acceptodds
Under review as a conference paper at ICLR 2027

SeaStar-TS: Self-Supervised Light-Curve Representations with Semantic Adaptation

11 views

Abstract

Astronomical surveys collect irregular time series of brightness measurements, called light curves, for far more variable stars and transient events than can receive detailed follow-up. Hand-crafted features remain strong baselines for classifying these light curves: they summarize periodicity, amplitude, and shape even when observations are irregular and noisy. We present SeaStar-TS, which learns representations directly from photometry by predicting masked observations while accounting for measurement uncertainty. Across ELAsTiCC, MACHO, StarEmbed, and LEAVES, we evaluate the approach under different label budgets and observing conditions. It surpasses hand-crafted-feature baselines on ELAsTiCC and LEAVES. On StarEmbed, its linear and MLP readouts exceed the corresponding feature-based readouts, and on MACHO it has the highest listed scores at two label budgets. As an additional use of the learned representation, alignment with class names and descriptions enables top-k recognition of classes omitted from supervised training. These results show that raw-photometry representations can compete with engineered features and place held-out classes among the top-ranked labels.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Astronomy, Astrophysics, Time series, Machine Learning, LLM

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_2S42OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Light curves are difficult inputs for machine learning: observations arrive at irregular times and with unequal uncertainties. Surveys also differ in filters, calibration offsets, and temporal coverage. The challenge is to extract a useful description of the underlying variability from these incomplete measurements and to reuse it when labels are limited or the observing survey changes. This paper focuses on the challenge of extracting a useful description of the underlying variability from these incomplete measurements and reusing it when labels are limited or the observing survey changes. To solve the problem, the authors present SEASTAR-TS, which learns representations directly from photometry by predicting masked observations while accounting for measurement uncertainty. The proposed method is evaluated on four datasets under different label bugdets and observing conditions. The experimental results shows the effectiveness of the proposed method.

Critical strengths
  1. The paper addresses the challenges in astronomical tasks. The proposed framework are designed for the characteristics of astronomical data through its input encoding and self-supervised pretraining objective, rather than simply application of machine learning methods.

  2. Unlike conventional MSE-based reconstruction, the faussian KL objective uses the measurement uncertainty. The ablation study in Appendix also shows that Gaussian KL achieves the highest performance across all label budgets.

Critical weaknesses

My main concern is the limited methodological novelty, as most components are based on established techniques. But the conbination of the methods is effective on an important scientific problem.

Questions for authors
  1. What is the main novelty beyond existing masked reconstruction, uncertainty modeling, and semantic alignment methods?

  2. Can the authors provide more direct evidence that the Gaussian KL objective is more effective than MSE, especially under different measurement uncertainties and noise levels?

  3. How much of the improvement comes from the astrophysical information in class descriptions rather than the pretrained text encoder or class names?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 2S42 (1/2)

We thank the reviewer for the constructive questions.

Q1 On methodological novelty

We agree that masked reconstruction, Gaussian likelihoods and contrastive alignment are established, and we do not claim a new learning principle. The contribution lies in what their combination achieves for irregular astronomical photometry:

  • Input and objective. Each observation (time, band, value, reported uncertainty) is one token, with no interpolation, patching or per-band grid, and the reported uncertainty enters both the input and the pretraining objective. Under the same split, this encoder exceeds a classifier on 427 hand-crafted features on ELAsTiCC (0.759 against 0.739, Table 1) and the published feature-based baselines on LEAVES (0.944 and 0.949 against 0.93, Table 5).
  • One recipe across four datasets that differ in passbands, cadence and label budget, supported by a controlled ablation in which all configurations share the same hyperparameters.
  • Class-level text targets for light curves. To our knowledge they have not been used before. They let the model rank classes without labeled examples (Table 2) and, as shown below, classify stars under a different taxonomy without labels.

