SD-MAR: Multi-image Analytical Reasoning via Synthetic Data and Reinforcement Learning

arXiv cs.CL · 6 hours ago

Researchers introduced SD-MAR, a framework using synthetic data and reinforcement learning to improve vision language models' ability to perform analytical reasoning across multiple images, such as comparing visual states and detecting changes. Fine-tuning Qwen2.5-VL-7B and InternVL3-8B with GRPO-lite training achieved up to 36.95% improvement in accuracy on the SD-MAR benchmark, with Qwen2.5-VL-7B surpassing GPT-4.1. The trained models maintained or improved performance on existing benchmarks while gaining stronger multi-image reasoning capabilities for real-world applications requiring visual comparison and inference.