.8670 ±\pm 0.00720.6401 ±\pm 0.05190.6163 ±\pm 0.08440.5139 ±\pm 0.05440.5620 ±\pm 0.106
20.8231 ±\pm 0.01580.7381 ±\pm 0.01520.6183 ±\pm 0.10800.6226 ±\pm 0.03070.7136 ±\pm 0.0112
30.6234 ±\pm 0.06070.4551 ±\pm 0.03500.4897 ±\pm 0.05570.4069 ±\pm 0.04490.4954 ±\pm 0.1055
90.7886 ±\pm 0.00760.6987 ±\pm 0.02770.6951 ±\pm 0.00600.6437 ±\pm 0.01800.5857 ±\pm 0.0328
100.7543 ±\pm 0.04490.7322 ±\pm 0.02740.6155 ±\pm 0.12040.6478 ±\pm 0.02090.6722 ±\pm 0.0388
140.7893 ±\pm 0.02780.7459 ±\pm 0.01790.6897 ±\pm 0.05230.6976 ±\pm 0.01000.7182 ±\pm 0.0357
150.8195 ±\pm 0.01080.7608 ±\pm 0.02030.7608 ±\pm 0.03720.6804 ±\pm 0.03230.7447 ±\pm 0.0070
160.6711 ±\pm 0.09590.7602 ±\pm 0.01670.7158 ±\pm 0.05490.6706 ±\pm 0.05140.7313 ±\pm 0.0217
360.7524 ±\pm 0.01240.7302 ±\pm 0.05290.6797 ±\pm 0.02940.6255 ±\pm 0.02390.6780 ±\pm 0.0222
370.7320 ±\pm 0.10270.8239 ±\pm 0.01720.6975 ±\pm 0.03710.6835 ±\pm 0.02840.7623 ±\pm 0.0157
380.7903 ±\pm 0.05430.7734 ±\pm 0.01710.6086 ±\pm 0.08120.6824 ±\pm 0.01820.7455 ±\pm 0.0436
Reviewer_LcPMOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes Distribution-based Drift Adaptation (DiDA) for cross-session neural decoding. A model is pretrained on Day 0 recordings and four synthetically perturbed versions of those recordings. Shared latent dynamics and a behavioral decoder are then frozen. Adaptation updates the mean and diagonal variance of a low-dimensional Gaussian drift state, which generates session-specific neural transformations. The framework supports behavioral supervision, Gaussian latent-distribution matching, and direct preference optimization using rankings of candidate decoded trajectories.

The experiments use macaque center-out and random-target reaching recordings. Adaptation proceeds through a prefix of a new session, while performance is measured on its final 20 trials. DiDA generally improves early-session velocity decoding over Align-seqVAE, ERDiff, and CycleGAN.

Critical strengths
  1. The evaluation targets a useful calibration regime. Keeping the behavioral decoder fixed and measuring performance after 10, 25, and 50 observed trials addresses a practical question that an offline endpoint alone would miss. At 25 trials, CO unsupervised R2 is 0.6362 for DiDA versus 0.3592 for the strongest listed baseline in that setting; at 50 trials, RT supervised R2 is 0.5920 versus 0.4043 (Tables 1–2, pp. 7–8). These are meaningful observed improvements under the stated protocol.

  2. The appendix investigates when DiDA is useful. The authors vary the drift-code size and the number of trials collected between updates. They also show that Align-seqVAE can outperform DiDA when all available adaptation trials and movement labels are used. These experiments help identify the conditions under which DiDA offers an advantage.

Critical weaknesses
  1. The method combines several training objectives and adaptation components, but is evaluated only on two tasks from the same data source, with a single real pretraining session per task. This limited setup makes it difficult to judge whether the added complexity brings a general benefit over simpler approaches. Broader evaluation and simpler baselines using the same data and augmentations are needed to establish its overall usefulness.

  2. Important dataset details are missing, including a clear breakdown of subjects, sessions, and trials per task, the number of neural channels, and trial duration. This makes the amount of training and evaluation data difficult to assess. In addition, preference feedback is simulated using ground-truth trajectories rather than collected from real users. Although the authors acknowledge this limitation, the practical value of user-guided adaptation remains untested.

Questions for authors
  1. Does DiDA outperform simpler baselines using the same data and augmentations, and do the gains persist across other pretraining sessions or datasets?

