acceptodds
Under review as a conference paper at ICLR 2027

Resolving Latent Representation Collapse in Schrödinger Bridges for Multi-Task Inference and Calibrated Forecasting of Panel Time Series: An Application to Astrophysics

Abstract

Conditional generative modeling via paired Schrödinger Bridges faces severe limitations in sequential forecasting of sparse, irregularly sampled panel time series. In streaming contexts – such as real-time multi-band supernova light curve monitoring – continuous, noisy non-detections during the critical pre-peak phase creates an asymmetric many-to-one mapping deficiency. This sparse conditioning triggers latent representation collapse, leading to miscalibrated uncertainty intervals and severe classification prior saturation toward dominant classes. In this work, we present a framework that maps panel time series into a continuous hidden space, advancing a hidden-layer dynamic variance Image-to-Image Schrödinger Bridge (I2SB) solver to model extension intervals without data-level noise corruption. We solve the uncertainty quantification task by generating parametric prediction intervals derived directly from the hidden-space I2SB trajectories paired with a post-hoc fine-tuned UQNet uncertainty network, cleanly decoupling trajectory synthesis from downstream tasks. Crucially, categorical classification is resolved via a novel pre-softmax logit fusion technique at the terminal boundary. This architecture ensembles an abstract latent-space energy-based model prior with a standalone temporal neural network trained directly on uncompressed temporal flux trajectories using class-explicit inverse-frequency weights. Concurrently, light-curve-driven redshift estimation operates as a completely independent, parallel channel. Evaluated on highly imbalanced astrophysical benchmarks, our joint pipeline significantly improves early-stage calibration – substantially improving empirical coverage in the pre-peak regime – while maximizing macro classification F1-score and reducing prior saturation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Uncertainty Quantification, Schrödinger Bridges, Energy-Based Model, Panel Time Series, Calibrated Forecasting, Multi-Task Inference, AI for Science, Astrophysics

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_Lby1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors proposed an SB method operated in latent space for forecasting of irregularly sampled timeseries in supernovae data. They used SELDON to handle the irregularity and operate most of the probabilistic treatments in the latent space. They used an SB to transport partial LC encodings to full LC encodings, and a separate UQnet for uncertainty quantification. For classification task they poposed a pre-softmax mixing for better calibration.

Critical strengths
  • Operating probablistic inference part in latent space is a good idea and SB seems somewhat natural for the partial to full task.
  • The authors made good attempt to make sure the inference is calibrated rather than simply being able to generate similar data.
Critical weaknesses
  • I think the advantage of doing SB is not well explaind esp. compare to conditional sampling.
  • Calibration claims are less supported than point performance.
Questions for authors
  • Advantage of SB in place or a conditional diffusion/flow method. The partial to full in SELDON space is done by learning a direct coupling that moves partial representations to full representations.
    • There are some computation/learning costs associated with the generality of SB compared to e.g., diffusion which has a known forward process. And the transportation is on the population scale rather than trailered towards any single event. (i.e., the transportation will transport distribution of partial LC's representation to full LC's representation, which is different than trying to get posterior's given any single event).
    • Is there a particular advantage so the authors uses SB rather than a conditional sampling method conditioning on partial and sample full? This also closer to a Bayesian inference in the latent space.
    • The many to one issue seems can be solved more naturally with conditional sampling than learning a transport?
  • Re: Temporal completeness. Will this give the method some look ahead information? Suppose we have a new event can we get it before we observe the full LC?
  • Ordering of averaging and decoding in point forecast. The decoding process is non-linear so it does not commute with expectations. So for point forecasting we at least have two options, 1) average the representation, then decode as the authors did or 2) decode all sampled representation, then average. Is the second tried?
  • Testing of calibration: the authors claimed that SB's uncertainty is not enough thus a separate UQNet is needed. I am not sure I fully follow if there is an empirical test for this. Please correct me if I missed it.
  • Pre-softmax mixing: I assume the hold out data is not used for choosing the 0.85. The authors said it is using a validation data but I am unsure about the relationship between this validation dataset, training set and hold out test set. Also is this tested for claimed better calibration?
3 comments

We sincerely thank the reviewer for their thoughtful feedback, constructive questions, and Weak Accept recommendation. We are glad you found latent-space Schrödinger Bridge transport natural for partial-to-full forecasting, and we also appreciated the connection to rapidity addition in special relativity and hyperbolic geometry.