Q2 Gaussian KL versus MSE

Our evidence is the controlled ablation in Appendix C.1. The pretraining objective is varied while the data, the schedule and all hyperparameters are kept identical across configurations. Averaged over the other design axes, Gaussian KL has the highest test macro-F1 of the four objectives at every label budget (Table 12):

Pretraining objective1% labels5%10%98%
Gaussian KL0.5760.6420.6650.707
Simple MSE0.5560.6370.6610.701
RoMAE MSE0.5290.6190.6450.701
Normalized MSE0.5100.6060.6480.682

Gaussian KL is also stable: none of the ten collapsed runs in Table 13 uses Gaussian KL. We agree that Gaussian KL and simple MSE are close, especially once 5% or more of the labels are available, and we will say so explicitly in the revision. We also conducted a new experiment on noisier data, adding noise of 2 and 5 times the reported error to half of the observations, with the reported errors enlarged accordingly, for pretraining, fine-tuning and evaluation alike. With the recipe of Table 1 (light curves only), test macro-F1 over three seeds is:

Noise added to half of the observationsGaussian KLSimple MSE
2 times the reported error0.732 ± 0.0060.725 ± 0.004
5 times the reported error0.713 ± 0.0020.708 ± 0.004

The two objectives remain close at both noise levels. We therefore present Gaussian KL as a principled way to use the reported measurement uncertainty that is at least as good as MSE in our experiments, and not as a major source of the gains. We do not claim that it makes the model more robust to additional noise.

AuthorsOpenReview ↗

Response to Reviewer 2S42 (2/2)

Q3 Description content versus the text encoder and class names

For held-out classes on ELAsTiCC, Table 2 already separates these: random vectors reach 0.055 top-3 accuracy, class names 0.465 and descriptions 0.559 (AUROC 0.460, 0.771 and 0.780). Most of the ranking signal is thus available from the class name through the text encoder, and the descriptions add a further gain in top-3 and top-5 accuracy.

To isolate each factor we conducted a new experiment, a cross-taxonomy test. A model aligned with LEAVES labels classifies StarEmbed stars (7 classes, 3,477 stars that the model never saw with labels) from the StarEmbed class texts alone, without any StarEmbed label. The class descriptions of the paper (Appendix G) contain a physical part and a light-curve morphology part, and we test each part alone as well as the class names. Every row uses the same kind of text for the LEAVES classes during alignment and for the StarEmbed classes at test time. The reference row maps the predicted LEAVES classes to StarEmbed classes with a hand-written table instead of text (balanced accuracy, three seeds):

TextText encoderBalanced accuracy
Full descriptionQwen3-Embedding-8B0.705 ± 0.002
Class nameQwen3-Embedding-8B0.617 ± 0.012
Physics-only descriptionQwen3-Embedding-8B0.706 ± 0.004
Morphology-only descriptionQwen3-Embedding-8B0.678 ± 0.002
Full descriptionQwen3-Embedding-0.6B0.692 ± 0.016
Full descriptione5-mistral-7b-instruct0.707 ± 0.003
Class namee5-mistral-7b-instruct0.607 ± 0.008
Reference: expert-written class mappingnone0.694 ± 0.006

Descriptions assigned to the wrong classes score 0.04 to 0.10. The class name alone carries substantial signal (0.61 to 0.62). The description content adds to it (0.705 to 0.707), and the physical part of the description is sufficient to reach the full score. The result does not depend on the text encoder (Qwen3-Embedding, Zhang et al., 2025): a model from a different family (e5-mistral-7b-instruct, Wang et al., 2024) gives the same two numbers, and a model more than ten times smaller gives 0.692. Descriptions are also the only text that reaches the accuracy of an expert-written mapping between the two taxonomies.

Reviewer_M5eaOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Classification for light curves is typically done via expensive but sample-efficient methods like Gaussian Processes or inexpensive but hand-crafted methods that require feature-engineering. In this paper, the authors propose SeaStar-TS, a self-supervised algorithm that first performs unsupervised feature learning on light curves before adapting its representation to a downstream task. The work shows that this approach outperforms many current deep-learning based methods for this task.

