AI
New Dataset and Model Improve How AI Tackles Visual Math Problems with Multiple Solution Paths
Researchers introduce a dataset capturing diverse reasoning approaches for multimodal math problems, boosting both accuracy and solution variety in vision-language AI systems.
A team has unveiled a new approach to teaching multimodal AI systems—those that process both images and text—to solve mathematical problems more effectively by incorporating multiple valid reasoning strategies into their training.
The Challenge with Current Methods#
Existing multimodal large language models (MLLMs) designed for math reasoning typically learn from single solution paths paired with image-question combinations. This narrow training regimen misses a fundamental reality of mathematics: many problems admit multiple legitimate approaches, each revealing different insights. By training only on one canonical solution per problem, these systems fail to develop the flexible thinking needed for robust reasoning across varied contexts.
A Richer Dataset and Training Approach#
The researchers introduced MathV-DP, a dataset that captures multiple distinct solution trajectories for each image-question pair, enriching the supervision signal with diverse solving perspectives. Building on this foundation, they developed Qwen-VL-DP, an enhanced vision-language model fine-tuned through a two-stage process combining supervised learning with group relative policy optimization (GRPO), a rule-based reinforcement learning method. The GRPO approach incorporates two key elements: a correctness discriminator that evaluates solution validity, and a diversity-aware reward function that encourages the model to learn from different yet correct reasoning paths.
Results and Impact#
Testing on standard benchmarks—MathVista's minitest and Math-V—demonstrated that Qwen-VL-DP substantially outperformed prior base multimodal models on both accuracy and generative diversity. The findings underscore how exposing AI systems to multiple valid reasoning perspectives during training strengthens their mathematical problem-solving capabilities and produces more varied, thoughtful outputs across different solution strategies.
The work highlights a broader principle: richer, more diverse training signals that reflect the actual plurality of valid approaches can meaningfully improve how artificial intelligence learns to reason about complex domains.
