acceptodds
Under review as a conference paper at ICLR 2027

BiXformer: A Bidirectional Cross Attention Transformer for Disentangling Inter-Regional Neural Dynamics

Abstract

Advances in high-throughput neural recording technologies enable simultaneous measurement of activity across multiple brain regions in behaving animals, producing datasets of unprecedented scale and richness. Interpreting these data remains challenging due to the bidirectional and temporally offset nature of inter-regional communication, where feedforward and feedback signals are superimposed within neural populations. We introduce BiXformer, a bidirectional cross-attention transformer that disentangles these interactions by decomposing inter-regional communication into temporally leading and lagging streams using directionally masked attention. By enforcing temporal constraints within attention heads, BiXformer recovers low-dimensional, directed latent dynamics and estimates communication delays without relying on linearity or stationarity assumptions. Validated on synthetic data with known ground-truth delays, BiXformer accurately recovers both the latent structure and inter-regional timing, outperforming existing methods in both accuracy and speed. Applied to simultaneous neural-behavioral recordings and multi-region neural recordings during a movement task, BiXformer reveals interpretable, temporally structured components consistent with the coexistence of sensory feedback and motor-related signals. These results establish BiXformer as a flexible framework for uncovering dynamic, directed communication in complex neural circuits.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Multi-region communication, Inter-regional delays, Neural population dynamics, Cross-attention transformers, Systems neuroscience

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_ayzrOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes BiXformer, a cross-attention transformer that separates inter-regional neural dynamics into leading and lagging streams using temporal attention masks and an orthogonal decoder. The model is evaluated on synthetic data for latent and delay recovery, and on real neural-behavioral and S1–M1 recordings to demonstrate its ability to identify interpretable temporal interactions.

Critical strengths
  1. Interesting approach: Separating leading and lagging interactions using masked cross-attention and orthogonal decoding is a useful contribution. The synthetic results show promising performance in recovering latent dynamics and temporal delays.

  2. Real-data applications: I appreciate the evaluation on both neural-behavioral recordings and simultaneous S1–M1 recordings. These demonstrate the potential of the method across different types of data and provide interpretable results. The presentation of these experiments is also clear.

Critical weaknesses
  1. Limited synthetic validation: One of the main motivations for BiXformer is overcoming the linearity and stationarity assumptions of existing methods such as DLAG. However, the synthetic data are generated using stationary Gaussian processes and linear mixing. I would expect experiments with nonlinear and nonstationary dynamics to demonstrate the claimed advantages. It is also unclear why DLAG performs substantially worse on data that seem compatible with its assumptions.

  2. Insufficient baseline comparisons: All baseline comparisons are limited to one synthetic dataset. Important related methods such as MR-SDS (ICLR 2024), MRDS-IR (ICLR 2025), and GLM-Transformer are not included in the experiments. Quantitative comparisons on the real datasets are also missing. These comparisons would help establish the advantages of BiXformer over existing methods.

  3. Interpretation of directionality: Both streams reconstruct the target population using past or future activity from the source population. While this identifies temporal relationships, it does not necessarily establish physiological communication direction. This is particularly concerning for the S1–M1 experiment, where the attention profiles peak at zero lag (Appendix D.6), yet the learned components are interpreted as S1→M1 and M1→S1 communication.

  4. Limited validation of the architectural constraints: The Stiefel decoder imposes orthogonality between the two streams, but it is unclear whether this assumption is appropriate for neural populations with potentially overlapping feedforward and feedback representations. The existing ablation focuses on reconstruction accuracy and removes both constraints together. It does not demonstrate how each constraint contributes to recovering the true latent streams and delays.

Questions for authors
  1. Given the zero-lag attention profiles in the S1–M1 experiment, what evidence supports interpreting the two streams as separate communication directions?

  2. Are the estimated sensory and motor delays in Section 4.2 consistent with previous physiological measurements? How sensitive are these estimates to temporal smoothing?