Critical strengths
  1. The work is incredibly clearly communicated, even for somebody with minimal astronomy background.
  2. The algorithm itself is quite principled and simple, enabling scalability and clear comparisons. Having learning algorithms that require such little domain knowledge is not typical in my experience reading AI for Science papers.
  3. The ablations for the proposed algorithms are quite comprehensive, I was most impressed by Appendix C.1
Critical weaknesses
  1. I don't see where hyperparameters are tuned, which could confound a lot of the evaluations (specifically the ablations in Appendix C.1)
  2. It seems like the pre-training offers very little for this task when looking at Appendix C.5 Table 18? I thought it'd be a lot more helpful, but other design decisions (such as including metadata) seem more helpful? The fact that your method has such an improvement over the existing methods feels suspicious to me if random initialization itself does so well.
Questions for authors
  1. Is this method of using text descriptions for classes typical for astronomy? I've seen similar things used for CLIP but wanted to understand if this was a novel contribution.
  2. Have you considered standard principles from pre-training to improve performance, including (1) compute-optimal pre-training from proper model size and data size (2) data-optimal pre-training with epoching, regularization, and ensembling
4 comments
AuthorsOpenReview ↗

Response to Reviewer M5ea (1/2)

We thank the reviewer for the positive assessment and the precise questions.

Q1 Hyperparameter tuning

In the Appendix C.1 ablation nothing is tuned per configuration. All pretraining runs share one setting: 40,000 steps, base learning rate 3e-4 at a reference batch of 256 with square-root scaling to the global batch, 3% warmup, weight decay 0.05, gradient clipping 1.0, mask ratio 0.5, and the final checkpoint is used. All 384 fine-tuning runs share one setting as well: learning rate 1e-3, 300 warmup steps, batch 256, early stopping on validation macro-F1 with patience 10. Only the ablated component changes. The maximum number of fine-tuning epochs depends on the label budget alone (200, 100, 60 and 20 epochs for 1%, 5%, 10% and 98%). The label subsets are class-stratified, nested and identical for every configuration, the checkpoint is selected on a separate 2% stratified validation split, and the test set is read once at the selected epoch. The comparison is therefore controlled. The other side of this choice is that a component could do better with its own tuning, and we will say so in Appendix C.1. We will also clarify that the 96 configurations come from 72 pretraining runs, because autoregressive pretraining with last-token pooling reuses the checkpoint of autoregressive pretraining with mean pooling, and that each configuration has one seed.

For the main results, recipes are fixed per dataset or selected on validation data only: for example the MACHO schedule is fixed across sizes, label budgets and folds, and the XGBoost baselines are selected by validation macro-F1 among 8 candidate configurations with light curves only and among 24 with metadata. We will add a table that lists, for every stage and dataset, the fixed values, the searched ranges and the split used for selection.

Q2 Small gain from pretraining in Table 18

Table 18 is the setting least favourable to pretraining: ELAsTiCC provides 1.46M labeled light curves, so supervised fine-tuning alone already reaches 0.754 and pretraining adds a consistent but small gain (0.770). We conducted a new experiment that repeats this comparison with fewer labels. It uses the recipe of Table 18 (light curves only, class-description fine-tuning, linear readout), class-stratified subsets of the training labels and three seeds, and reports test macro-F1:

LabelsRandom initializationPretrained
1%0.438 ± 0.0040.575 ± 0.006
5%0.597 ± 0.0240.695 ± 0.000
10%0.641 ± 0.0230.723 ± 0.003

Training the randomly initialized model three times longer does not close the gap (0.438, 0.573 and 0.646 at 1%, 5% and 10% of the labels). Two further, independently pretrained checkpoints give 0.591 and 0.603 at 1% labels. The benefit of pretraining therefore grows as labels become scarce, which is the regime that matters for rare classes. It is also larger on the observed surveys than on ELAsTiCC: on StarEmbed, continued self-supervised pretraining raises the linear score from 0.714 to 0.740 (Table 4).

On the comparison with existing methods: all methods in Table 1 use the ATAT split (Cabrera-Vives et al., 2024), the same balanced test set and the same metric. Appendix C.1 uses the ATCAT-style downstream protocol, in which the fine-tuned network predicts through its own classification head, and under it our backbones reach up to 0.729 with 98% of the labels (Table 11), the same range as ATCAT (0.718; Tung, 2025) and RoMAE (0.721; Zivanovic et al., 2025). The remaining gain in Table 1 comes from the fine-tuning procedure of Section 3 and the inverse-frequency-weighted readout, which matters because the training set is highly imbalanced while the test set is balanced. The strong result of the randomly initialized model reflects this downstream procedure and the encoder design, and it is not an artifact of the evaluation.

AuthorsOpenReview ↗

Response to Reviewer M5ea (2/2)

Q3 Are class descriptions typical in astronomy?

