ScienceOne-AI/S1-VL-32B-RL

S1-VL-32B-RL: Scientific Multimodal Reasoning Model

S1-VL-32B-RL, developed by the ScienceOne AI team at the Chinese Academy of Sciences, is a 33.4 billion parameter multimodal large language model optimized for scientific applications. It integrates two core reasoning paradigms: Scientific Reasoning for complex, multi-step problem analysis, and Thinking with Images, which enables the model to invoke code tools for image manipulation (cropping, zooming, enhancement, annotation) during inference.

Key Capabilities

• Scientific Multimodal Reasoning: Achieves state-of-the-art performance across diverse scientific benchmarks including MMMU, MathVision, and VRSBench-MINI, covering mathematics, physics, chemistry, astronomy, earth sciences, and biology.
• Image Operation Reasoning: Ranks first across five benchmarks (HRBench-4K, HRBench-8K, MME-RealWorld-CN, MME-RealWorld-Lite, V*), demonstrating superior ability in high-resolution image understanding and real-world visual reasoning.
• Code Tool Invocation: Can proactively use code to enhance visual information, as demonstrated in case studies involving medical imaging where it crops and magnifies regions of interest for clearer analysis.
• Progressive Post-training: Utilizes a four-stage training procedure, including Scientific Reasoning SFT, Thinking-with-Images Cold-Start SFT, and two stages of Reinforcement Learning (RL) with the SAPO algorithm, to progressively unlock and optimize its scientific reasoning and image operation capabilities.

Good For

• Analyzing and solving complex scientific problems involving both text and images.
• Interpreting dense scientific charts, high-resolution remote sensing, microscopic, and astronomical images.
• Applications requiring dynamic image manipulation during the reasoning process, such as medical image analysis and academic figure Q&A.