  2. Can you provide the subject, session, and trial counts for each task and data split, together with the neural channel count and trial duration?

  3. How would preferences be collected from real users, and what evidence supports the method's effectiveness with limited or noisy feedback rather than oracle rankings?

6 comments

We sincerely thank the reviewer for the constructive feedback and valuable suggestions. We will carefully revise the manuscript accordingly to improve its clarity and rigor.

W1 & Q1 Comparison with simpler baselines

We thank the reviewer for this suggestion. Our baselines are selected to cover representative cross-session adaptation strategies: ERDiff uses diffusion-based alignment, CycleGAN employs adversarial distribution matching, and Align-seqVAE serves as a direct comparison using a full transformation matrix. We further add SPINT[1], which enables adaptation without test-time fine-tuning, and Procrustes[2], which provides a simple closed-form unsupervised alignment reference.

Importantly, while these approaches differ in their alignment objectives and adaptation mechanisms, DiDA primarily improves what is optimized during adaptation, rather than introducing another alignment objective. Specifically, it replaces high-dimensional transformation updates with optimization of a compact probabilistic drift state, enabling more reliable adaptation with limited data.

For fair comparison, we apply the same augmentations to methods supporting end-to-end decoder pretraining (ERDiff, Align-seqVAE, and SPINT). In contrast, CycleGAN relies on adversarial alignment to a fixed reference neural distribution with a separately trained decoder, while Procrustes computes a closed-form geometric alignment. Therefore, neither directly supports joint representation learning across augmented conditions in its original formulation. We have also updated the experimental results by conducting additional runs with different random seeds, following Reviewer 4pgy's suggestion.

Tab. 1 Performance Comparison on CO.

Supervision ModeMethod102550
SupervisedDiDA0.6961 ±\pm 0.10270.7646 ±\pm 0.06980.7968 ±\pm 0.0539
SupervisedAlign-SeqVAE0.6492 ±\pm 0.08640.7144 ±\pm 0.09740.7637 ±\pm 0.0588
SupervisedERDiff0.5213 ±\pm 0.07330.6534 ±\pm 0.07350.7209 ±\pm 0.0601
SupervisedCycleGAN0.5832 ±\pm 0.09420.6250 ±\pm 0.08830.6824 ±\pm 0.0505
UnsupervisedDiDA0.5262 ±\pm 0.17400.6107 ±\pm 0.14140.6637 ±\pm 0.1002
UnsupervisedAlign-SeqVAE-0.1624 ±\pm 0.2546-0.0651 ±\pm 0.2584-0.0144 ±\pm 0.2712
UnsupervisedERDiff0.0224 ±\pm 0.10820.0026 ±\pm 0.2440-0.1264 ±\pm 0.3328
UnsupervisedCycleGAN0.3540 ±\pm 0.08460.3469 ±\pm 0.08910.3486 ±\pm 0.0832
UnsupervisedProcrustes0.3754 ±\pm 0.14600.4307 ±\pm 0.12480.4467 ±\pm 0.1226
UnsupervisedSPINT0.2508 ±\pm 0.22900.2648 ±\pm 0.20790.2852 ±\pm 0.2255

Tab. 2 Performance Comparison on RT.

Supervision ModeMethod102550
SupervisedDiDA0.4923 ±\pm 0.16420.5302 ±\pm 0.17340.5798 ±\pm 0.0970
SupervisedAlign-SeqVAE0.1068 ±\pm 0.94460.0266 ±\pm 0.96320.1807 ±\pm 0.7208
SupervisedERDiff0.2414 ±\pm 0.27070.2145 ±\pm 0.49570.3383 ±\pm 0.2983
SupervisedCycleGAN0.0562 ±\pm 1.05580.2078 ±\pm 0.63410.2313 ±\pm 0.5721
UnsupervisedDiDA0.1909 ±\pm 0.54710.3331 ±\pm 0.31590.3625 ±\pm 0.3009
UnsupervisedAlign-SeqVAE0.0606 ±\pm 0.3700-0.0612 ±\pm 0.60670.0588 ±\pm 0.4284
UnsupervisedERDiff-0.0749 ±\pm 0.1805-0.0209 ±\pm 0.0512-0.0105 ±\pm 0.0336
UnsupervisedCycleGAN0.2268 ±\pm 0.23540.2785 ±\pm 0.16090.2495 ±\pm 0.2042
UnsupervisedProcrustes-0.6452 ±\pm 1.9468-5.4434 ±\pm 17.55740.1050 ±\pm 0.2731
UnsupervisedSPINT0.0543 ±\pm 0.15480.1556 ±\pm 0.12780.2072 ±\pm 0.1413

[1] Le, T., Fang, H., Li, J., Nguyen, T., Mi, L., Orsborn, A. L., ... & Shlizerman, E. (2026). SPINT: Spatial permutation-invariant neural transformer for consistent intracortical motor decoding. Advances in Neural Information Processing Systems, 38, 25815-25835.

