SD-MAR boosts VLM multi-image reasoning by 37%, beating GPT-4.1

Vision Language Models (VLMs) like GPT-4V and Gemini excel at single-image understanding but struggle when reasoning across multiple images—think comparing two photos to spot differences, tracking changes over time, or performing multi-step visual inference. This limitation hampers real-world applications in fields like medical imaging, autonomous driving, and surveillance, where analytical reasoning across visual contexts is critical. To address this gap, researchers from a team led by Shiyu Yuan developed SD-MAR (Synthetic Data for Multi-image Analytical Reasoning), a framework that generates paired visual scenarios through controlled perturbations and creates reasoning tasks requiring semantic change attribution and quantitative comparison.

The core innovation lies in the training methodology: GRPO-lite with Backward Discounted Allocation (BDA). This reinforcement learning approach removes traditional KL regularization to encourage stronger policy optimization and allocates greater credit to later reasoning steps—where analytical conclusions are formed. The team fine-tuned Qwen2.5-VL-7B and InternVL3-8B on SD-MAR data, achieving up to 36.95% accuracy gains on in-domain tasks. Remarkably, the Qwen2.5-VL-7B model outperformed GPT-4.1 on the SD-MAR benchmark, demonstrating that smaller, specialized models can surpass larger general-purpose ones for multi-image reasoning.

Out-of-domain generalization was preserved or even improved: performance on established benchmarks like MME, MMMU-Pro, and MathVista remained within 1% of baseline, while MMBench saw up to 4% improvement. An LLM-as-judge evaluation confirmed consistent gains in logical coherence and explanation quality. The work opens the door for cost-effective, high-performance multi-image reasoning without relying on massive proprietary models, making advanced visual analytics more accessible to developers and researchers.