Think-Verify-Revise: Neuro-Symbolic Visual Reasoning with Vision-Language Models and Dynamic Logic Tensor Networks

Bridging perceptual recognition and formal relational reasoning remains a central challenge in neuro-symbolic AI. This paper introduces an iterative 'Think-Verify-Revise' architecture that pairs a Vision-Language Model with Dynamic Logic Tensor Networks (D-LTN). The VLM formulates candidate First-Order Logic rules conforming to formal grammar, while runtime-assembled D-LTNs verify them against CNN embeddings and feed failures back for hypothesis revision. On visual Sudoku benchmarks across four image datasets, the system successfully discovers valid symbolic constraints using only three training examples.