<p>Complex structural zones in petroliferous basins, with multi-stage superimposed deformation and significant spatiotemporal lithological heterogeneity, present intricate well log features, multi-scale lithology, and class imbalance, challenging high-precision model development. To address this, we propose the Hierarchical Fusion Transformer (HyLiFT), a parameter-efficient model maintaining high classification accuracy while significantly reducing model complexity. The innovations of HyLiFT encompass: (1) constructing a Unified Feature Mechanism (UFM) to capture multi-dimensional features from well log data; (2) developing a three-level fusion mechanism across temporal, feature, and cross-scale dimensions to enable hierarchical integration of multi-model features; (3) constructing a dynamic weight generator for adaptively adjusting the contribution of various sub-modules; and (4) incorporating a parameter sharing strategy to reduce model size. To validate the model’s performance, eight different neural network architectures were employed to predict the lithology of the Kangcun Formation (N_1-2&#xa0;K) and Jidike Formation (N₁j) within the Kuqa Depression of the Tarim Basin. Results show HyLiFT achieved 92.6% accuracy in six-class lithology classification with 198&#xa0;K parameters (43% fewer than iTransformer). Its F1-score surpassed Random Forest and XGBoost by 14.5% and 12.6%, respectively. Furthermore, its precision and recall exceeded state-of-the-art deep models, notably improving identification of minority classes like shaly conglomerate. Model training took 22&#xa0;min, with inference 37% faster than iTransformer. Feature importance analysis revealed different lithologies’ varying dependencies on well log features, explaining its performance advantages. Employing hierarchical fusion and parameter sharing, HyLiFT balances high accuracy with reduced parameters offering an efficient oil and gas exploration solution.</p>