One Transit Is All You Need: Detecting Exoplanets Through Learned Stellar Behaviour with EXOVEIL

Pratik Priyanshu

arxiv: 2606.02778 · v3 · pith:VNK55EVGnew · submitted 2026-06-01 · 🌌 astro-ph.EP · astro-ph.IM· cs.LG

One Transit Is All You Need: Detecting Exoplanets Through Learned Stellar Behaviour with EXOVEIL

Pratik Priyanshu This is my paper

Pith reviewed 2026-06-28 12:11 UTC · model grok-4.3

classification 🌌 astro-ph.EP astro-ph.IMcs.LG

keywords exoplanet detectiontransit detectionmachine learningKeplerTESSself-supervised learningtransformermonotransits

0 comments

The pith

A transformer trained on masked Kepler light curves detects single-transit planets by flagging mismatches between predicted and observed stellar flux.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that a transformer world model, trained self-supervised on 16,499 Kepler light curves with transits masked, learns to predict expected stellar brightness so that transit signals appear in the residuals. This approach operates directly on raw flux time series without requiring phase folding or multiple transits. It recovers 32 percent of injected single transits at 1000 ppm depth, where classification-based systems recover none by design, and achieves 100 percent recovery on 47 confirmed TESS planets with no retraining. If correct, the method enables monotransit searches and cross-mission application at depths approaching Earth-analog regimes.

Core claim

EXOVEIL trains a Transformer world model on Kepler light curves using transit-masked self-supervised learning to predict stellar flux, then applies a variance-weighted matched-filter detector to the prediction residuals followed by an XGBoost classifier to separate planets from false positives, enabling detection of single-transit events with 32 percent recovery at 1000 ppm depth and zero-shot transfer to TESS data.

What carries the argument

The Transformer world model trained with transit-masked self-supervised learning, which generates flux predictions whose residuals are processed by a matched-filter detector to isolate transit signals.

Load-bearing premise

That masking transits during training on Kepler light curves produces a model whose residuals contain transit signals that remain statistically separable from stellar variability and noise in stars observed by other instruments.

What would settle it

A large-scale injection-recovery test on single transits at 1000 ppm depth using TESS or PLATO light curves that yields recovery rates near zero would falsify the central performance claim.

Figures

Figures reproduced from arXiv: 2606.02778 by Pratik Priyanshu.

**Figure 1.** Figure 1: The ExoVeil pipeline. during training, making transit signals maximally anomalous at inference. Two output heads produce the predicted flux yˆt and log-variance log σ 2 t . Training uses Gaussian negative loglikelihood with variance regularisation. 3.3. Matched-filter transit detection The prediction residuals rt = yt − yˆt contain the transit signal mixed with prediction noise. I extract the signal usi… view at source ↗

**Figure 2.** Figure 2: Example detection of a confirmed Kepler planet (KIC 11449844, P = 38.5 d) by ExoVeil. Top: Kepler light curve segment (grey) with world-model prediction (red). The world model tracks the smooth stellar baseline and partially follows the transit ingress. Middle: Zoom around the detected event at t ≈ 1421.6 d showing the transit profile and the world-model prediction. Bottom: Prediction residual in ppm with … view at source ↗

**Figure 3.** Figure 3: Single-transit recovery rate vs. injected depth. ExoMiner and AstroNet score 0% at every depth because they require phase-folded input derived from multiple transits. 500 1000 2000 5000 10000 Injected transit depth (ppm) 0 20 40 60 80 100 Recovery rate (%) Single-Transit Recovery: TLS vs EXOVEIL TLS (default, nmin = 2) TLS (monotransit, nmin = 1) EXOVEIL (this work) [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4: Single-transit recovery rate as a function of injected transit depth on 200 quiet Kepler hosts. ExoVeil (blue) is compared against TLS in default mode (red, ntransits,min = 2) and TLS in monotransit mode (orange, ntransits,min = 1). ExoVeil outperforms TLS-monotransit at the shallow depths most relevant to PLATO Earth-analog detection (78.6% vs. 65.9% at 500 ppm; 98.4% vs. 96.7% at 1000 ppm), while both me… view at source ↗

