UniCAR-RL improves visual math reasoning in AI models
UniCAR-RL: Seeing Better before Thinking Deeper in Visual Mathematics
Artificial Intelligence
Summary
Visual language AI models have a hard time understanding math problems with images because they often misinterpret what they see, leading to wrong answers. The authors created a new training approach called UniCAR-RL that separately improves the model’s ability to see details and to think logically. This method trains the AI without needing extra labels and uses three coordinated steps to help it first describe images accurately, then reason correctly, and finally answer questions well. Their approach makes these AI models much better at solving math problems involving pictures, even when tested on different types of models.
What this means in practice
- •For ai developers: Improve AI systems that handle mathematical problems involving images by separately training visual perception and reasoning capabilities.
- •For technology product teams: Build better educational or assistive tools that require visual math problem solving by integrating improved MLLM perceptual and reasoning training.
Authors
Yuzhe Li, Hao Yan, Hao Wang, Xingchen Liu, Ya-Qi Yu, Jihao Wu, Minghui Liao, Wei Chen, Yuliang Liu
Abstract
Multimodal Large Language Models (MLLMs) often struggle with complex mathematical visual reasoning primarily due to a lack of fine-grained perception, causing initial visual hallucinations to directly trigger cascading reasoning failures. In traditional end-to-end reinforcement learning (RL), sparse rewards fail to decouple perceptual hallucinations from logical missteps, hindering targeted perception optimization. Alternatively, fine-tuning with perception-enhanced CoT data incurs high costs and hallucinations. In this paper, we address these challenges by proposing UniCAR-RL, an annotation-free RL framework. By explicitly decoupling the optimization of perception and reasoning during the training process, it achieves isolation and optimization of both capabilities. Specifically, UniCAR-RL consists of three synergistic branches: 1) a Caption-RL branch that optimizes perception capabilities through verifier-guided reasoning validation; 2) a Reasoning-RL branch that performs logical reasoning based on a gold image description to halt cascading errors; 3) a QA-RL branch that retains native end-to-end alignment to ensure robust question-answering performance. Experiments show that UniCAR-RL substantially improves MLLMs' mathematical and visual reasoning using only raw short-answer data. Furthermore, it demonstrates strong generalization across diverse architectures and scales.