Below, we address each technical question.

1. Advantage of I2SB Over Conditional Diffusion / Flow Matching

You raise an important distinction between population-level transport and event-level conditional sampling p(xfull∣xpartial)p(x_{\mathrm{full}}\mid x_{\mathrm{partial}}).

Why I2SB for paired partial-to-full completion. Unlike general Schrödinger Bridge formulations requiring iterative proportional fitting between unpaired marginals, I2SB trains on explicitly paired partial/full representations from the same object using an analytically tractable Brownian-bridge process. Each reverse trajectory starts directly from xpartialx_{\mathrm{partial}}, allowing multiple samples for the same partial observation without iterative marginal matching.

Direct comparison with Conditional Flow Matching (CFM). We additionally trained a conditional CFM baseline in the same frozen 256-dimensional SELDON hidden space, with

x0∼N(0,I256)x_0\sim\mathcal{N}(0,I_{256})

transported toward xfullx_{\mathrm{full}} while conditioning on xpartialx_{\mathrm{partial}}.

  • Few-step forecasting: I2SB maintains strong forecasting accuracy at NFE=5NFE=5, with pre-peak NRMSE 0.1895, compared with 0.1953 for CFM at NFE=5NFE=5 and 0.1946 for CFM at NFE=20NFE=20. CFM remains faster at matched NFE (1.13 s vs. 2.33 s at NFE=5NFE=5).
  • Early-stage classification and calibration: I2SB improves Macro-F1 and ECE in both early completeness regimes. At 0–20%, Macro-F1 improves from 22.36% to 27.42% and ECE from 7.01% to 4.75%. At 20–40%, Macro-F1 improves from 27.52% to 31.22% and ECE from 7.51% to 2.80%. Accuracy improves at 0–20% (40.34% →\rightarrow 41.56%) and is essentially unchanged at 20–40% (44.38% vs. 44.35%).
  • Late-stage sharpness: At 80–100% completeness, I2SB produces narrower bridge-only intervals (6.19 vs. 6.81) with comparable coverage (8.50% vs. 8.44%).

These results suggest that CFM is a strong alternative rather than a uniformly weaker baseline, while endpoint-paired I2SB provides favorable behavior in several sparse and low-NFE operating regimes.

2. Temporal Completeness rir_i Without Look-Ahead Information

Thank you for flagging this issue. Your question prompted us to review our data-loading pipeline, where we found that the original implementation inadvertently read a per-object cached tmax⁡t_{\max} rather than a fixed global value.

We corrected the protocol so that temporal completeness uses no future information at inference. For a new streaming event,

\operatorname{clip} \left( \frac{t_{\mathrm{last},i}-t_{\min}} {t_{\max}-t_{\min}}, 0,1 \right)$$ with $t_{\min}=0$ and a fixed $t_{\max}=221.61\ \text{days}$ defined as the 95th percentile of completed **training-set** light-curve durations. Thus, $r_i$ depends only on the currently observed timestamp $t_{\mathrm{last},i}$. Using this corrected definition, we reran the affected pipeline: - **Three-seed retraining:** I2SB, UQNet, and the temporal classifier were retrained across all three seeds. - **Validation-only fusion selection:** The optimal fusion weight shifted from $\alpha=0.825$ to $\alpha=0.800$. - **Updated results:** Numerical values changed slightly, while the main qualitative trends, UQNet calibration gains, early-stage separation results, and baseline comparisons remained consistent. We will clarify this inference protocol and update the corresponding results in the manuscript.

