Dr. Seg: Revisiting GRPO Training for Visual Large Language Models through Perception-Oriented Design

Researchers introduce Dr. Seg, a new framework that improves Group Relative Policy Optimization (GRPO) training for Visual Large Language Models by addressing key differences between language reasoning and visual perception tasks. The framework includes a Look-to-Confirm mechanism and Distribution-Ranked Reward module that enhance performance in complex visual scenarios without requiring architectural changes.

• →Current GRPO training methods designed for language models don't transfer seamlessly to visual perception tasks in VLLMs.
• →Two critical factors for visual perception were identified: broader output space requirements and fine-grained stable rewards.
• →Dr. Seg is a plug-and-play framework that integrates with existing GRPO-based VLLMs without architectural modifications.
• →The framework demonstrates improved performance in complex visual scenarios while maintaining strong generalization capabilities.
• →Research focuses on reasoning segmentation as a representative case for visual perception optimization.