When One Transit Is Enough: A Quiet Coup in the Search for Other Worlds
03 Jun 2026, Yanjiang
A Transformer-based model learns a star’s normal variability, then detects planets that transit only once — worlds invisible to traditional surveys.
What if the way we’ve been searching for planets has been teaching us to ignore the very worlds we most want to find?
That question sits uneasily at the heart of a new preprint (arXiv:2606.02778) by P. Priyanshu at SRH Hochschule in Heidelberg. The paper introduces EXOVEIL, a detection system that doesn’t hunt for planets the way astronomers have for decades. It doesn’t wait for a planet to betray itself through repetition — three transits, each identical, each regular as a metronome. Instead, it learns what a star should look like when nothing interesting is happening, and then it listens for the silence to break.
Here’s the problem in its starkest form. Most exoplanet surveys are built on a kind of trust: they assume that if something looks like a transit only once, it’s probably not a planet. Noise. A glitch. A cosmic false alarm. To earn the status of candidate, a signal must repeat. Must fold back on itself across multiple orbits until the evidence becomes undeniable. This is reasonable. It’s also, viewed from another angle, a form of systematic blindness. A planet that transits only once during a telescope’s observing window — because its orbit is long, because the window is short, because the geometry is unlucky — is invisible by definition. Not hard to see. Impossible to see.
Priyanshu’s paper confronts this with a question that feels less like a technical improvement and more like a philosophical intervention. What would happen, it asks, if we stopped demanding that the universe repeat itself before we’re willing to believe it?
The answer, it turns out, is that planets start appearing where none were supposed to exist.
The machinery behind this idea is a Transformer — the same architecture that, in slightly different form, has been quietly remaking linguistics, biology, and now astronomy. But the way EXOVEIL deploys it is genuinely unusual. Most machine learning approaches to planet detection are classifiers. Show them a folded light curve, the star’s brightness compressed into a neat periodic signal, and they’ll tell you: planet, or not planet. This works remarkably well when there is a folded light curve to show. When there isn’t — when the planet transits exactly once — the classifier has nothing to classify. It scores zero percent, by construction, at every depth of transit.
Think of it like trying to identify a song from a single note. A classifier trained on melodies is helpless. But a different kind of system — one that has learned the statistical texture of silence — might notice that the note is there at all.
EXOVEIL’s “world model” does exactly this. Priyanshu trained it on 16,499 Kepler light curves using a technique called transit-masked self-supervised learning. The idea is elegant in its inversion: during training, the model sees real stellar light curves with their actual transits deliberately hidden. It learns to predict what the brightness should be at those masked moments — not by recognizing planets, but by understanding stars. The variability of starspots. The gentle undulations of rotation. The thousand small rhythms that make a star’s light curve as individual as a heartbeat.
Once trained, the model generates a prediction for every moment in a new star’s timeline. Then Priyanshu’s pipeline subtracts that prediction from reality. What remains — the residual — is everything the world model couldn’t account for. Instrumental noise. Cosmic rays. And, just possibly, a planet.
Just one dip in starlight is all EXOVEIL needs to detect an exoplanet. This allows astronomers to find new planets with much less telescope time. (Source: arXiv:2606.02778)
Example detection of the strongest monotransit candidate KIC 11706231. Top: Full Kepler light curve (grey) with world model prediction (red). Middle: Zoom around the detected event at t = 1272.8 d, showing the 227ppm dip against the predicted baseline. Bottom: Prediction residual in ppm with pm 3sigma local noise threshold. The transit signal is clearly resolved in the residual. (Source: arXiv:2606.02778)
This is where the matched-filter detector enters. It scans these residuals not for periodic signals, but for the characteristic shape of a transit: a brief, symmetric dip with the right duration and depth. A learned classifier — an XGBoost model — then separates the plausible planets from the false alarms, achieving a classification performance that, on standard tasks, reaches an AUC of 0.938 on the Kepler DR25 dataset.
But the truly startling result comes from the single-transit injection-recovery test. When Priyanshu injected artificial transits into real Kepler data — planets that transit just once — EXOVEIL recovered 32 percent of them at a depth of about one part in a thousand — a 1000-ppm dip, roughly a Jupiter-sized planet crossing a Sun-like star. For context: no existing classification-based system can recover any. The improvement is not incremental. It is categorical. A door opens where before there was only a wall.