3. Empirical Evidence That Bridge Uncertainty Alone Is Insufficient

We apologize if this connection was unclear in the text. The direct empirical test comparing raw I2SB uncertainty against I2SB + UQNet is shown in Table 3 and Table 4:

  • Table 3 (UQNet is not used): Even with all bridge variance components active (Config D, which includes completeness-dependent variance and the flux-error floor), raw I2SB trajectories substantially under-cover at the 0.80 nominal level, achieving only 0.305 ± 0.000 all-stage coverage and 0.493 ± 0.007 pre-peak coverage (largely because held-out points in the tails consist of noisy non-detections and the hidden-space bridge does not model observation-level heteroscedastic noise).
  • Table 4 (I2SB + UQNet): Passing the hidden-space bridge intervals into our decoupled post-hoc UQNet substantially reduces this under-coverage, lifting empirical coverage at the 80% nominal level from 30.53% to 81.04% (all-stage) and from 49.34% to 74.16% (pre-peak), while leaving the I2SB point forecast unchanged.

4. Ordering of Averaging and Nonlinear Decoding

You are correct that the SELDON decoder D(⋅)D(\cdot) is nonlinear, so in general D(E[xfull])≠E[D(xfull)]D(\mathbb{E}[x_{\mathrm{full}}]) \neq \mathbb{E}[D(x_{\mathrm{full}})].

Uncertainty quantification. For prediction intervals and coverage metrics, we already decode every sampled trajectory individually, D(xfull(k)),k=1,…,KD(x_{\mathrm{full}}^{(k)}), \quad k=1,\ldots,K

and compute empirical quantiles directly in flux space, preserving nonlinear decoder effects.

Point forecasting. We additionally compared the two estimators directly:

EstimatorNRMSEHeld-out RMSELatency/objectThroughputIncremental GPU memory
D(E[x_full])0.26070 ± 0.0027647.96 ± 1.547.38 ms135.5 objects/s12.5 MB
E[D(x_full)]0.25961 ± 0.0024447.35 ± 1.0015.34 ms65.2 objects/s3,206.8 MB

The sample-then-decode estimator gives a small accuracy improvement, whereas decoding the mean representation requires only one decoder pass and is substantially more efficient. We therefore retain D(E[xfull])D(\mathbb{E}[x_{\mathrm{full}}]) for the main point-forecast experiments while using per-sample decoding for all uncertainty estimates. The two estimators achieve nearly identical forecasting accuracy, while D(E[xfull])D(\mathbb{E}[x_{\mathrm{full}}]) roughly halves latency and reduces GPU memory by over 250×.

5. Validation Split Hygiene and Pre-Softmax Mixing (α=0.800\alpha=0.800)

Disjoint splits. The dataset is split at the astrophysical-object level into disjoint training (80%), validation (10%), and held-out test (10%) sets, preventing object or partial-view overlap across splits. The fusion weight α=0.800\alpha=0.800 was selected exclusively on validation pre-peak Macro-F1 and frozen before test evaluation.

Calibration Impact of Pre-Softmax Mixing. Fusion substantially improves calibration relative to the energy-only branch in the pre-peak regime. On the held-out test set, pre-peak ECE decreases from 30.68% at α=0\alpha=0 to 3.82% at α=0.800\alpha=0.800. Model B alone (α=1\alpha=1) achieves a similar ECE of 3.85%. Thus, we interpret fusion as combining complementary signals and substantially reducing the poor calibration of the energy-only classifier, rather than claiming uniform superiority over both single-branch models on every calibration metric.

We will clarify both the validation/test separation and the calibration interpretation in the manuscript.

Thank the authors for the detailed response. It addresses most of my concerns. I would usually increase score once my questions are addressed but this time we only have 4 possible scores and I think general ML contribution-wise the current paper is shy of the top score of 4 so I will keep my 3. To me the paper deserves a raise, just not to 4.

Reviewer_TVi3OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper aims to predict trajectories of supervnovae observations from partially observed data, particularly focussing on pre-peak data. The existing trained classifiers suffer from latent representation collapse under sparse observations that the authors seek to solve by augmenting existing procedures with a Schrodinger Bridge formalism

Critical strengths

The paper builds on recent prior work for solving a topical science problem with strong motivation from the upcoming LSST experimental data. The application is an interesting potential target for scientific machine learning.