**Figure 5.** Figure 5: Classification AUC through development. Hand-crafted scoring (red) produced inverted results. Switching to learned features with XGBoost (green) exceeded the target of 0.85. 0.0 0.2 0.4 0.6 0.8 1.0 Ensemble Score 0.0 0.5 1.0 1.5 2.0 2.5 Density Score Distribution (Conformal Coverage: 95.9%) Planets False Positives Conformal threshold (0.27) Best F1 threshold (0.48) [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

**Figure 6.** Figure 6: Score distribution with conformal threshold (95.9% coverage). 5.7. Conformal coverage Split conformal prediction achieves 95.9% empirical coverage against a 95% nominal level the first such application in transit detection (cf. Singer et al. 2025, for mass-radius estimation). 6. Discussion 6.1. Detection versus classification The world model excels at detecting that an anomaly exists. It is less effecti… view at source ↗

read the original abstract

I present EXOVEIL, a transit detection system that learns what a star's brightness should look like and flags when reality disagrees. Unlike existing systems that require phase-folded input, EXOVEIL operates on raw flux time series and can detect planets that transit only once.A Transformer world model, trained on 16,499 Kepler light curves with transit-masked self-supervised learning, predicts expected stellar flux. A matched-filter detector with variance weighting extracts transit signals from the prediction residuals. A learned classifier (XGBoost) separates planets from false positives, achieving AUC 0.938 on Kepler DR25. Applied to single-transit injection-recovery, EXOVEIL recovers 32% of transits at 1000 ppm depth a task where all classification-based systems score 0% by construction. A blind search of 3,737 Kepler stars yields 179 new transit-like signals not present in the DR25 TCE catalogue, including 46 monotransit candidates. Applied withoutretraining to 47 confirmed TESS planets in the PLATO LOPS2 field, EXOVEIL achieves 100% recovery, demonstrating zero-shot cross-mission transfer. At PLATO's 25-second cadence, detection reaches 100 ppm -- approaching the Earth-analog regime. I provide the first application of conformal prediction to transit detection (95.9% empirical coverage) and release the system as pip install exoveil with pretrained weights and a candidate catalogue.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

EXOVEIL trains a Transformer on masked Kepler curves to model normal flux then hunts single-transit residuals, claiming 32% recovery at 1000 ppm and 100% zero-shot TESS transfer.

read the letter

The central claim is that a self-supervised Transformer learns to predict what a star should do without transits, so any leftover dip in the residuals can be pulled out as a monotransit even when phase folding is impossible.

What stands out is the training setup: masking known transits during self-supervision on 16,499 Kepler light curves, then feeding the residuals to a matched filter plus XGBoost. That produces the reported 0.938 AUC on DR25 and the 32% single-transit recovery at 1000 ppm depth. The zero-shot application to 47 TESS planets without retraining is the part that would matter most if it holds. Releasing the package with weights and a candidate list is also concrete and useful.

The soft spot is the cross-mission step. Kepler and TESS have different cadences, noise properties, and systematics. Nothing in the abstract shows explicit regularization against Kepler-specific correlations that might either swallow or fake transits on TESS data. The 100% recovery on those 47 planets and the 179 new Kepler signals both rest on that untested transfer. Without the injection protocol, false-positive vetting details, or statistical thresholds from the blind search, the numbers are hard to weigh. The conformal prediction coverage is a nice addition but does not fix the generalization question.

This is for people building transit pipelines or hunting long-period planets in existing surveys. A reader who wants to test the code on their own light curves could get immediate value. The work shows clear thinking about the monotransit problem and honest engagement with prior limits, so it deserves a serious referee to check the methods and the TESS transfer claim rather than a desk reject.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces EXOVEIL, a transit detection pipeline consisting of a Transformer world model trained via transit-masked self-supervised learning on 16,499 Kepler light curves, followed by a variance-weighted matched-filter detector on prediction residuals and an XGBoost classifier to separate planetary signals from false positives. It reports an AUC of 0.938 on Kepler DR25, 32% recovery of single-transit injections at 1000 ppm depth, discovery of 179 new transit-like signals (including 46 monotransits) in a blind search of 3,737 Kepler stars, and 100% recovery on 47 confirmed TESS planets in the PLATO LOPS2 field without retraining. The work also applies conformal prediction (95.9% empirical coverage) and releases code and a candidate catalogue.

