Think-Verify-Revise: Neuro-Symbolic Visual Reasoning with Vision-Language Models

Forum topic · 小凯 · 2026-09-09

Summary

This paper introduces a neuro-symbolic framework for visual reasoning that couples a Vision-Language Model (VLM) with a Dynamic Logic Tensor Network (D-LTN) in an iterative Think-Verify-Revise loop. The VLM receives a small set of labelled visual examples and proposes candidate First-Order Logic (FOL) rules conforming to a strict grammar (Think). The D-LTN is automatically assembled from these rules at runtime and verifies them against CNN-generated visual embeddings (Verify). Verification failures are fed back to guide the VLM toward its next rule hypothesis (Revise). Evaluated on the ViSudo-PC benchmark across four visual domains (MNIST, EMNIST, KMNIST, FMNIST), the system induces valid Sudoku constraint rules from only three training examples. The approach achieves AUC scores matching or exceeding prior methods such as NeuPSL and LTN, demonstrating the potential of automatic rule discovery via VLMs for tasks requiring joint visual perception and formal relational reasoning.

Paper Overview

Field: Computer Vision Authors: Homayoun Afshari, Pietro Basci, Alessandro Russo, Lia Morra Published: 2026-09-04 arXiv: 2609.05388

Introduction

Visual reasoning tasks require a system to jointly perceive visual content and apply formal relational constraints — a combination that neither pure neural nor purely symbolic approaches handle well in isolation.

The Think-Verify-Revise Framework

This paper proposes a Neuro-Symbolic (NeSy) framework that closes this gap by tightly coupling:

Think: A Vision-Language Model (VLM) receives a small set of labelled visual examples and proposes candidate First-Order Logic (FOL) rules conforming to a strict grammar.
Verify: A Dynamic Logic Tensor Network (D-LTN) is automatically assembled from these rules at runtime and evaluates them, grounded on CNN-produced visual embeddings.
Revise: Verification failures are fed back to guide the VLM in generating its next hypothesis, forming a closed iterative feedback loop.

Results

Evaluated on the ViSudo-PC benchmark across four visual domains (MNIST, EMNIST, KMNIST, FMNIST), the system induces valid Sudoku constraint rules using only three training examples. The proposed method achieves AUC scores that match or exceed previous approaches (NeuPSL, LTN), demonstrating the potential of automatic rule discovery via VLMs.

*Auto-collected on 2026-09-09*