Update: Most of my concerns were addressed and authors performed valuable experiments during rebuttal hence I increased my score from 2 to 3 at the end of rebuttal session.

20 comments

Response to Reviewer ayzr (1/4)

We thank the reviewer for the encouraging and helpful feedback. We respond to each point below. We hope these responses address the reviewer's concerns, and that the reviewer will consider revisiting their score.

The rebuttal figures referenced below (rebuttal_figure1–5) are in the rebuttal figures folder of our anonymous repository: https://anonymous.4open.science/r/BiXformer-EB1B/


W1

One of the main motivations for BiXformer is overcoming the linearity and stationarity assumptions of existing methods such as DLAG. However, the synthetic data are generated using stationary Gaussian processes and linear mixing. I would expect experiments with nonlinear and nonstationary dynamics to demonstrate the claimed advantages.

We thank the reviewer for this feedback. To address it, we ran the model comparison on two additional synthetic datasets, which we will include in the appendix:

  • Nonlinear mixing (our synthetic dataset with ReLU mixing). This dataset uses the same GP-driven latents and ground-truth delays as our main synthetic dataset, but maps them to each population through a ReLU MLP instead of a linear projection. The latent dynamics remain linear; the relationship between the latents and the observed activity, and therefore between the two populations, is nonlinear, which violates the linear assumptions of methods such as DLAG. BiXformer achieves the highest held-out R² and the lowest delay error (see table below).

    We view nonlinear mixing (as proposed by the reviewer) as the central test of inter-regional communication methods beyond the stationary-linear regime, because the complexity and nonlinearity of the signals passed between regions is what varies across brain areas and tasks, and because DLAG and related linear-Gaussian methods cannot accommodate it. The reviewer may additionally have intended to ask about nonlinear latent dynamics, which this benchmark does not test. Benchmarking on nonlinear latent dynamics probes a separate question, the form of within-region temporal structure, that is distinct from the inter-regional communication claim of the submitted model. We expect BiXformer's architecture to extend to nonlinear latent dynamics because its attention mechanism carries no linear-Gaussian prior on latent structure. If the nonlinear mixing benchmark does not fully address the reviewer's concern and the reviewer would like an additional benchmarking dataset, we would be glad to run it.

MethodR² (mean ± std)Delay MAE (ms)
DLAG0.265 ± 0.01618.9 ± 1.1
MR-LFADS0.790 ± 0.003N/A
BiXformer0.813 ± 0.0021.8 ± 1.0
  • Non-stationary dynamics (ADM dataset; Li et al., 2025, ICML). This dataset has two regions of 50 neurons each and two across-region latents, one in each direction, and each latent's delay changes within a trial. ADM achieves a slightly higher held-out R², while BiXformer trains in half the time. We are further investigating this difference and will report our findings in the final revision.
MethodR² (mean ± std)Train time / fold (min)
ADM0.652 ± 0.0081.88 ± 0.02
BiXformer0.632 ± 0.0160.94 ± 0.03

Please let us know if these comparisons are sufficient to alleviate the reviewer's concerns, or if additional benchmarking (time permitting) is necessary.

It is also unclear why DLAG performs substantially worse on data that seem compatible with its assumptions.

Regarding DLAG's poor performance, the delays and GP timescales in our synthetic dataset were sampled independently from a uniform distribution over a biologically plausible range; we will revise Appendix B.1 to state this explicitly. As a result, some of our cross-area latents had a delay/timescale ratio of less than or equal to 1. Under DLAG's squared exponential kernel, the cross-area covariance in this regime changes little between zero lag and the true delay, so the likelihood is nearly flat around the true delay and EM has little gradient to recover it. Although DLAG's log-likelihood and delay estimates converged in all folds (rebuttal_figure3), this flatness limits the precision of its delay estimates and, therefore, its prediction of the target population.