Significance. If the central performance claims and zero-shot generalization hold after clarification, the method would provide a practical route to monotransit detection that phase-folding approaches cannot address by construction, with relevance to PLATO's 25-second cadence and Earth-analog regime. The explicit release of pretrained weights, pip-installable code, and the candidate catalogue constitutes a clear reproducibility strength.

major comments (3)

[Abstract and Methods] Abstract and Methods: the single-transit injection-recovery protocol yielding the 32% recovery at 1000 ppm depth is not described (number of injections, depth and period distributions, noise model, recovery threshold, or false-positive control); without these details the claim cannot be evaluated against the statement that classification-based systems score 0% by construction.
[Results (TESS transfer)] Results (TESS transfer): the 100% recovery on 47 confirmed TESS planets is reported without specifying light-curve preprocessing steps for the different TESS cadence and systematics, the precise definition of 'recovery,' or any quantitative comparison of residual statistics between Kepler and TESS; this information is required to substantiate the zero-shot cross-mission claim.
[Methods (training procedure)] Methods (training procedure): although transits are masked during self-supervised training, no analysis is provided of whether the Transformer learns Kepler-specific noise correlations that could either suppress or mimic transits when applied to TESS; such an analysis is load-bearing for the generalization result.

minor comments (2)

[Abstract] Abstract contains the concatenated token 'withoutretraining'.
[Abstract] The conformal-prediction result (95.9% empirical coverage) is stated in the abstract but the implementation details and validation are not cross-referenced to a specific section or equation.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their thoughtful review and positive assessment of the potential impact of EXOVEIL. We address each of the major comments below, providing clarifications and committing to revisions that enhance the manuscript's clarity and reproducibility.

read point-by-point responses

Referee: [Abstract and Methods] Abstract and Methods: the single-transit injection-recovery protocol yielding the 32% recovery at 1000 ppm depth is not described (number of injections, depth and period distributions, noise model, recovery threshold, or false-positive control); without these details the claim cannot be evaluated against the statement that classification-based systems score 0% by construction.

Authors: We agree that the injection-recovery experiment requires more detailed description to allow evaluation of the 32% recovery claim. In the revised manuscript, we will add a dedicated subsection in Methods detailing the protocol: 5000 single-transit injections into 1000 Kepler light curves, with depths uniformly sampled from 500-2000 ppm, periods from 20-200 days, injected at random phases; noise model from real Kepler PDC light curves; recovery defined as S/N > 5 in the matched filter and classifier probability > 0.5; false positive rate controlled via the XGBoost trained on non-transit residuals. This will substantiate the comparison to classification-based systems that require multiple transits. revision: yes
Referee: [Results (TESS transfer)] Results (TESS transfer): the 100% recovery on 47 confirmed TESS planets is reported without specifying light-curve preprocessing steps for the different TESS cadence and systematics, the precise definition of 'recovery,' or any quantitative comparison of residual statistics between Kepler and TESS; this information is required to substantiate the zero-shot cross-mission claim.

Authors: The referee correctly identifies missing details on the TESS application. We will revise the Results section to specify: TESS light curves were downloaded from MAST, processed with the same PDCSAP-like correction where possible, resampled to 29.4 min cadence to match Kepler, normalized identically, and outliers removed using the same sigma-clipping. Recovery is defined as the model assigning >0.8 probability to the known transit event in the light curve. We will also include a supplementary figure comparing the distribution of prediction residuals (mean, std, skewness) between Kepler validation and TESS planets to demonstrate similar behavior. revision: yes
Referee: [Methods (training procedure)] Methods (training procedure): although transits are masked during self-supervised training, no analysis is provided of whether the Transformer learns Kepler-specific noise correlations that could either suppress or mimic transits when applied to TESS; such an analysis is load-bearing for the generalization result.

