2025MNRAS.542.2245A

The faint radio-source population includes sources dominated both by star formation and active galactic nuclei (AGNs), encoding the evolution of activity in the Universe. To investigate its nature, we probabilistically classified 4471 radio sources at $z < 0.947$ using low-frequency radio data from the LoTSS (LOFAR Two-metre Sky Survey) Deep Fields alongside a multicomponent model for nebular emission, sampled by spectra obtained with the Dark Energy Spectroscopic Instrument (DESI). This was done by combining three tools: (i) the identification of a radio excess, (ii) the BPT diagram, and (iii) a modified mass-excitation diagram, alongside Monte Carlo methods to estimate the probability that each source is either a star-forming galaxy (SFG), a radio-quiet AGN (RQ AGN), or a high-/low-excitation radio galaxy (HERG or LERG). This approach extends the probabilistic classification framework of previous works by nearly doubling the redshift range, such that we can now probabilistically classify sources over the latter half of cosmic history. Often regarded as the 'gold standard' method, spectroscopic classifications allow us to evaluate the performance of other methods. Using a 90 per cent reliability threshold, we find reasonable overall agreement ($\sim 77$ per cent) with state-of-the-art photometric classifications, but significant differences remain, including that we identify 2-5 times more RQ AGN. Furthermore, our high-confidence spectroscopic classifications show that radiatively efficient and inefficient AGN exhibit clearly distinct Eddington-scaled accretion rate distributions, contrary to recent findings in the literature. Overall, our results highlight the need for new and forthcoming spectroscopic campaigns targeting radio sources, on the pathway to the SKA.