Response to Reviewer ayzr (2/4)

W2

All baseline comparisons are limited to one synthetic dataset. Important related methods such as MR-SDS (ICLR 2024), MRDS-IR (ICLR 2025), and GLM-Transformer are not included in the experiments.

  • We thank the reviewer for these references. We will revise the Related Work section to discuss MR-SDS (Karniol-Tambour et al., 2024) and MRDS-IR (Dowling & Savin, 2025). The methods in our original comparison were chosen to span the two properties BiXformer combines: DLAG, which estimates delays explicitly but assumes linear and stationary interactions, and MR-LFADS, which captures nonlinear dynamics but does not estimate delays.
  • Following the reviewer's suggestion, we have added MRDS-IR (Dowling & Savin, 2025) to the model comparison (see table below). Although MRDS-IR does not explicitly model time delays, a delay estimate can be read post hoc from the peak of its learned cross-region impulse response filters. In both directions (X → Y and Y → X) and in all five cross-validation folds, these filters peaked at 5 ms. We report the resulting delay MAE in the table, noting that this readout is ours and that MRDS-IR was not designed to explicitly model delays. Like MR-LFADS, MRDS-IR is a recurrent state-space model, which likely explains its longer training time relative to BiXformer. The table also includes ADM (Li et al., 2025), which builds on DLAG's time-delay kernel by allowing delays to vary within a trial, and reduces training time by approximating each GP latent as a state-space model. The comparison now also spans a nonlinear and a time-varying dataset (see our response to W1). We will add this table to the appendix of the revised manuscript, together with a full description of each model's training procedure and hyperparameters. No model's hyperparameters, including BiXformer's, were tuned for these comparisons.
  • We did not include MR-SDS or GLM-Transformer, for the following reasons. MR-SDS adds switching between discrete dynamical regimes to multi-region nonlinear latent dynamics. Our main synthetic datasets contain no contextual changes that would engage this switching, so we expect MR-SDS to perform similarly to the nonlinear dynamical models already in the comparison (MRDS-IR and MR-LFADS). GLM-Transformer models cross-area communication through a stationary GLM coupling term, a property shared with methods already in our comparison (DLAG, MRDS-IR). It also uses a Poisson likelihood on spike counts, so it cannot be applied to our continuous-valued synthetic data without modification. We are happy to run either method if the reviewer would find it useful, although given the limited time and the reasons above, we hope the current comparison is sufficient.
MethodR² (mean ± std)Delay MAE (ms)Train time / fold (min)
RRR (r=4)0.341 ± 0.005N/AN/A
DLAG0.590 ± 0.07413.5 ± 8.993.9 ± 3.9
ADM0.591 ± 0.01719.3 ± 1.214.4 ± 0.2
MRDS-IR0.842 ± 0.00130.0 ± 0.050.72 ± 6.15
MR-LFADS0.865 ± 0.001N/A78.82 ± 2.85
BiXformer0.887 ± 0.0181.0 ± 2.00.44 ± 0.01

Quantitative comparisons on the real datasets are also missing.

  • We thank the reviewer for this suggestion. Since the real datasets have no ground truth for latents or delays, we compared predictive performance: we applied MR-LFADS to the neural–behavioral dataset and report its held-out R² and training time alongside BiXformer (table below). BiXformer achieved a higher held-out R² and trained far faster.
  • We limited this to one method because MR-LFADS and the other methods take far longer to train than BiXformer (see the synthetic-data comparison table), which makes running all of them infeasible within the rebuttal period. These models are also sensitive to hyperparameters, and we did not run an extensive search, so the MR-LFADS R² is likely an underestimate. Please let us know whether this is sufficient.
MethodR² (mean ± std)Train time / fold (min)
MR-LFADS0.434 ± 0.041224.8 ± 10.6
BiXformer0.656 ± 0.0131.40 ± 0.04

Response to Reviewer ayzr (3/4)

W3 and Q1

Both streams reconstruct the target population using past or future activity from the source population. While this identifies temporal relationships, it does not necessarily establish physiological communication direction. This is particularly concerning for the S1–M1 experiment, where the attention profiles peak at zero lag (Appendix D.6), yet the learned components are interpreted as S1→M1 and M1→S1 communication.

