OpenVLThinker: Complex Vision-Language Reasoning via Iterative SFT-RL Cycles

OpenVLThinker, introduced as one of the first open-source large vision-language models, showcases sophisticated chain-of-thought reasoning, achieving notable performance gains on challenging visual reasoning tasks. While traditional text-based reasoning models like Deepseek R1 have shown promise in text-only contexts, they often struggle when adapted to visual tasks due to imprecise visual grounding. OpenVLThinker addresses this by employing a novel approach that alternates between supervised fine-tuning (SFT) and reinforcement learning (RL). This iterative process not only surfaces latent reasoning behaviors in the model but also narrows the RL search space, leading to significant performance improvements. Specifically, OpenVLThinker improved metrics on benchmarks such as MathVista by 3.8%, EMMA by 2.4%, and HallusionBench by 1.6%. The model, with 7 billion parameters, represents a significant advancement in the field, indicating a promising direction for future AI developments that r…