[2] Degenhart, A. D., Bishop, W. E., Oby, E. R., Tyler-Kabara, E. C., Chase, S. M., Batista, A. P., & Yu, B. M. (2020). Stabilization of a brain–computer interface via the alignment of low-dimensional spaces of neural activity. Nature biomedical engineering, 4(7), 672-685.

Q1 Generalization across pretraining sessions.

We acknowledge that our evaluation is limited to two tasks from the same data source. Nevertheless, CO and RT represent distinct reaching paradigms, with longitudinal recordings spanning 38 days (CO) and 79 days (RT), allowing us to assess adaptation across different behavioral and recording conditions. To further evaluate generalizability, we have conducted additional experiments using different pretraining sessions, examining the robustness of DiDA to the choice of source session. We will include these results in the revised manuscript, while recognizing that broader cross-dataset validation remains valuable.

Tab 3. Performance under supervised adaptation across pretraining sessions on CO.

Source / TargetDay 0Day 1Day 2Day 3Day 9Day 10Day 14Day 15Day 16Day 36Day 37Day 38
Day 0—0.86690.82950.63070.79080.79630.82070.82380.58520.75050.61450.7295
Day 1——0.89460.81500.82240.79960.81980.85420.74300.73640.74200.8572
Day 2———0.82660.86740.82070.83430.89210.79160.73350.81260.8207
Day 3————0.80520.83920.86020.83250.79490.77720.69170.7741
Day 9—————0.82130.83180.83090.85920.82100.85520.8692
Day 10——————0.82510.83880.74080.81960.72280.8523
Day 14———————0.84490.80160.78280.83010.8320
Day 15————————0.83160.82450.80280.7789
Day 16—————————0.79320.81660.8133
Day 36——————————0.87950.9031
Day 37———————————0.9037
Day 38————————————

Tab 4. Performance under unsupervised adaptation across pretraining sessions on CO.

Source / TargetDay 0Day 1Day 2Day 3Day 9Day 10Day 14Day 15Day 16Day 36Day 37Day 38
Day 0—0.84780.71520.55710.60700.64990.62150.71490.43850.70870.55610.5523
Day 1——0.84310.64970.80300.68150.62660.76570.46640.67310.55840.6686
Day 2———0.72030.84830.69780.73310.7696-0.09640.57030.68070.7107
Day 3————0.78220.61200.78590.70720.65760.55240.57230.5162
Day 9—————0.78220.73930.81530.74700.75260.81730.7920
Day 10——————0.80560.78110.69290.76180.71400.6889
Day 14———————0.76030.73000.61350.79960.7071
Day 15————————0.52060.63720.59360.5428
Day 16—————————0.45670.62120.7193
Day 36——————————0.80710.8550
Day 37———————————0.7999
Day 38————————————

Tab 5. Performance under preference-guided adaptation across pretraining sessions on CO.

Source / TargetDay 0Day 1Day 2Day 3Day 9Day 10Day 14Day 15Day 16Day 36Day 37Day 38
Day 0—0.85800.73000.38570.74250.63830.72210.70150.68350.61950.59850.6596
Day 1——0.85470.63390.85140.72720.71040.61550.62280.70540.57510.7604
Day 2———0.77350.85410.71120.82240.84200.61640.43600.67260.7631
Day 3————0.78950.67850.81030.64930.75880.55370.61310.7283
Day 9—————0.79740.77020.79970.76330.59180.67770.8442
Day 10——————0.71860.76780.52050.72480.69540.6644
Day 14———————0.63490.59520.60840.67550.7933
Day 15————————0.78880.42360.67740.6645
Day 16—————————0.74790.80160.7490
Day 36——————————0.83250.8739
Day 37———————————0.8373
Day 38————————————

Tab 6. Performance under supervised adaptation across pretraining sessions on RT.

