Sparse supervision turns polar pretraining into transferable molecular response fields

Electronic polarization shapes intermolecular energies and forces, yet a molecular dipole discards much of the spatial response experienced by nearby molecules. Recomputing that response with quantum chemistry for every environment limits its reuse. We ask whether polar pretraining can instead be redirected toward spatial response with only tens of labelled environments. Our model, Geometry-Learned Induced Dipole and Electrostatic Response (GLIDER), learns a constrained atom-centred representation of the counterpoise-consistent interaction-induced response. We train on 14 chemistries and 48 three-water environments, then freeze one checkpoint before testing it on 56 solutes absent from response supervision in 224 water environments and, without retraining, on 12 of those solutes with three new neighbour species. Our model lowers response-ESP error by 32–40% against the strongest eligible public baseline used without response-specific fitting, and by 27% after the change in neighbour identity. A separate test couples a water molecule excluded from both the model input and base-response calculation to the predicted field: GLIDER lowers response-coupling energy error from 0.114 to 0.049 kcal mol$^{-1}$ and improves near-field torque and orientation. Sparse supervision can therefore turn polar pretraining into a reusable spatial electronic response across solutes, neighbours, and geometries.