Given the zero-lag attention profiles in the S1–M1 experiment, what evidence supports interpreting the two streams as separate communication directions?

  • We thank the reviewer for raising this point, and we agree: the two streams capture components of the target population explained by past versus future source activity, and this temporal ordering alone does not establish physiological communication direction. This holds for all methods that infer communication from observational recordings; establishing direct coupling would require perturbing specific pathways, which we view as important future work.
  • To address the zero-lag concern, we retrained BiXformer with strictly directional masks (zero lag excluded from both streams) and compared the results with the original model (rebuttal_figure4). Held-out R² was essentially unchanged, and the PC1 time course of each stream was nearly identical between the two models. The decomposition therefore does not depend on the shared zero-lag bin.
  • We also performed a condition-matched trial-shuffled control on the neural–behavioral dataset, pairing behavior from each trial with neural activity from a different trial of the same condition. Both reconstruction and directional structure collapse under the shuffle (rebuttal_figure1), indicating that the recovered streams depend on trial-specific correspondence between the two signals and not on task-locked responses alone. This provides further support for a communication channel.
  • In light of this, the model's design and these controls support interpreting the two streams as separate communication channels. This does not unequivocally prove that communication is present, but it indicates a putative communication channel in each direction. We hope this addresses the reviewer's concern. If the reviewer would like to see additional analyses, we would be happy to include them.
Reviewer_cmArOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

BiXformer reconstructs one neural population from another using parallel past- and future-masked cross-attention streams. Rotary positional encoding supports lag-sensitive attention, while orthogonal decoders separate the streams into distinct output subspaces. Attention-profile peaks are interpreted as communication delays. Experiments include a synthetic GP dataset, neural–behavioral recordings from the medulla, and simultaneous S1–M1 recordings during licking.

Critical strengths
  1. The architecture provides an explicit, inspectable decomposition through directional attention masks and orthogonal decoders. Individual heads expose lag preferences, and the reported synthetic-data training time is under one minute per fold.
  2. BiXformer accurately recovers latent trajectories and delays in the tested synthetic setting. The paper also reports neutral or unfavorable findings transparently, including untuned MR-LFADS performance, zero-lag S1–M1 profiles, and similar constrained/unconstrained reconstruction accuracy.
  3. The applications combine rich neural and behavioral recordings, including medulla activity with high-speed kinematics and simultaneous S1–M1 recordings. These provide relevant settings for generating sensorimotor hypotheses, and the manuscript supplies implementation details and a code link.
Critical weaknesses
  1. Common inputs and lick-locked responses can produce lead/lag signatures without inter-regional communication. No condition-matched trial-shuffle control is reported. Rhythmic behavior and acausal smoothing further complicate interpretation, so the streams are not yet established as distinct motor and sensory communication pathways.

  2. Identifiability remains insufficiently demonstrated. Both masks admit zero lag, and decoder orthogonality imposes separation rather than validating its biological meaning. The combined unconstrained ablation evaluates reconstruction alone, while latent/subspace recovery and decomposition stability across initializations remain untested.

  3. The quantitative benchmark uses one linear, stationary GP setting, leaving nonlinear and time-varying claims untested. A lagged linear baseline is absent, MR-LFADS is untuned, and DLAG convergence is insufficiently documented. The CPU/GPU runtime comparison also limits conclusions about architectural efficiency.

  4. The S1–M1 conclusions rely mainly on qualitative examples despite all attention profiles peaking at zero lag. Held-out reconstruction performance, trial/session counts, and quantitative support for stream-specific findings are missing. The neural– behavioral GMM modes also require stronger validation before being interpreted as distinct communication channels.

