<p>During vanadium extraction in a basic oxygen furnace, semi-steel temperature, carbon content, vanadium content, and V₂O₃ content in vanadium slag must be controlled within appropriate ranges to obtain high-grade vanadium slag and high-carbon semi-steel for subsequent steelmaking. Most existing studies adopt data-driven models with multivariate inputs and single-variable outputs, neglecting variable coupling and reaction mechanisms, which limits consistency with metallurgical principles in multi-task prediction. In this study, a thermodynamic mechanism-constrained Feature Tokenizer (FT)-Transformer multi-task prediction model (MC-FTMT) is proposed. Numerical and categorical features are mapped into a unified feature sequence through a Feature Tokenizer and encoded by a Transformer to extract globally shared features, followed by four independent output layers. Inequality-based mechanism loss functions are embedded during training to ensure coordinated multi-task prediction with physical interpretability. Results show hit rates of 94% for semi-steel temperature within ± 15°C, 96.67% for carbon content within ± 0.04 wt.%, 97.34% for vanadium content within ± 0.0025 wt.%, and 93.34% for slag V₂O₃ content within ± 0.5 wt.%.</p>