jihochoi/DR-MV3D-R-GRPO
DR-MV3D-R-GRPO by jihochoi is a 3 billion parameter multi-view 3D reasoning model, fine-tuned from Qwen2.5-VL-3B-Instruct. It specializes in answering questions about 3D scenes presented from multiple viewpoints, utilizing global maps and local views. The model was optimized using GRPO with dense, verifiable rewards, making it particularly effective for complex spatial reasoning tasks.
Loading preview...
DR-MV3D-R-GRPO Overview
DR-MV3D-R-GRPO is a 3 billion parameter model developed by jihochoi, designed for multi-view 3D reasoning. It is an official checkpoint from the paper "Dense Reward for Multi-View 3D Reasoning with Global Maps and Local Views" (ECCV 2026). This model is a fine-tuned version of Qwen2.5-VL-3B-Instruct, specifically optimized for understanding and answering questions about 3D scenes from various perspectives.
Key Capabilities
- Multi-View 3D Reasoning: Processes information from multiple viewpoints to construct a comprehensive understanding of a 3D scene.
- Map-Grounded Reasoning: Utilizes both global and per-view ego-centric maps to facilitate complex spatial reasoning.
- Structured Output: Produces a cognitive map, per-view maps, a reasoning chain, and a final answer, reflecting its structured approach to problem-solving.
- GRPO Optimization: Enhanced through Global Reward Policy Optimization (GRPO) with dense, verifiable rewards for improved performance.
Good For
- Complex 3D Scene Understanding: Ideal for applications requiring detailed analysis and question-answering based on multi-view 3D data.
- Robotics and Autonomous Navigation: Potentially useful for systems that need to interpret and reason about their environment from diverse sensor inputs.
- Research in Vision-Language Models: Serves as a strong baseline and advanced model for researchers exploring dense reward mechanisms and multi-view reasoning in VLMs.