Questions for authors
  1. Please report condition-matched trial-shuffled controls that preserve task-locked responses while removing trial-specific correspondence. How much reconstruction and apparent directional structure survives? Common-drive simulations with unequal observation delays would also help assess whether BiXformer reports bidirectional communication without direct coupling.
  2. How sensitive are the delays to unsmoothed and causally smoothed spike counts, accounting for known filter delays? For rhythmic licking, how do you distinguish competing lag explanations separated by approximately one lick cycle?
  3. How is zero-lag variance allocated between streams, and are the learned subspaces reproducible across initializations and splits? Please clarify head-selection rules and how the 16-dimensional latents per stream were aligned with the four ground-truth latents, including whether alignment was fitted using training data only.
  4. Please include a lagged linear baseline and document DLAG initialization, delay bounds, EM convergence, and source-only held-out prediction. A better-tuned MR-LFADS comparison and synthetic tests with nonlinear or time-varying coupling would strengthen the claimed comparative advantages.
  5. For S1–M1, please report held-out R², session/trial counts, and quantitative evidence supporting stream-specific interpretations despite zero-lag peaks. For the neural–behavioral analysis, please assess uncertainty in the pooled GMM modes and clarify whether reported means denote delay magnitudes or signed offsets.
11 comments

Response to Reviewer cmAr (1/3)

We thank the reviewer for the detailed and helpful feedback. We hope these responses address the reviewer's concerns, and that the reviewer will consider revisiting their score.

The rebuttal figures (rebuttal_figure1–5) are in our anonymous repository: https://anonymous.4open.science/r/BiXformer-EB1B/

Q1

Please report condition-matched trial-shuffled controls that preserve task-locked responses while removing trial-specific correspondence. How much reconstruction and apparent directional structure survives?

Thanks for this suggestion. We performed a condition-matched trial-shuffled control on the neural–behavioral dataset, pairing behavior from each trial with neural activity from a different trial of the same condition. Both reconstruction and directional structure collapse under the shuffle (rebuttal_figure1): held-out R² drops from 0.653 ± 0.018 to 0.166 ± 0.028, and the delay distributions within each stream lose the distinct peaks seen in the original data (Fig. 4e). If the reviewer would like this added as an appendix to the revised manuscript, we are happy to do that.

Common-drive simulations with unequal observation delays would also help assess whether BiXformer reports bidirectional communication without direct coupling.

We thank the reviewer for raising an important confound. The trial-shuffle control above rules out common input mediated by condition-average task structure. What it cannot rule out is common input from an unobserved region that fluctuates trial by trial, because such common input is destroyed by the shuffle in the same way direct coupling is. This residual limitation is intrinsic to inferring communication from observational recordings and is not specific to BiXformer; other methods (e.g., DLAG, MR-LFADS) face the same ambiguity, because an unobserved source can only be ruled out by recording from it or perturbing it. We will expand the Limitations section to discuss this. Future work will apply BiXformer to datasets with perturbations, which would allow us to test for direct coupling.

Q2

How sensitive are the delays to unsmoothed and causally smoothed spike counts, accounting for known filter delays?

To address these concerns, we will apply BiXformer to the unsmoothed data and repeat the analysis in Fig. 4e, reporting the attention heads' diagonal profiles and delay distributions alongside the original results. We will also investigate this in synthetic data with different filtering profiles. If the reviewer would like to see additional analyses, we would be happy to include them.

For rhythmic licking, how do you distinguish competing lag explanations separated by approximately one lick cycle?

Inter-lick intervals in mice are generally from 100–200 ms (Boughter et al., 2007), and all reported delays here are about 50 ms or less. In addition, variability in inter-lick intervals within trials blurs aliased peaks while leaving true delays sharp, which is consistent with the well-localized attention profiles we observe (Figs. 9–10).

Q3

How is zero-lag variance allocated between streams?

