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

In September 2026, arXiv released a neural symbolic visual reasoning framework named Think-Verify-Revise. This research combines visual language models with dynamic logical tensor networks to achieve automatic induction of first-order logical rules, rule verification based on CNN embeddings, and feedback correction through a closed-loop iteration mechanism. The aim is to solve the challenge of jointly considering content and form constraints in visual reasoning. In the ViSudo-PC benchmark, the system can effectively induce rules under single-instance constraints in four visual domains: MNIST, EMNIST, KMNIST, and FMNIST using only three labeled examples, and its AUC score is comparable to or exceeds that of existing methods such as NeuPSL and LTN. The related code has been made open-source.