COSMOS2025: A Machine Learning Census of Massive Quiescent Galaxies at $2.5<\text{z}<5$

The existence of massive quiescent galaxies at high redshifts ($\text{z}\gtrsim 2$) strongly constrains the rapid quenching mechanisms in galaxy evolution models. We present a machine learning framework to identify massive ($\log ({\text{M}}_{\ast }\mathrm{/}{\text{M}}_{\odot })>9.5$) quiescent galaxies at $2.5<\text{z}<5$ in the COSMOS2025 catalog. We train a $\mathtt{\text{CatBoostClassifier}}$ on mock photometry from the Santa Cruz semi-analytic models (SAMs), incorporating key JWST NIRCam bands and realistic noise to transfer the SAM-derived quiescent label (based on specific star-formation rate) to the observational space. When validated against the SAM ground truth, our classifier achieves a significantly higher recall (completeness) of 78% (compared to 53% for spectral energy distribution (SED)-fitting), while maintaining a high purity of 82%. Applied to the COSMOS2025 sample, and assuming the SAM definition of quiescence transfers to the real Universe, the model identifies 1111 quiescent candidates, a population 2.6 times larger than the 427 candidates identified via the catalog’s simple SED-fitting configuration. Under the SAM definition of quiescence, this consistent pattern of high purity but poor completeness suggests that the SED-fitting methods, constrained by simplified parametric star-formation histories, may miss a significant fraction of the quiescent population, likely galaxies in crucial transitional evolutionary stages. The trained classifier and classified COSMOS2025 sample are publicly available.