Critical weaknesses

For an applied physical-sciences submission at ICLR, I would expect either (1) a substantial methodological contribution of interest to the ML community, or (2) evidence that the application of ML enables a qualitatively new scientific capability.

On (1), the presentation is of a bespoke pipeline with many components, building on existing pretrained work, with little clear isolation of the interesting challenges being solved. The individual architectural choices are not sufficiently motivated or compared against simpler alternatives, making it difficult to identify the methodological contribution. The exposition in general is very unclear and hard to read. I think significant work is needed on the presentation, experimental comparisons, and positioning to make this of interest to ICLR.

On (2), I have little sense of what the headline scientific result or new capability is supposed to be. While the paper reports improvements across several metrics, the scientific significance of these improvements is not established. The paper does not even include a conclusion or discussion of limitations, which makes it particularly difficult to understand what generalizable methodological insights or headline scientific results the authors intend the reader to take away.

Questions for authors

The paper doesn’t have a conclusion or discussion, the paper would greatly benefit from a clear exposition of the strengths and weaknesses of what has been proposed

The paper claims that the astrophysical dataset is primarily for illustrative purposes and the framework is directly transferable to other contexts. I think in this case either a second example or even a synthetic example to illustrate what each component is doing would greatly help the paper. For example, without significant prior knowledge of what papers like SELDON are doing, I can’t see how a framework so coupled to this domain specific solution does generalise

Could the authors also provide comparisons with simpler methods for mapping partial observations to complete representations? In particular, why is a Schrödinger Bridge necessary rather than a standard supervised regression model or alternative conditional generative model.

The paper is quite hard to follow. Phrases such as "asymmetric many-to-one mapping deficiency", "latent representation collapse" and "prior saturation" are introduced without convincing demonstrations of the underlying problems. Most importantly, I don't see evidence that representation collapse has actually been diagnosed, and it instead just asserted. This paper could benefit from a considered rewrite especially if the target is a general ML conference like ICLR.

2 comments

We sincerely thank the reviewer for the thorough and candid assessment. We agree that the original manuscript did not sufficiently isolate the core machine-learning contribution or empirically demonstrate several terms used in our framing. We therefore added direct diagnostics and alternative baselines, as summarized below.

1. Clear Definitions and Empirical Diagnosis of Core Failure Modes

A. Asymmetric Partial-to-Full Mapping

During training, progressively longer partial light curves from the same object share one complete target, whereas at inference a sparse partial observation may admit multiple plausible future completions. This motivates a probabilistic rather than purely deterministic partial-to-full mapping. We have revised the manuscript to describe this more directly and avoid overstating deterministic-regression failure without a dedicated benchmark.

B. Early Latent Representation Degradation

To directly test whether sparse observations reduce class-discriminative structure, we compute the normalized centroid separation

Scd=∥μc−μd∥2(σc2+σd2)/2,S_{cd}=\frac{\|\mu_c-\mu_d\|_2}{\sqrt{(\sigma_c^2+\sigma_d^2)/2}},

and evaluate Scd/ScdFullS_{cd}/S_{cd}^{\mathrm{Full}} across class pairs and completeness levels.

CompletenessPartial / FullI2SB / FullGain
0–20%0.6010.642 ± 0.023+0.042 ± 0.023
20–40%0.9020.924 ± 0.006+0.021 ± 0.006
40–60%0.9830.985 ± 0.002+0.002 ± 0.002
60–80%1.0000.998 ± 0.003−0.002 ± 0.003
80–100%1.0000.997 ± 0.004−0.003 ± 0.004

At 0–20% completeness, the partial representation retains only 0.601 of the full-curve class separation, while I2SB increases this to 0.642. We therefore revised our claim from “resolving latent collapse” to partially recovering early-stage class-discriminative separation.

