VeriEvol: Scaling Multimodal Mathematical Reasoning via Verifiable Evol-Instruct

VeriEvol is a new iterative framework designed to enhance multimodal mathematical reasoning by addressing the challenges of scaling data and ensuring answer reliability. It features a type-aware evolution module for generating harder prompts and an HTV-Agent verifier that confirms answer validity through multi-source evidence checks. In experiments, scaling the dataset from 10K to 250K samples improved accuracy from 35.42% to 54.73%, and the framework can be integrated into existing reinforcement learning policies, providing practitioners with a robust method for reliable data construction and verification in visual reasoning tasks.