An important question raised by earlier work on transit detection, however, complicates this picture. Researchers who have built classification systems like ExoMiner++, as Valizadegan and colleagues demonstrated in their TESS catalogue work, have invested enormous effort in vetting their candidate lists against instrumental artifacts and stellar variability. The threshold for claiming a new planet is exacting, because the penalty for being wrong is that the catalogue becomes untrustworthy. The EXOVEIL paper reports that a blind search of over three thousand Kepler stars yielded a crop of 179 new transit-like signals not present in the existing DR25 catalogue — including more than forty that appear to transit only once. But what fraction of these are planets and what fraction are artifacts remains an open and urgent question.
The methodological tension here is genuinely productive. Cádiz-Leyton and colleagues, working on uncertainty estimation for time-series classification of variable stars, have emphasized that predictive models must be transparent about when they do not know something. Their work on conformal prediction — a framework for attaching rigorous uncertainty estimates to machine learning outputs — identified that coverage guarantees can be fragile when applied to astronomical time series with complex noise structures. Priyanshu explicitly adopts conformal prediction for the classifier stage and reports empirical coverage of 95.9%, meaning the predicted uncertainty intervals contain the true classification label with near-nominal reliability. But the harder test, as the uncertainty estimation literature makes clear, is whether such coverage holds when the data distribution shifts — when moving from Kepler to TESS, say, or from TESS to the future PLATO mission.
EXOVEIL offers a partial answer here, and it’s one of the paper’s most striking empirical claims. Applied without retraining to forty-seven confirmed TESS planets in the PLATO LOPS2 field, the system detected all forty-seven — a perfect recovery on this sample. Zero-shot cross-mission transfer — the ability of a model trained on one telescope’s data to work on another’s without adaptation — is notoriously difficult in astronomy, where every instrument has its own systematic quirks. That EXOVEIL managed it at all suggests that the world model has learned something genuinely general about stellar behavior, not just the particular fingerprint of Kepler’s detectors.
The deeper challenge, however, concerns a different kind of baseline. Salinas and colleagues, in their work on Transformer-based candidate identification in TESS full-frame images, built a system that could find transits without pre-folded input. But their validation strategy included comparison against simple, classical algorithms — dip-searchers, box-least-squares — to establish that the machine learning was adding genuine value over methods that require no training data at all. The EXOVEIL paper does not include such a comparison for the single-transit detection claim, and this absence matters. It is possible that a straightforward matched-filter scan of raw light curves, without any world model at all, would recover some fraction of the same signals. How much? The paper doesn’t say.
What the paper does say, though, is that its approach points toward a sensitivity that current methods cannot match. At PLATO’s planned 25-second cadence — a sampling rate fast enough to catch the detailed shape of a transit’s ingress and egress — Priyanshu estimates that detection could reach depths approaching a hundred parts per million — the regime where Earth-like planets around Sun-like stars become visible. That is, in the careful language of the field, approaching the regime where Earth-like planets around Sun-like stars become visible. The faint dip of a small, rocky world crossing a quiet star. The signal we’ve been hunting since Kepler first opened its shutter.
To frame this in terms borrowed from the history of astronomy: most detection methods are sieves. They let through only what matches the pattern of the mesh. EXOVEIL, in its aspiration, is something closer to a loom — it weaves a model of the star from its own light, and what doesn’t fit the weave stands out as thread that doesn’t belong. The difference is subtle but fundamental. A sieve can’t find things it wasn’t designed to catch. A loom reveals any thread that isn’t part of the intended fabric. Though of course, unlike an actual loom, the model’s weave is statistical — it can miss faint threads and sometimes sees patterns where none exist.
This raises a question that the paper does not fully answer, and that the adversarial dialogue between its claims and the existing literature sharpens considerably. The classification stage of EXOVEIL, despite the world model’s raw, un-folded architecture, still uses folded features when making its final planet-versus-false-positive decision. The system does not operate entirely in the period-free regime, because the XGBoost classifier is shown statistics derived from multiple transits whenever they happen to be available. For the single-transit case, this is not a contradiction — there is nothing to fold. But for the broader detection task, the claim of paradigm-shifting novelty must be tempered by the recognition that EXOVEIL’s final stage borrows from the very tradition it seeks to overturn.