Because both attention masks include the zero-lag bin, any shared variance at zero lag is in principle accessible to both streams, and its allocation between them is not constrained by the model. In the synthetic dataset, however, there is no shared zero-lag variance to allocate: all cross-area latents have non-zero delays.

and are the learned subspaces reproducible across initializations and splits?

On the synthetic dataset, we computed the correlation between recovered and ground-truth latents across 5 folds and 5 random initializations (n = 25 runs). The average correlation across all four latents was 0.89 ± 0.02 (mean ± s.d. across runs), and the correlations for each individual latent are reported in rebuttal_figure2. If the reviewer would like this added as an appendix to the revised manuscript, we are happy to do that. Note that these folds and random initializations differ from those used in Fig. 3, so the values are not expected to match that figure exactly.

Please clarify head-selection rules and how the 16-dimensional latents per stream were aligned with the four ground-truth latents, including whether alignment was fitted using training data only.

Regarding head selection, we used Hungarian matching to pair the estimated delays from the attention heads with the ground-truth delays (Table 2). Regarding latent alignment, each ground-truth latent was decoded post hoc from the corresponding stream's 16-dimensional latents using ordinary least-squares regression, and the decoded trace was temporally aligned to the ground truth via sliding cross-correlation. The readout is fitted on the training data, and the reported correlations are evaluated on the held-out test data. We will revise the manuscript to describe both procedures in detail (Appendix A).

Response to Reviewer cmAr (2/3)

W3 and Q4

A lagged linear baseline is absent, MR-LFADS is untuned, and DLAG convergence is insufficiently documented. The CPU/GPU runtime comparison also limits conclusions about architectural efficiency.

Please include a lagged linear baseline and document DLAG initialization, delay bounds, EM convergence, and source-only held-out prediction. A better-tuned MR-LFADS comparison and synthetic tests with nonlinear or time-varying coupling would strengthen the claimed comparative advantages.

  • Regarding the lagged linear baseline, DLAG models linear interactions between the two populations with explicit time delays, and it is already included in our model comparison (Table 2). We therefore see it as serving this role. If the reviewer had something else in mind, please let us know and we would be happy to include it.
  • We will expand Appendix C to document DLAG's fitting procedure. We used DLAG's default pCCA initialization and default delay bound (|δ| ≤ 500 ms). Across all folds, the log-likelihood and delay estimates plateaued well before the 5000-iteration limit (rebuttal_figure3). We also clarify that the R² reported for all methods in Table 2 is a held-out, source-only prediction of population Y from population X.
  • We revisited the MR-LFADS comparison and identified the source of its poor performance. In our original setup, the model had access to Y's activity during training but not during inference, so it was evaluated under conditions different from those it was trained on. We have now zeroed out Y's activity during training as well, which matches the evaluation setting and provides a fair comparison with BiXformer. MR-LFADS now achieves a held-out R² of 0.865 ± 0.001 (Table 2 will be updated accordingly).
  • Regarding runtime, we agree that the comparison with DLAG reflects practical runtime on each method's standard implementation rather than architectural efficiency alone. MR-LFADS and BiXformer, however, were trained on the same GPUs, so their comparison more directly reflects architectural efficiency. We will revise the text to clarify this (Appendix C.1). If the reviewer has a specific request concerning the most appropriate way to reflect differences in runtime, we are happy to consider it. We think our approach is the most honest to the constraints of each method.

The quantitative benchmark uses one linear, stationary GP setting, leaving nonlinear and time-varying claims untested.

