Towards a unified scheme of blazar evolution

Abstract

Machine learning (ML) and deep learning (DL) techniques are increasingly used across astrophysics, enabled by the growing availability of data and improved acquisition methods. These approaches now support tasks from redshift estimation to source classification. In this work, we aim to (i) classify blazars from the Fermi 4LAC-DR3 catalogue, in particular to identify the likely origin of blazars of uncertain type (BCUs), and (ii) investigate the full blazar sample to study their structure and redshift-luminosity evolution. We focus especially on the transition region between Flat Spectrum Radio Quasars (FSRQs) and BL Lacertae objects (BL Lacs), which may yield insights into accretion disk evolution. We examine Changing-Look Blazars (CLBs) as potential intermediates in this transition. We implement a classification pipeline using both a strong benchmark model (XGBoost) and a foundation model pre-trained on millions of tabular datasets (TabPFN). By extracting and reducing the high-dimensional latent space of the best model, we provide a 2D representation of the blazar population. This reveals a continuum between FSRQs and BL Lacs, including CLBs as transitional sources. These results support a scenario of gradual evolution from radiatively efficient (FSRQ-like) to inefficient (BL Lac-like) accretion. Ultimately, we show that a single probability score, combined with the latent space, offers a new framework for interpreting blazar diversity beyond discrete classes.