It is not typical. Class-level text targets are established in computer vision (DeViSE, Frome et al., 2013; CLIP-style zero-shot prompts, Radford et al., 2021). In astronomy, existing contrastive models align instance-level pairs: galaxy images with spectra (AstroCLIP, Parker et al., 2024), photometry with spectra and metadata (AstroM3, Rizhko & Bloom, 2025), supernova light curves with spectra (Maven, Zhang et al., 2024), and Hubble images with proposal abstracts (PAPERCLIP, Mishra-Sharma et al., 2024). To our knowledge, no prior work aligns light curves with written descriptions of astrophysical classes. We do not claim the alignment loss as new. What is new is the use of expert-written class descriptions as targets for a light-curve encoder, and the evidence that this lets the model rank classes without labeled examples (Table 2) and, in a new experiment, classify stars under another dataset's taxonomy without any of its labels: a model aligned with LEAVES labels classifies StarEmbed stars from the StarEmbed class descriptions alone, without any StarEmbed label, and reaches 0.705 balanced accuracy, against 0.694 for an expert-written mapping between the two taxonomies.

Q4 Compute-optimal and data-optimal pretraining

We have not optimized pretraining along these lines, and we agree that it is a natural next step. What we can report:

  • Model size. The L-size backbone (21.5M parameters) improves over S-size (3.7M) on MACHO at every label budget (0.634 to 0.697 at 20 samples per class, Table 3), on LEAVES (0.9441 to 0.9491, Table 5) and on StarEmbed (0.770 to 0.780 with the random-forest readout, Table 19).
  • Schedule and regularization. Pretraining length, dropout and weight decay are fixed and were not tuned. The backbones are small and cheap to pretrain: the ELAsTiCC backbone takes about 70 minutes on four GPUs, so a systematic study of data size and schedule is feasible.
  • Ensembling. In a new experiment, averaging the predicted probabilities of fine-tuning seeds on ELAsTiCC gives 0.767 for single models, 0.773 for four seeds and 0.773 for eight.

We will add the size and ensembling results to the appendix and list data-size scaling as future work.

Appreciate the clarifications, maintain my (high) score!

We sincerely thank the reviewer for the positive assessment and constructive feedback throughout the review process.

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

This paper introduces SeaStar-TS, a self-supervised framework for astronomical light-curve representation learning using uncertainty-aware masked reconstruction and semantic alignment with class descriptions. Evaluations on four datasets (ELAsTiCC, MACHO, StarEmbed, and LEAVES) demonstrate competitive or superior classification performance compared with existing neural models and hand-crafted feature baselines, including promising few-shot and cross-survey adaptation results. The semantic alignment approach also enables top-k recognition of classes unseen during supervised training.

Critical strengths
  1. The paper evaluates SeaStar-TS on four astronomical datasets covering different observation conditions and label budgets. The competitive or superior performance against existing neural models and hand-crafted feature baselines provides convincing evidence of the framework's practical effectiveness.

  2. The combination of uncertainty-aware masked reconstruction and class-description alignment offers a useful framework for astronomical representation learning. Particularly, the ability to recognize classes excluded from supervised training demonstrates potential beyond conventional closed-set classification.

Critical weaknesses
  1. Although semantic alignment enables promising top-k recognition, the uncalibrated top-1 accuracy is only 3.5%, below random chance (5%). Calibration substantially improves held-out accuracy but significantly degrades seen-class performance. This limits the practical applicability of one of the paper's key contributions.

  2. Cross-survey transfer relies heavily on additional supervised fine-tuning, with substantial performance degradation when the pretrained encoder is used directly. Furthermore, the cross-survey readout-transfer experiments involve joint supervised alignment across surveys. Consequently, the results do not convincingly establish that the learned representations generalize effectively to new surveys without labeled data.

Questions for authors

Given the very low uncalibrated top-1 accuracy (3.5%) and the substantial degradation in seen-class accuracy after calibration, is there a practical calibration strategy that does not require labeled examples of unseen classes while maintaining reasonable performance on both seen and unseen classes?

3 comments
AuthorsOpenReview ↗

Response to Reviewer PBRZ

We thank the reviewer for the careful reading and the constructive comments. Both concerns are fair, and we address each of them below with additional results.