Source / TargetDay 0Day 1Day 38Day 39Day 40Day 52Day 53Day 67Day 69Day 77Day 79
Day 0—0.71130.72930.66610.57490.50900.47960.5381-0.47310.52410.5896
Day 1——0.73060.63140.57610.52200.60260.55130.14140.58310.6218
Day 38———0.67300.69200.42940.55690.5751-1.00570.52700.5879
Day 39————0.66580.38140.55490.4595-0.02980.49220.5507
Day 40—————0.42030.46290.56370.14440.50150.6075
Day 52——————0.60070.65400.42280.60550.6755
Day 53———————0.50400.23410.55710.6569
Day 67————————0.46120.62080.6653
Day 69—————————0.58740.6649
Day 77——————————0.6884
Day 79———————————

Tab 7. Performance under unsupervised adaptation across pretraining sessions on RT.

Source / TargetDay 0Day 1Day 38Day 39Day 40Day 52Day 53Day 67Day 69Day 77Day 79
Day 0—0.63730.63380.61020.43090.38820.40030.3909-0.28180.46530.3996
Day 1——0.67990.56290.48950.47230.47330.4822-0.32310.42390.4341
Day 38———0.56220.63480.28250.41830.4776-1.77510.36040.4590
Day 39————0.60720.16080.44820.1151-0.73010.22070.3026
Day 40—————0.22180.39480.46980.21490.21130.4780
Day 52——————0.59690.57200.27960.51790.5390
Day 53———————0.3580-0.98640.49270.6141
Day 67————————0.24780.40120.5820
Day 69—————————0.49150.5940
Day 77——————————0.6952
Day 79———————————

Tab 8. Performance under preference-guided adaptation across pretraining sessions on RT.

Source / TargetDay 0Day 1Day 38Day 39Day 40Day 52Day 53Day 67Day 69Day 77Day 79
Day 0—0.62900.64530.44220.46340.42730.49400.46930.20830.39240.4116
Day 1——0.58820.60610.42440.50300.42620.4377-0.55530.45360.2488
Day 38———0.54220.41660.31110.39680.4067-0.11990.31850.4078
Day 39————0.62330.13610.37740.22810.03320.15500.3424
Day 40—————0.21460.39920.42810.00520.26220.4126
Day 52——————0.46720.50820.26140.53360.5049
Day 53———————0.34480.20610.54790.5606
Day 67————————0.36480.46970.5798
Day 69—————————0.48890.4645
Day 77——————————0.5618
Day 79———————————
Reviewer_H87qOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

I am not an expert in BCI.

This paper attempts to build an algorithm to correct for drift in neural representations, sampling, and coding across time in BCI control. The attempt here is to describe the drift process as such in stead of one that requires full retraining. The main idea is that one can retrain the neural to latent map with out retraining the latent space and its dynamics. They map across session drift as a probabilistic drift state.

I have raised my score to 3

Critical strengths

The authors take a set of existing approaches and link them here to do something new.

Critical weaknesses

The drift subspace is 64 dimensional but trained with 5 sessions. It does not seem like this can not be done such that much of the Wc must be unconstrained.

Most of the approaches are described in the literature - they are just strung together here.

The stats seem problematic Standard deviations across sessions are large, and no paired tests are reported. Several of the wins fall within noise. RT supervised is 0.51 at 10 trials, drops to 0.47 ± 0.39 at 25, then rises to 0.59 at 50. Ablations often improve when a component is removed (CO sequential)

Questions for authors

Can you report paired per-session differences?

DiDA seems to do well only in the low data region. Is this due to the regularization?

There is no true online performance tested

128 dimensions seems to do better than 64. Why was it not used throughout?

3 comments

We appreciate the reviewer's helpful suggestions, which will be incorporated into the revised manuscript to strengthen its presentation and technical clarity.

W1 Drift dimensionality and limited pretraining sessions

We thank the reviewer for raising this important concern. We would like to clarify that the five training sessions do not correspond to only five fixed drift vectors used to learn WcW_c. In DiDA, each session is associated with a Gaussian drift distribution qη(s)(c)q_{\eta^{(s)}}(c), rather than a single deterministic drift state. During pretraining, drift states are repeatedly sampled from these distributions, and the shared mapping WcW_c is jointly optimized through neural reconstruction and behavioral decoding objectives. Therefore, the training of WcW_c is not restricted to the span of five fixed vectors. Moreover, the KL regularization toward a standard Gaussian prior encourages the drift distributions to remain within a common, non-degenerate sampling space, allowing the shared mapping to receive training signals across different drift directions.