C. Downstream Classification Under Sparsity

The dataset is strongly class-imbalanced, and classification becomes particularly difficult at low completeness. We therefore report Macro-F1 and class-wise recall in addition to accuracy. In the 0–20% regime, I2SB + fusion improves Macro-F1 from 22.36% with CFM to 27.42%, while rare SNIIb recall increases from 0.45% to 5.86%. At 20–40%, Macro-F1 increases from 27.52% to 31.22%. We now describe these results as improved balanced classification rather than elimination of class imbalance or “prior saturation.”

2. Comparison with Simpler and Alternative Baselines

To address why a Schrödinger Bridge is used instead of an alternative conditional generative model, we trained a Conditional Flow Matching (CFM) baseline in the same frozen 256-dimensional SELDON hidden space. CFM transports

x0∼N(0,I256)x_0\sim\mathcal{N}(0,I_{256})

toward xfullx_{\mathrm{full}} while conditioning on xpartialx_{\mathrm{partial}}, whereas I2SB directly pairs partial and full representations and initializes the reverse process from the partial endpoint.

The comparison shows that CFM is a strong alternative rather than uniformly weaker:

  • At 0–20% completeness, I2SB improves Macro-F1 from 22.36% to 27.42% and ECE from 7.01% to 4.75%.
  • At 20–40%, Macro-F1 improves from 27.52% to 31.22% and ECE from 7.51% to 2.80%, while accuracy is essentially unchanged.
  • At 80–100% completeness, I2SB gives narrower bridge-only intervals (6.19 vs. 6.81) at comparable coverage (8.50% vs. 8.44%).
  • At NFE=5, I2SB achieves pre-peak NRMSE 0.1895, lower than CFM at NFE=5 (0.1953) and NFE=20 (0.1946). CFM remains faster at matched NFE.

These results motivate I2SB as an endpoint-paired alternative with favorable behavior in several sparse and low-NFE regimes, rather than establishing universal superiority over conditional flow methods.

3. Isolating Individual Architectural Choices

We agree that the original manuscript did not sufficiently isolate the contribution of each component. Our ablations show distinct roles:

  • Completeness-dependent variance: Adding dynamic variance increases pre-peak 80%-nominal coverage from 0.469±0.0120.469\pm0.012 to 0.520±0.0090.520\pm0.009.
  • Observational flux-error floor: Adding the flux-error floor gives the lowest all-stage future-flux RMSE in the bridge ablation (46.92±0.2946.92\pm0.29).
  • Post-hoc UQNet: Raw bridge intervals substantially undercover. At 80% nominal coverage, raw I2SB achieves approximately 30.5% all-stage and 49.3% pre-peak coverage, whereas I2SB + UQNet increases these to approximately 81.0% and 74.2%, respectively.
  • Pre-softmax fusion: The latent-energy and decoded-curve classifiers provide complementary signals. The final α=0.800\alpha=0.800 was selected exclusively using validation pre-peak Macro-F1 before held-out test evaluation.

These ablations clarify that trajectory completion, interval calibration, and downstream classification are deliberately decoupled rather than forming an undifferentiated pipeline.

4. Scientific Capability, Scope, and Limitations

We agree that the original submission did not clearly state the scientific implication or limitations. We have therefore added dedicated Discussion, Conclusion, Limitations, and Future Work sections, summarized in our General Comment.

The main scientific motivation is earlier probabilistic forecasting and classification of transient candidates from sparse pre-peak observations. Our experiments show that I2SB partially recovers early-stage class separation, UQNet substantially improves pre-peak interval calibration, and the framework maintains strong forecasting accuracy with five reverse steps. We have revised the manuscript to present these as potential ingredients for earlier survey triage rather than claiming demonstrated autonomous rare-event identification within a fixed number of hours.

We also explicitly acknowledge three limitations:

  1. Rare classes remain difficult at very low completeness. For example, SNIIb recall is only 5.86±6.93%5.86\pm6.93\% at 0–20% completeness and 4.00±2.00%4.00\pm2.00\% at 20–40%.
  2. Nominal interval coverage requires post-hoc UQNet calibration; bridge uncertainty alone is insufficient.
  3. Performance is bounded by the frozen SELDON representation, motivating future work on joint encoder adaptation.