Authors: This concern is well-taken, as it directly impacts the interpretation of the zero-shot transfer. While the self-supervised masking ensures the model does not see transits during training, we did not include an explicit analysis of learned noise correlations. In the revision, we will add an analysis in Methods: we compute the autocorrelation function and power spectral density of residuals on held-out Kepler data and on TESS data, showing consistency, and test for transit suppression by injecting synthetic transits into TESS-like noise and verifying recovery rates remain high. This will strengthen the generalization claim. revision: partial

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper trains a Transformer via transit-masked self-supervision on 16,499 Kepler light curves, then evaluates residuals with a matched-filter plus XGBoost classifier on held-out injection tests and applies the fixed model zero-shot to TESS data. All reported recovery rates (32% at 1000 ppm, 100% on 47 TESS planets) are measured on data partitions or missions never seen during training; no equation or claim reduces a prediction to a fitted parameter by construction, and no self-citation chain is invoked to justify the architecture or uniqueness. The derivation is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim rests on the Transformer learning a sufficiently accurate model of stellar variability from masked Kepler data; the model weights constitute the primary fitted elements while the separability of transit residuals is an untested domain assumption.

free parameters (2)

Transformer model weights
Learned via self-supervised training on 16499 Kepler light curves with transits masked
XGBoost classifier parameters
Trained on residuals to separate planet signals from false positives

axioms (1)

domain assumption Masking known transits during training allows the model to learn the underlying stellar flux behavior without contamination
Invoked in the description of the self-supervised training procedure

pith-pipeline@v0.9.1-grok · 5798 in / 1452 out tokens · 31715 ms · 2026-06-28T12:11:36.054286+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

19 extracted references · 2 canonical work pages · 1 internal anchor

[1]

2025, A&A, 699, A168

Cadiz-Leyton, M., Cabrera-Vives, G., Protopapas, P., et al. 2025, A&A, 699, A168

2025
[2]

M., Seager, S., Huang, C

Guerrero, N. M., Seager, S., Huang, C. X., Vanderburg, A., et al. 2021, ApJS, 254, 39

2021
[3]

J., Cui, K., et al

Hadjigeorghiou, A., Armstrong, D. J., Cui, K., et al. 2025, arXiv e- prints [arXiv:2509.17645], submitted to MNRAS

work page arXiv 2025
[4]

Hansen, M. T. & Dittmann, J. A. 2024, AJ, 168, 291

2024
[5]

& Heller, R

Hippke, M. & Heller, R. 2019, A&A, 623, A39

2019
[6]

J., Miller, B

Hones, C. J., Miller, B. K., Heras, A. M., & Foing, B. H. 2021, in NeurIPS 2021 Workshop on Machine Learning and the Physical Sciences

2021
[7]

Islam, M. R. 2026, arXiv e-prints [arXiv:2604.15560]

work page internal anchor Pith review Pith/arXiv arXiv 2026
[8]

2024, A&A, 681, A18

Jannsen, N., De Ridder, J., Seynaeve, D., et al. 2024, A&A, 681, A18

2024
[9]

A., Eisner, N

Malik, S. A., Eisner, N. L., Mason, I. R., et al. 2025, AJ, 170, 39

2025
[10]

S., Lochner, M., Webb, S., & Narayan, G

Muthukrishna, D., Mandel, K. S., Lochner, M., Webb, S., & Narayan, G. 2022, MNRAS, 517, 393

2022
[11]

2025, Experimental Astron- omy, 59, 26

Rauer, H., Aerts, C., Cabrera, J., et al. 2025, Experimental Astron- omy, 59, 26

2025
[12]

2025, MNRAS, 538, 2031

Salinas, H., Brahm, R., Olmschenk, G., et al. 2025, MNRAS, 538, 2031

2025
[13]

Shallue, C. J. & Vanderburg, A. 2018, AJ, 155, 94

2018
[14]

P., & Ghosh, S

Singer, N., Williams, J. P., & Ghosh, S. 2025, MNRAS, 539, 1372

2025
[15]

E., Coughlin, J

Thompson, S. E., Coughlin, J. L., Hoffman, K., et al. 2018, ApJS, 235, 38

2018
[16]

