AI News

Neuro-symbolic models are often supervised only on final labels, so they can predict correctly while assigning the wrong intermediate concepts. This reasoning shortcut occurs because multiple concept assignments may be compatible with the same label. The authors introduce Soft-PNet, a prototypical-network method that replaces hand-crafted, task-specific differentiable losses with a Metropolis walk over a precomputed cache of feasible symbolic solutions. A prototype distribution is built from one labeled anchor per concept, and training uses a single KL-divergence objective between that distribution over cached solutions and the network's concept predictions. The same objective is intended to work across tasks, including settings where the full solution space cannot be enumerated. Under scarce supervision on MNIST-EvenOdd, Visual Sudoku, and Kand-Logic, Soft-PNet matched loss-engineered prototypical networks on concept and label performance. It also recovered concepts that soft-grounding baselines missed, while requiring less training time and no task-specific loss design.