The characterization of the pre-main-sequence (PMS) stars from the Gaia spectrophotometric data is one of the new objectives in preparation of the fourth Gaia Data Release (DR4) expected in 2026. The mass accretion taking place in PMS stars induces a continuum radiation excess that is detectable in the ultraviolet, optical, and infrared ranges covered by the Gaia Blue (BP) and Red (RP) Photometer low-resolution spectra (300–1100 nm). The analysis of the BP/RP spectra allows us to determine their stellar and accretion parameters (effective temperature \(T_{\mathrm {eff}}\) , logarithm of the surface gravity \(\log {g}\) , metallicity [M/H], stellar mass M, extinction \(A_0\) , mass accretion rate \(\log \dot {M}\) and filling factor f). We present here an inference method based on the Conditional Variational Autoencoders (CVAE) trained on a precomputed grid of BP/RP model spectra parametrized according to the magnetospheric accretion model. The accuracy in recovering the parameters of the precomputed grid is better than \({\sim } \) 5 K in \(T_{\mathrm {eff}}\) , \(\sim \) 0.005 dex in \(\log {g}\) , \(\sim \) 0.001 in [M/H], \(\sim \) 1.5 in \(\log {\dot {M}}\) , \(\sim \,\) 0.003 in f, \(\sim \,\) 0.001 in M, \(\sim \,\) 0.007 in \(A_0\) . We demonstrate the capability of the method to capture the correlations between parameters.