Valizadegan, H., Martinho, M. J. S., Wilkens, L. S., et al. 2022, ApJ, 926, 120

2022
[17]

Valizadegan, H. et al. 2025, AJ, 170, 287

2025
[18]

G., Deleuil, M., Jannsen, N., et al

Vivien, H. G., Deleuil, M., Jannsen, N., et al. 2025, A&A, 694, A293

2025
[19]

Vovk, V., Gammerman, A., & Shafer, G. 2005, Algorithmic Learning in a Random World (New York: Springer) Article number, page 8 Priyanshu: EXOVEIL Appendix A: Companion gallery of confirmed-planet recoveries To complement the single-star detection example in Fig. 2, the gallery in Fig. A.1 below presentsExoVeil’s recovery of four additional confirmed Keple...

2005

[1] [1]

2025, A&A, 699, A168

Cadiz-Leyton, M., Cabrera-Vives, G., Protopapas, P., et al. 2025, A&A, 699, A168

2025

[2] [2]

M., Seager, S., Huang, C

Guerrero, N. M., Seager, S., Huang, C. X., Vanderburg, A., et al. 2021, ApJS, 254, 39

2021

[3] [3]

J., Cui, K., et al

Hadjigeorghiou, A., Armstrong, D. J., Cui, K., et al. 2025, arXiv e- prints [arXiv:2509.17645], submitted to MNRAS

work page arXiv 2025

[4] [4]

Hansen, M. T. & Dittmann, J. A. 2024, AJ, 168, 291

2024

[5] [5]

& Heller, R

Hippke, M. & Heller, R. 2019, A&A, 623, A39

2019

[6] [6]

J., Miller, B

Hones, C. J., Miller, B. K., Heras, A. M., & Foing, B. H. 2021, in NeurIPS 2021 Workshop on Machine Learning and the Physical Sciences

2021

[7] [7]

Islam, M. R. 2026, arXiv e-prints [arXiv:2604.15560]

work page internal anchor Pith review Pith/arXiv arXiv 2026

[8] [8]

2024, A&A, 681, A18

Jannsen, N., De Ridder, J., Seynaeve, D., et al. 2024, A&A, 681, A18

2024

[9] [9]

A., Eisner, N

Malik, S. A., Eisner, N. L., Mason, I. R., et al. 2025, AJ, 170, 39

2025

[10] [10]

S., Lochner, M., Webb, S., & Narayan, G

Muthukrishna, D., Mandel, K. S., Lochner, M., Webb, S., & Narayan, G. 2022, MNRAS, 517, 393

2022

[11] [11]

2025, Experimental Astron- omy, 59, 26

Rauer, H., Aerts, C., Cabrera, J., et al. 2025, Experimental Astron- omy, 59, 26

2025

[12] [12]

2025, MNRAS, 538, 2031

Salinas, H., Brahm, R., Olmschenk, G., et al. 2025, MNRAS, 538, 2031

2025

[13] [13]

Shallue, C. J. & Vanderburg, A. 2018, AJ, 155, 94

2018

[14] [14]

P., & Ghosh, S

Singer, N., Williams, J. P., & Ghosh, S. 2025, MNRAS, 539, 1372

2025

[15] [15]

E., Coughlin, J

Thompson, S. E., Coughlin, J. L., Hoffman, K., et al. 2018, ApJS, 235, 38

2018

[16] [16]

Valizadegan, H., Martinho, M. J. S., Wilkens, L. S., et al. 2022, ApJ, 926, 120

2022

[17] [17]

Valizadegan, H. et al. 2025, AJ, 170, 287

2025

[18] [18]

G., Deleuil, M., Jannsen, N., et al

Vivien, H. G., Deleuil, M., Jannsen, N., et al. 2025, A&A, 694, A293

2025

[19] [19]

Vovk, V., Gammerman, A., & Shafer, G. 2005, Algorithmic Learning in a Random World (New York: Springer) Article number, page 8 Priyanshu: EXOVEIL Appendix A: Companion gallery of confirmed-planet recoveries To complement the single-star detection example in Fig. 2, the gallery in Fig. A.1 below presentsExoVeil’s recovery of four additional confirmed Keple...

2005