We agree, however, that this does not guarantee full identifiability of all directions in WcW_c. With only five training conditions, some directions may remain weakly constrained, particularly if they have limited influence on the training objectives. Our approach does not assume that every direction is equally identifiable; rather, it learns a shared transformation family whose behaviorally relevant directions are constrained by the reconstruction and decoding objectives, while the prior regularization limits excessive deviations in the drift distributions. The effectiveness of this parameterization is supported by adaptation to held-out real recording sessions spanning 38 days for CO and 79 days for RT. We will clarify this distinction between the number of training sessions, stochastic drift-state sampling, and parameter identifiability in the revised manuscript.

W2 Novelty beyond combining existing methods

While DiDA builds upon established techniques, its novelty lies in addressing several practical challenges of long-term BCI adaptation. First, unlike conventional approaches that primarily focus on decoding accuracy after sufficient calibration, we emphasize rapid performance recovery during the early calibration stage, when only a few trials are sequentially available. Second, since precise behavioral labels are not always accessible in practical BCI use, we consider different levels of available information, including supervised, unsupervised, and preference-guided adaptation. DiDA introduces a probabilistic drift representation that enables efficient sequential adaptation while preserving the pretrained decoder, and naturally supports preference-guided optimization. Rather than simply combining existing objectives, our contribution is a unified adaptation framework designed around the data scarcity and feedback constraints of practical BCI use, evaluated across different reaching tasks and longitudinal recording sessions.

W3 & Q1 Statistical reliability and paired per-session comparisons

We appreciate this concern. We have added per-session results to enable direct paired comparisons between methods. Following Reviewer 4pgy's suggestion, we have also repeated the experiments with multiple random seeds and updated the reported results by averaging across seeds. These additions provide a more reliable assessment of performance variability and help address the previously observed fluctuations, particularly on RT.

Tab. 1 RT velocity decoding performance of DiDA under the supervised mode across seeds

Day offset / nadaptn_{adapt}102550
10.6832 ±\pm 0.02060.6998 ±\pm 0.01810.7069 ±\pm 0.0142
380.6720 ±\pm 0.02520.7109 ±\pm 0.02590.7329 ±\pm 0.0194
390.5976 ±\pm 0.00380.6386 ±\pm 0.00520.6430 ±\pm 0.0031
400.5430 ±\pm 0.01350.5906 ±\pm 0.02500.6070 ±\pm 0.0128
520.4845 ±\pm 0.01170.5035 ±\pm 0.01230.5101 ±\pm 0.0157
530.5113 ±\pm 0.02130.5122 ±\pm 0.02100.5402 ±\pm 0.0096
670.3585 ±\pm 0.16370.4271 ±\pm 0.18060.4720 ±\pm 0.0653
690.1161 ±\pm 0.22350.1068 ±\pm 0.10240.4417 ±\pm 0.0502
770.4699 ±\pm 0.02870.5288 ±\pm 0.00290.5329 ±\pm 0.0084
790.4866 ±\pm 0.03290.5837 ±\pm 0.04120.6114 ±\pm 0.0443

Tab. 2 RT velocity decoding performance of DiDA under the unsupervised mode across seeds

Day offset / nadaptn_{adapt}102550
10.5978 ±\pm 0.04340.6135 ±\pm 0.03190.6360 ±\pm 0.0191
380.5935 ±\pm 0.00710.6302 ±\pm 0.00320.6462 ±\pm 0.0123
390.4629 ±\pm 0.06040.5352 ±\pm 0.01120.5580 ±\pm 0.0060
400.3480 ±\pm 0.08950.4259 ±\pm 0.07090.4554 ±\pm 0.0750
520.1627 ±\pm 0.09250.2971 ±\pm 0.09860.3276 ±\pm 0.1022
530.3174 ±\pm 0.05050.3741 ±\pm 0.05820.3974 ±\pm 0.0661
670.1595 ±\pm 0.13240.1910 ±\pm 0.11260.2190 ±\pm 0.0542
69-1.3014 ±\pm 0.6768-0.4724 ±\pm 0.1295-0.4032 ±\pm 0.3519
770.3371 ±\pm 0.11170.4280 ±\pm 0.03600.4234 ±\pm 0.0160
790.2311 ±\pm 0.35140.3084 ±\pm 0.26560.3647 ±\pm 0.2262

