2025A&A...700A.260O

Context. Machine learning (ML) and deep learning (DL) techniques are increasingly being adopted across many fields of astrophysics. With the growing availability of data and refined acquisition methods, these approaches can now be applied to a wide range of tasks, from redshift estimation and light curve variability studies to astrophysical source classification.
Aims. For this work, our goal was twofold. Firstly, we wanted to classify blazars from the Fermi 4LAC-DR3 catalogue in order to identify the most probable origin of those with currently unknown classifications (BCUs); secondly, we wanted to explore the full sample of blazars to investigate the structure and the redshift-luminosity evolution of the blazar population. Particular attention was given to the transition region between flat spectrum radio quasars (FSRQs) and BL Lacertae (BL Lacs), which may provide key insights into the nature and development of the accretion disk activity. Building on recent studies, we explored the role of changing-look blazars (CLBs) as potential intermediates in this transition.
Methods. We implemented a classification approach based on a strong benchmark model (XGBoost) and a state-of-the-art foundation model, pre-trained on millions of tabular datasets (TabPFN). This constitutes, to the best of our knowledge, the first application of such a pre-trained model to high-energy astrophysics. By extracting the high-dimensional latent space provided by the pre-trained model and by reducing its dimensions, we provided a two-dimensional representation of the blazar population. This enables a nuanced interpretation of the characteristics of sources that lie at the boundary between FSRQs and BL Lacs.
Results. By analysing the reduced latent representation of our data given by the pre-trained model, we identified a clear continuum between FSRQs and BL Lacs, both in terms of high-energy properties and central engine characteristics. This continuous structure reveals a population of sources with intermediate properties, CLBs, which represent a transitional evolutionary stage between FSRQs and BL Lacs. These findings support the scenario of a gradual evolution from FSRQs, with radiatively efficient accretion disks and high Compton dominance, towards BL Lacs characterized by radiatively inefficient flows.
Conclusions. A key outcome of our study is that a single probability score, combined with the pre-trained model's latent space, can robustly describe any blazar in the sample, offering a new framework for visualizing and interpreting blazar diversity beyond discrete class boundaries. The use of a pre-trained model without the need for domain-specific optimization offers a fast and scalable tool, particularly well-suited for identifying and characterizing ambiguous or transitional sources in current and future blazar catalogues.