Approximately 50% of TESS transit candidates are false positives aris-ing from eclipsing binaries, instrumental artefacts, or background contamination.Existing deep learning classifiers—including CNN-based systems such as ExoMinerand ExoMiner++, achieve high recall but provide neither calibrated uncertaintyestimates nor statistically validated interpretability, properties that catalog-buildingastronomers require for prioritised follow-up. We present QViT-Exo, a hybridquantum-classical Vision Transformer that addresses both limitations simultane-ously. The model processes dual-channel 2D representations of phase-folded Keplerlong-cadence light curves, Recurrence Plots and Gramian Angular Fields—fusedwith five auxiliary diagnostic features through a ViT-B/16 backbone enhancedwith Quantum Orthogonal Neural Network (QONN) attention layers. Uncertaintyquantification is provided by Adaptive Quantum Conformal Prediction (AQCP),which produces prediction sets with provable finite-sample marginal coverage guar-antees. On the Kepler DR25 catalog (7585 KOIs), QViT-Exo achieves 91.2% recalland 51.4% precision with an Expected Calibration Error of 0.046, reducing false-positive rate by 56% via selective abstention. Quantum attention is statisticallymore focused than classical ViT (Mann–Whitney p < 0.001; entropy HQ = 4.24vs. HC = 5.27), though transit-aligned concentration was not significant at thecurrent training scale—a finding reported transparently as a limitation
Shyan Paul (Mon,) studied this question.