W3 Ablation results and component effectiveness

These regularization terms are designed to encourage a more robust and session-invariant latent space, which is particularly important for unsupervised adaptation where behavioral labels are unavailable. Their benefits are therefore more evident in the unsupervised setting. In contrast, supervised adaptation already benefits from strong behavioral supervision, making these additional constraints less critical and sometimes slightly restrictive, resulting in small performance differences or occasional drops.

Q2 Effectiveness in the low-data regime

Our focus on the low-data regime is intentional, as rapid adaptation with minimal calibration data is particularly important for practical BCI applications. The key challenge is the trade-off between adaptation flexibility and data efficiency: highly flexible transformations can achieve better performance with sufficient data, but are difficult to estimate reliably from only a few trials. DiDA addresses this trade-off by restricting adaptation to a pretrained, structured drift space, enabling more reliable updates with limited observations. Regularization primarily helps learn a robust and session-invariant latent space during pretraining, rather than being the main reason for DiDA's advantage in the low-data regime.

Q3 Lack of true online evaluation

We agree that true closed-loop evaluation is important for validating DiDA in practical BCI systems, particularly because decoder updates can influence subsequent user behavior and neural activity. However, sequential replay provides a necessary and controlled setting for investigating key challenges before closed-loop deployment. By adapting progressively without accessing future data, we systematically evaluate early-session performance recovery under limited observations and different levels of supervision. This protocol also enables fair comparisons across methods without confounding effects from user–decoder interactions. While closed-loop experiments are needed to assess real-time performance and interactive adaptation, our results establish an important foundation by demonstrating data-efficient sequential adaptation, a prerequisite for practical closed-loop BCI systems.

Q4 Drift-state dimensions

We clarify that 128 dimensions do not consistently outperform 64 dimensions, particularly under unsupervised adaptation as shown in Appendix Table 14 and 15. This reflects a trade-off between transformation flexibility and estimation reliability. Under supervised adaptation, behavioral labels provide strong optimization signals, allowing higher-dimensional drift states to benefit from greater flexibility. In contrast, unsupervised adaptation relies on distribution statistics estimated from limited trials. Increasing the drift dimension introduces more degrees of freedom without additional information to constrain them, making adaptation more sensitive to estimation noise.

Comments on the paper

1 comment

Dear Area Chair,

We sincerely thank all reviewers for their constructive feedback. As the discussion period concludes, we would like to summarize our core contribution and the additional evidence provided during the rebuttal.

Core contribution. DiDA addresses few-trial cross-session sequential adaptation by shifting the optimization space from high-dimensional neural transformations to a compact probabilistic drift representation. Beyond supervised and unsupervised adaptation, it further introduces a preference-guided objective that leverages lightweight user feedback to directly optimize the drift distribution when explicit behavioral labels are unavailable.

Additional evidence. During the rebuttal, we conducted extensive additional experiments, including:

  • Controlled comparisons with deterministic drift and unconstrained transformation parameterizations, supporting the benefit of probabilistic modeling beyond dimensionality reduction.
  • Additional baselines (FDA, SPINT, and Procrustes), strengthening the comparative evaluation.
  • Multiple-seed experiments and per-session performance comparisons, improving experimental reliability.
  • Adaptation across different pretraining sessions, demonstrating robustness to source-session selection.
  • Robustness evaluation under noisy preference feedback, demonstrating tolerance to imperfect preference signals.

Resolution of reviewers' concerns. These additional experiments directly address the reviewers' major concerns regarding methodological novelty, baseline comparisons, experimental reliability, and robustness. We also provided detailed clarifications on drift-state dimensionality, parameterization, and the sequential adaptation protocol.

Reviewer LcPM explicitly acknowledged that our responses and additional experiments answered their questions. We have not yet received further feedback from Reviewers H87q and 4pgy following our detailed responses.

Overall, the additional evidence strengthens our central methodological contribution and further demonstrates the effectiveness of structured probabilistic drift modeling for data-efficient cross-session adaptation. We believe these findings provide a stronger empirical foundation for evaluating the novelty and technical merits of DiDA.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.