Finally, because our empirical evaluation uses a single astrophysical benchmark, we have removed claims of demonstrated domain-agnostic transfer and revised the manuscript to distinguish potential general applicability from empirically validated generalization.

We appreciate the reviewer’s comments, which led us to substantially strengthen the diagnostics, baseline comparisons, positioning, and limitations of the work.

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

The paper proposes to solve challenges in time series prediction for astronomical monitoring. The particular challenges are latent representation collapse due to sparse observations and noisy non-detections creating a many-to-one degeneracy. The approach should fix problems with miscalibrated uncertainty intervals and over-predicting of dominant classes. Novel contributions include a pre-softmax logit fusion for categorical classification. Evaluation is carried out on one imbalanced astrophysical benchmark. Claims are early-stage calibration, maximizing macro classification, and reducing prior saturation. Evaluation of prediction accuracy and interval coverage is performed against ablated versions of the approach and simple baselines.

Critical strengths

The framework is well-motivated by a real scenario of sparse, noisy data that requires calibrated forecasting. The multi-task approach preserves learning of a rich representation space while also improving classification accuracy. Overall, the paper addresses several important challenges in real scientific data and carries out reasonable evaluations to validate the approach.

Evaluation demonstrates that the use of UQNet significantly improves uncertainty quantification without negatively impacting the accuracy of classification. Classification bias due to class imbalance is significantly reduced by the novel fusion approach compared to not using the fusion approach.

Critical weaknesses

The organization and clarity of the paper makes it difficult to assess the novelty and impact of the work. Notably, the paper ends abruptly with specific details of empirical results but without any discussion of domain-agnostic generalization or conclusions of the impact of the results. The abstract and intro state that the approach is evaluated on “benchmarks” (plural), yet only one benchmark (singular) are actually evaluated. Citations to existing literature are missing and makes it difficult to ascertain the novelty of the approach; e.g., UQNet should be cited (line 080 instead of line 139).

The claim of avoiding dominant-class saturation is not substantiated. My understanding is that pre-peak classification is the minority class in this problem; yet the an accuracy of 45% in this regime versus 71% for overall accuracy indicates that it is still major-class dominant. Handling class imbalance is a well-studied problem in machine learning, and so further exploration of approaches is warranted. It seems that the lesson is that the fusion approach (replacing softmax) improves the dominant-class issue yet the specific choice of the weight of the fusion impacts the results.

Questions for authors

Where in Figure 1 is Model B? It is not clear to me exactly what Model B is, yet it provides significantly better results. Similarly, where is the \alpha fusion weight in Figure 1? What tradeoff is being made between Model B and fusion?

Minor questions/suggestions:

Define acronyms including NLL and ECE.

Bold lines in Table 1 should include ties (I2SB only for NRMSE and RMSE in table (a); and, I2SB only NRMSE and RMSE in table (b)). The ties in accuracy indicate that the addition of UQNet does not change the overall accuracy (for better or worse) but it does significantly improve coverage. What does the PI width column tell us, and why is there no bold for the “best” of that column?

1 comment

We sincerely thank the reviewer for their constructive feedback and for recognizing the strengths of our multi-task representation learning, decoupled UQNet calibration, and pre-softmax logit fusion. We are encouraged that you found the problem well motivated and the evaluations effective. Below, we address each of your comments and summarize the corresponding revisions.

1. Clarification on “Pre-Peak” Regime vs. Minority Classes & Prior Saturation

Temporal Regimes vs. Class Labels: In our evaluation, “pre-peak” is not a class label, but a temporal observation subset consisting of partial light curves observed before maximum brightness. Within both pre-peak and post-peak subsets, the classifier predicts the same six astrophysical transient classes (SNIa, SNII, SNIb/c, SNIIn, SNIax, SNIIb), which are highly imbalanced.

Why Pre-Peak Accuracy is Lower: Pre-peak accuracy is lower than overall accuracy because early light curves contain fewer detections, higher photometric uncertainty, and less class-discriminative information, making all six classes harder to distinguish before peak brightness.