One way to understand this is through the lens of what the paper calls “conformal prediction.” The reported coverage is tight: predictions carry a threshold that ensures the classification is wrong no more than a small fraction of the time, within the uncertainty budget. In a field where false positives have historically outnumbered real planets by enormous margins, this represents a meaningful step toward trustworthiness. But the conformal guarantee depends on the assumption that the calibration data and the deployment data are drawn from the same distribution. When Priyanshu demonstrates zero-shot transfer to TESS, the empirical success is encouraging. When future data arrives from PLATO, with its faster cadence and different noise profile, the coverage may not survive intact. This is not a criticism of the method; it is the nature of moving from demonstration to deployment.
What should we make, then, of the 179 new candidates? The paper releases them openly, with pretrained weights and a catalogue, inviting the community to inspect, verify, and challenge. This is the correct scientific posture. Some of them will be planets. Some, perhaps most, will turn out to be instrumental artifacts, uncorrected systematics, or the subtle fingerprints of stellar activity that the world model has not yet learned to distinguish. The ratio between the two categories is, for now, unknown. The paper offers not a definitive answer but a bet — that the approach is sound enough that follow-up observations, particularly radial velocity measurements, will confirm a subset of candidates as genuine.
Astronomers are about to be inundated with data from PLATO, the European Space Agency’s next-generation planet hunter. Its rapid cadence and wide field will generate light curves for thousands of stars at a resolution that makes single-transit detection theoretically possible for the first time on a large scale. The bottleneck will not be data. It will be methods. Methods that can find what existing pipelines miss, not because the pipelines are poorly designed, but because they were designed for a different kind of search.
The question, then, is not whether EXOVEIL works. It does work, in the sense that it demonstrably finds signals that other systems cannot. The question is whether it works well enough — whether the signals it finds are clean enough, the false positives manageable enough, the calibration transferable enough — to justify making it a standard tool rather than an intriguing proof of concept. That answer, like so much in science, will emerge not from any single paper but from the slow, communal process of testing, failure, correction, and occasional vindication.
There is a deeper current running beneath this technical discussion. For decades, exoplanet astronomy has operated on an implicit assumption: that the signals worth paying attention to are the ones that repeat. This made perfect sense when the alternative was drowning in noise. But it also encoded a kind of aesthetic preference — a belief that nature’s most interesting messages would arrive not once, but again and again, polite enough to announce themselves clearly. EXOVEIL challenges that aesthetic by suggesting that the most important signals might be the ones we’ve been trained to dismiss. The lone transit. The single note. The planet whose orbit is so long that it will cross its star’s face only once during our watch.
If the approach proves robust, it will change not just detection statistics but the shape of the planetary catalogue itself. Long-period planets in wide orbits — worlds that take decades to complete a single revolution — will become findable. Cold gas giants. True Earth analogs whose years last exactly as long as ours. The census of planetary systems will expand in a direction that current methods, for all their extraordinary success, have systematically excluded.
The paper does not claim to have solved this problem. It claims, more modestly and more interestingly, to have opened a new direction for the solution. Priyanshu has released the code, the weights, the candidate list. The next moves belong to the community: independent validation, comparison against baselines, the slow work of separating detection from confirmation. We are left not with certainty but with a framework — a way of thinking about planetary signals that does not require them to repeat before they can be believed. And in the history of astronomy, sometimes the most important revolutions begin not with a discovery but with a change in what counts as worth looking for.
What the world model ultimately learns, after training on thousands of stars, is a kind of expectation. A sense of what is ordinary. And the detection of a planet, in this scheme, is simply the moment when the ordinary fails. Perhaps that is the deepest insight EXOVEIL offers — not a better classifier, but a different relationship to the data: one that listens not for a familiar melody but for the sudden, telling silence that precedes something new.
— Yanjiang
Yanjiang is an online editor of LoomSci.com.
References
- P. Priyanshu, One Transit Is All You Need: Detecting Exoplanets Through Learned Stellar Behaviour with EXOVEIL, arXiv:2606.02778
- Valizadegan et al., ExoMiner++: Enhanced Transit Classification and a New Vetting Catalog for 2-Minute TESS Data, arXiv:2502.09790
- Cádiz-Leyton et al., Uncertainty estimation for time series classification: Exploring predictive uncertainty in transformer-based models for variable stars, arXiv:2412.10528
- Salinas et al., Exoplanet Transit Candidate Identification in TESS Full-Frame Images via a Transformer-Based Algorithm, arXiv:2502.07542