Thanks for the helpful suggestion. We ran the model comparison on two additional synthetic datasets, which we will include in the appendix:

  • Nonlinear coupling (our synthetic dataset with ReLU mixing). This dataset uses the same GP-driven latents and ground-truth delays as our main synthetic dataset, but maps them to each population through a ReLU MLP instead of a linear projection, so the relationship between the two populations' activity is nonlinear. BiXformer achieves the highest held-out R² and the lowest delay error.
MethodR² (mean ± std)Delay MAE (ms)
DLAG0.265 ± 0.01618.9 ± 1.1
MR-LFADS0.790 ± 0.003N/A
BiXformer0.813 ± 0.0021.8 ± 1.0
  • Time-varying coupling (ADM dataset; Li et al., 2025, ICML). This dataset has two regions of 50 neurons each and two across-region latents, one in each direction, and each latent's delay changes within a trial. ADM achieves a slightly higher held-out R², while BiXformer trains in half the time. We are further investigating this difference and will report our findings in the final revision.
MethodR² (mean ± std)Train time / fold (min)
ADM0.652 ± 0.0081.88 ± 0.02
BiXformer0.632 ± 0.0160.94 ± 0.03

Please let us know if these comparisons are sufficient to alleviate the reviewer's concerns, or if additional benchmarking (time permitting) is necessary. We hope this addresses the reviewer's concern.

Response to Reviewer cmAr (3/3)

Q5

For S1–M1, please report held-out R², session/trial counts, and quantitative evidence supporting stream-specific interpretations despite zero-lag peaks.

We thank the reviewer for these suggestions. We will report held-out R² for the S1–M1 experiment in the main text and session/trial counts in Appendix B.3. To address the zero-lag concern, we retrained BiXformer with strictly directional masks (zero lag excluded from both streams) and compared the results with the original model (rebuttal_figure4). Held-out R² was essentially unchanged, and the PC1 time course of each stream was nearly identical between the two models. The decomposition therefore does not depend on the shared zero-lag bin. We hope this addresses the reviewer's concern. If the reviewer would like to see additional analyses, we would be happy to include them.

For the neural–behavioral analysis, please assess uncertainty in the pooled GMM modes and clarify whether reported means denote delay magnitudes or signed offsets.

For the neural–behavioral GMM, we re-ran the 100 random train–test splits and now report 95% bootstrap confidence intervals for the component means, SD, and weight of each GMM mode (table below). In each of 1000 iterations we resampled the 100 train–test splits with replacement, keeping both attention heads of a split together, and refit the GMM. We will revise the manuscript accordingly and now report these confidence intervals. Regarding the reported means, we will revise the manuscript so that all delays are reported as signed offsets (negative: behavior (X) leads neural activity (Y); positive: neural activity (Y) leads behavior (X)). We hope this addresses the reviewer's concern; if further analyses would be helpful, please let us know.

StreamGMM modeMean (ms)SD (ms)Weight
Sensory (X→Y)1−14.2 [−14.9, −12.5]9.0 [7.0, 10.0]0.79 [0.72, 0.83]
Sensory (X→Y)2−53.2 [−53.8, −49.6]2.0 [1.4, 6.7]0.21 [0.17, 0.28]
Motor (Y→X)118.8 [18.4, 19.3]2.4 [2.1, 2.7]0.50 [0.50, 0.50]
Motor (Y→X)242.0 [40.9, 43.1]5.7 [4.6, 6.9]0.50 [0.50, 0.50]

Values are estimates with 95% bootstrap confidence intervals in brackets.

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

In this paper, the authors proposed a framework named BiXformer "A Bidirectional Cross Attention Transformer", which utilizes the transformer architecture that uses masked cross-attention, and disentangled neural latnet subspaces to separate temporally component dynamicss in inter-region neural recordings. Extensive experiments on synthetic datasets and mouse neural-behavioral and S1–M1 recordings demonstrate a good recovery of neural latent dynamics and temporal components.

