πReVisual-R1 is a 7B open-source multimodal language model that follows a three-stage curriculumβcold-start pre-training, multimodal reinforcement learning, and text-only reinforcement learningβto achieve faithful, concise, and self-reflective state-of-the-art performance in visual and textual reasoning.
reinforcement-learning visual-reasoning mathematical-reasoning data-efficiency multimodal-large-language-model prioritized-advantage-distillation cold-start-initialization efficient-length-reward open-source-7b-model self-reflective-chain-of-thought
-
Updated
Oct 13, 2025 - Python