PyVision-RL: Forging Open Agentic Vision Models via RL
This addresses a bottleneck for developing scalable multimodal agents, though it appears incremental as it builds on existing RL and multimodal methods.
The paper tackles the problem of interaction collapse in reinforcement learning for agentic multimodal models, where models reduce tool usage and multi-turn reasoning, by introducing PyVision-RL, a framework that stabilizes training and sustains interaction, resulting in strong performance and improved efficiency for image and video understanding.
Reinforcement learning for agentic multimodal models often suffers from interaction collapse, where models learn to reduce tool usage and multi-turn reasoning, limiting the benefits of agentic behavior. We introduce PyVision-RL, a reinforcement learning framework for open-weight multimodal models that stabilizes training and sustains interaction. Our approach combines an oversampling-filtering-ranking rollout strategy with an accumulative tool reward to prevent collapse and encourage multi-turn tool use. Using a unified training pipeline, we develop PyVision-Image and PyVision-Video for image and video understanding. For video reasoning, PyVision-Video employs on-demand context construction, selectively sampling task-relevant frames during reasoning to significantly reduce visual token usage. Experiments show strong performance and improved efficiency, demonstrating that sustained interaction and on-demand visual processing are critical for scalable multimodal agents.