Q1 On the low uncalibrated top-1 accuracy, and whether calibration is possible without labeled examples of the held-out class.

The low uncalibrated top-1 accuracy is the seen-class bias that is typical of generalized zero-shot recognition (Chao et al., 2016), not an absence of signal. The held-out class has AUROC 0.780 and average rank 4.5 of 20 (Table 2), but only 0.8% of the test objects are assigned to it although it makes up 5% of the test set. What is missing is a single offset on the held-out score.

We conducted a new experiment to answer this question. The offset can be set without any label of the held-out class: we choose it so that the fraction of unlabeled validation objects assigned to the new class equals its assumed prevalence, here 1/20 because the new class is one of 20 candidate classes (prior matching). Results on ELAsTiCC over the 20 held-out classes (new results are averaged over three seeds):

SettingInformation usedHeld-out top-1Seen top-1
Uncalibrated (Table 17)none0.0350.903
Prior matching (new result)unlabeled validation data0.2300.882
Prior matching, with each seen class represented by its training examples (new result)seen-class labels, unlabeled validation data0.3040.878
Offset selected with labels, diagnostic (Table 17)labels0.8150.640

So a label-free offset raises held-out top-1 from 0.035 to 0.23 while seen accuracy stays at 0.88, and it does not require the large loss in seen accuracy shown in Table 17, which is one extreme point of a trade-off curve. In the third row, the prototype of each seen class combines its description with the mean embedding of 200 of its training light curves, which are available in deployment, while the new class is still represented by its description alone. Over all offsets, the area under the seen versus held-out accuracy curve (Chao et al., 2016) is 0.763 for class descriptions, 0.751 for class names and 0.506 for random vectors. We also ran the same leave-one-class-out test on StarEmbed (7 classes, new result), where the setting of the third row gives 0.534 held-out and 0.795 seen top-1.

The limit is clear: without labels of the new class, held-out top-1 stays between 0.2 and 0.3 if seen accuracy is to be preserved, and reliable first-choice decisions need a few labeled examples. We will add these rows and the trade-off curve to Appendix C.4 and keep top-k recognition as the primary claim.

Q2 On transfer to a new survey without labeled data.

We agree that the frozen MACHO encoder transfers weakly to ZTF (Table 4) and that Appendix E, which aligns with labels from both surveys, does not test label-free transfer.

We conducted a new experiment on LEAVES that tests transfer without target labels directly. The encoder is aligned with labels from one survey only. Light curves of the other survey appear only without labels in pretraining, and its test set is classified by the nearest class-description prototype, with no label from that survey. For StarEmbed, whose taxonomy differs from that of LEAVES, the prototypes come from StarEmbed's own class descriptions. All numbers in the table are new results (balanced accuracy, S-size / L-size):

Labels used for alignment, evaluated onNo labels from the evaluation surveyReference
ASAS-SN labels, ZTF test set (9 classes)0.675 / 0.7190.886 / 0.906 with labels from both surveys
ZTF labels, ASAS-SN test set (9 classes)0.716 / 0.7970.938 / 0.944 with labels from both surveys
ASAS-SN labels, StarEmbed (ZTF photometry, a different 7-class taxonomy)0.666 / 0.6890.761 fully supervised on StarEmbed

Controls with mismatched descriptions or random targets score 0.06 to 0.12, no better than random guessing (balanced accuracy 0.11 for 9 classes). Label-free transfer is therefore far better than random guessing, and on StarEmbed it reaches 0.67 to 0.69 against 0.76 for full supervision, but within LEAVES it remains well below the reference that has labels from the evaluation survey. We will replace the readout-transfer statement of Appendix E with this experiment and report the remaining gap as a limitation.

References

  • Chao, Wei-Lun, et al. "An empirical study and analysis of generalized zero-shot learning for object recognition in the wild." ECCV 2016.

Thank you for the rebuttal. I am maintaining the scores.

We appreciate the positive assessment and constructive feedback throughout the review process.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank all reviewers for their positive assessment of the paper and for their constructive comments. The questions raised have helped us clarify several points and strengthen the evidence. We respond to each reviewer individually below, and we will incorporate the corresponding clarifications and additional results into the revised paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.