Critical strengths
  1. The paper is well-written and has a concrete flow for readers to follow.
  2. The proposed methodology and the Section 3 method is clear and easy-to understand.
  3. The experiments in this work is abundant and extensive, which are evaluateed on synthetic dataset, and also neural-behavioral recordings and multi-region neural recordings.
Critical weaknesses
  1. I agree that Transformer blocks are detinitely powerful deep learning architectures, and some work has also adopted them in decoding neural data. But I am a bit wondering that whether the use of Transformer in this inter-regional neural dynamics understanding problem, may lead the problem and task setting it be a bit more block-box (I might be wrong).
  2. The related work comparison of this project could be expanded to better contextualize the interpretation of the recovery of inter-region latent dynamics. Two relevant works not currently discussed are ADM [1] and MRM-GP [2]. These approaches model time-varying communication delays and directional interactions with more interpretable Gaussian Process-based methods.

[1] Learning Time-Varying Multi-Region Brain Communications via Scalable Markovian Gaussian Processes, ICML 2025.

[2] Multi-Region Markovian Gaussian Process: An Efficient Method to Discover Directional Communications Across Multiple Brain Regions, ICML 2024.

Questions for authors

I have no more questions here for affecting my assessment. For my other questions, please refer to my Weaknesses section to see my other concerns.

1 comment

Thank you again for your feedback. The rebuttal revision is now submitted.

W2 (ADM and MRM-GP). Both methods are now discussed in the Related Work (Section 2.1). ADM is benchmarked against BiXformer on our synthetic dataset (Table 2, Appendix C) and on a time-varying delay benchmark following the ADM paper (Table 5, Appendix C.3).

We hope these revisions address your concerns, and that you will consider revisiting your score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the rebuttal and revision

We thank the Area Chair and all three reviewers for their time and for the valuable feedback and suggestions, which have strengthened this paper.

Strengths noted by the reviewers. We were encouraged that the reviewers found the leading/lagging decomposition through masked cross-attention a useful contribution, with an explicit and inspectable architecture. The reviewers also noted the accurate recovery of latents and delays on synthetic data, described the experiments as extensive and rich, and found the presentation clear.

Concerns and how they were addressed. The rebuttal revision includes all analyses discussed with the reviewers.

  • Nonlinearity and non-stationarity claims were untested (ayzr, cmAr). We added a nonlinear-mixing benchmark and a time-varying delay benchmark, and scoped the claims accordingly: the nonlinearity claim now refers to nonlinear interactions between populations, and the non-stationarity claims are removed.
  • Baselines and related work were limited (gFrt, ayzr, cmAr). We added ADM and MRDS-IR to the model comparison (Table 2), corrected the MR-LFADS comparison, added a comparison on real data, and documented the DLAG fitting procedure. In addition, we've revised the related work section to include MRM-GP, ADM, MR-SDS and MRDS-IR.
  • Lead/lag structure may not reflect communication direction (ayzr, cmAr). We added a condition-matched trial-shuffle control and a strict-mask control for S1–M1. The two streams are now described as leading and lagging predictive components, with the physiological directions presented as hypotheses.
  • Identifiability and architectural constraints (ayzr, cmAr). We report latent recovery across folds and initializations, and ablate the masks and the Stiefel constraint independently, on both real and synthetic data.
  • Robustness and interpretation of the delay estimates (ayzr, cmAr). Delays are now reported on unsmoothed data with bootstrap confidence intervals, together with physiological context and the relation to inter-lick intervals.
  • Missing details (cmAr). We added S1–M1 reconstruction metrics and session and trial counts, described head selection and latent alignment, and clarified the runtime comparison between models.

Limitations. The new analyses also clarified limitations, which the revision states openly. BiXformer does not track within-trial delay changes, and we propose future directions that could address this. In the neural–behavioral application, the longer sensory lag does not persist without smoothing, so we now report the delays estimated on unsmoothed data (Fig. 4e). We also revised the Limitations section to explicitly address the common input confound, which cannot be ruled out from observational recordings.

Outcome of the discussion. Reviewer ayzr raised their rating from 2 to 3 and indicated that the expanded benchmarks, controls and ablations address their main concerns. Reviewer cmAr indicated that most points were largely or substantially addressed, and the revision includes the remaining items requested.

We thank everyone again for a constructive discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.