Substantiating Reduced Prior Saturation: Because overall accuracy alone can obscure poor performance on rare classes, we additionally report Macro-F1 and per-class recall. Pre-softmax fusion combines complementary information from the latent-energy and decoded-curve classifiers. The fusion weight α=0.800\alpha=0.800 was selected exclusively using validation pre-peak Macro-F1, before held-out test evaluation. We have also softened our original “prior saturation” language and now describe fusion as improving balanced classification rather than eliminating class imbalance.

2. Discussion, Domain-Agnostic Generalization, and “Benchmarks” (Plural)

Added Discussion, Conclusion, Limitations, and Future Work: We agree that the original manuscript ended too abruptly. We will add dedicated Discussion, Conclusion, Limitations, and Future Work sections, summarized in our General Comment. These sections clarify:

  1. the broader methodological interpretation of endpoint-paired Schrödinger Bridges relative to alternative conditional generative models;
  2. the potential scientific relevance for early-stage transient forecasting and triage; and
  3. key limitations, including poor recall for ultra-rare classes at very low completeness, dependence on post-hoc UQNet calibration, and the representational ceiling imposed by the frozen encoder.

“Benchmarks” (Plural): We agree that this wording was misleading because the current study evaluates one astrophysical benchmark. We will change “benchmarks” to “benchmark” throughout the manuscript and avoid claiming demonstrated domain-agnostic generalization beyond the evaluated setting.

3. Literature Citations

We will move the UQNet citation to its first mention and expand the related-work discussion to better contextualize class-imbalanced/long-tailed classification and pre-softmax logit adjustment methods.

4. Clarification of Figure 1 (“Model B” and the α\alpha Fusion Weight)

Thank you for pointing out the missing labels in Figure 1. We will revise Figure 1 and its caption to explicitly label Model B and the fusion junction.

  • Location of Model B and α\alpha: Model B is the temporal classifier operating on the decoded multi-band light curve. The final logits are combined as

(1−α) ℓenergy+α ℓModel B.(1-\alpha)\,\ell_{\mathrm{energy}}+\alpha\,\ell_{\mathrm{Model\,B}}.

  • Trade-Off Governed by α\alpha: At α=0\alpha=0, classification uses only the latent-energy branch, while α=1\alpha=1 uses only Model B. Intermediate values combine complementary latent and decoded-curve information. The final value α=0.800\alpha=0.800 was selected using validation pre-peak Macro-F1 only, prior to held-out test evaluation.

5. Minor Questions and Table Formatting Fixes

  • Acronym Definitions: We will explicitly define NLL (Negative Log-Likelihood) and ECE (Expected Calibration Error) at first appearance in Section 4.
  • Bold Ties in Table 1: We will correct the missing bold formatting for tied NRMSE and RMSE entries in Table 1.
  • Interpretation of PI Width: We intentionally do not bold the smallest PI width because interval width cannot be interpreted independently of empirical coverage. A narrow interval can appear favorable while being severely under-calibrated. PI width should therefore be interpreted jointly with coverage; among methods with comparable coverage, narrower intervals indicate sharper uncertainty estimates.

We greatly appreciate these comments, which helped us improve the clarity, positioning, and presentation of the work.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Rebuttal Summary and Major Updates

Dear Program Chairs, Area Chairs, and Reviewers,

We thank the reviewers for their thoughtful feedback. In response, we conducted additional experiments and prepared the revisions below (to be incorporated into the final manuscript) to better isolate the core problem, justify each component, and distinguish our contribution from the frozen SELDON backbone.

Summary of Major Updates

  • Leakage-Free Temporal Completeness & Full Re-Evaluation: We found that the original implementation inadvertently used a per-object cached tmax⁡t_{\max}. We corrected this using a fixed global tmax⁡=221.61t_{\max}=221.61 days, defined from the training set only, and retrained I2SB, UQNet, and the temporal classifier across three seeds. Validation-only model selection shifted the fusion weight from α=0.825\alpha=0.825 to α=0.800\alpha=0.800. Reported values changed slightly, while the main qualitative conclusions remained consistent.

