arxivcs.CVcs.CL2026-07-15
SD-MAR: Multi-image Analytical Reasoning via Synthetic Data and Reinforcement Learning
Shiyu Yuan, Sourav Sanjukta Bhabesh, Zhe Wang, Dmitriy Bespalov, Wesley Rose, Huzefa Rangwala
Vision Language Models (VLMs) demonstrate strong perceptual abilities but remain limited in tasks requiring analytical reasoning across multiple visual states, such as multi-image comparison, change detection, and multi-step visual inference. These capabilities are critical for r…