  • Empirical Diagnosis of Early Latent Degradation: We added a normalized class-centroid separation analysis in the frozen 256-dimensional hidden space. At 0–20% completeness, the Partial/Full separation ratio is 0.601, while I2SB increases it to 0.642, partially recovering the early-stage deficit.

  • Comparison with Conditional Flow Matching (CFM): We evaluated CFM in the same frozen SELDON space. CFM is a strong alternative, while I2SB shows stronger early-stage classification metrics, sharper late-stage intervals at comparable bridge-only coverage, and strong forecasting accuracy at low NFE. We frame these as differences in inductive bias and operating behavior rather than universal superiority.

  • Justification of UQNet & Logit Fusion: Raw bridge uncertainty substantially undercovers held-out observations, while post-hoc UQNet improves empirical coverage without altering point forecasts. We also clarify the complementary roles of the latent-energy and decoded-curve branches and the validation-based selection of α\alpha.

  • Added Discussion, Conclusion, and Limitations: We prepared dedicated sections clarifying the scope, interpretation, and limitations of the framework.

Discussion and Conclusion

Discussion. Our findings highlight two broader insights for sparse, irregular panel time series.

Endpoint Pairing vs. Unpaired Noise Transport: Unlike unpaired continuous-time models (e.g., CFM) that generate from isotropic Gaussian priors conditioned on partial states, anchoring transport at the partial hidden representation with completeness-modulated variance sis_i produces sharper late-stage intervals and maintains strong forecasting accuracy with as few as five sampling steps.

Utility for Real-Time Survey Triage: For widefield surveys such as the Vera C. Rubin Observatory, distinguishing rare subtypes from sparse early observations remains challenging and often motivates costly spectroscopic follow-up. By partially recovering early class separation, improving pre-peak uncertainty calibration, and maintaining forecasting accuracy with five reverse steps, our framework may support earlier probabilistic triage before peak brightness fades.

Conclusion. We addressed the asymmetric mapping between nested, incomplete observation prefixes during training and multi-modal future completions in streaming time-series forecasting. On supernova light curves, early partial encodings exhibit reduced class separation in the frozen hidden space, under-calibrated prediction intervals, and challenging imbalanced classification. Our hidden-space I2SB with completeness-aware bridge variance and an observational flux-error floor partially recovers early separation, maintains strong forecasting accuracy at five reverse steps, and produces sharper intervals at high completeness without data-level noise corruption. Decoupling trajectory generation from post-hoc UQNet calibration and using pre-softmax logit fusion with α=0.800\alpha=0.800 improves uncertainty calibration and balanced classification. Together, these components provide a framework for probabilistic early-stage forecasting and classification of transient alerts, with potential relevance to high-throughput streaming surveys.

Limitations and Future Work

Three limitations remain:

  1. Ultra-Rare Subtypes at Minimal Completeness: While our pipeline improves early macro-F1, recall on rare subtypes remains low in the earliest bins (e.g., SNIIb recall is 5.86±6.93%5.86 \pm 6.93\% at 0–20% and 4.00±2.00%4.00 \pm 2.00\% at 20–40%, rising to 6.37±3.43%6.37 \pm 3.43\% pre-peak and 60.92±5.49%60.92 \pm 5.49\% post-peak).
  2. Post-Hoc Calibration Dependence: Raw bridge trajectories substantially undercover held-out observations, motivating the decoupled UQNet stage.
  3. Frozen Encoder Bounds: Performance remains bounded by the information preserved in the frozen panel encoder; jointly adapting the encoder and bridge is a promising future direction.

The individual reviewer responses below provide the corresponding quantitative results and address each question in detail.

AuthorsOpenReview ↗

Revised Manuscript Uploaded

Dear Reviewers,

We have uploaded a revised version of the manuscript incorporating the corrections, additional experiments, and clarifications described in our rebuttal responses and general summary.

Thank you again for your thoughtful feedback. Please let us know if anything remains unclear before the deadline. Thank you very much.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.