VLM-R1: a stable and generalizable R1-style vision-language model trained with GRPO for multimodal reasoning

What it solves

VLM-R1 addresses the challenge of creating stable and generalizable Large Vision-Language Models (LVLMs) that can reason through visual tasks. It specifically aims to improve performance on out-of-domain data for tasks like Referring Expression Comprehension (REC), Open-Vocabulary Detection (OVD), and multimodal math, where traditional Supervised Fine-Tuning (SFT) often fails to generalize as well as reinforcement learning approaches.

How it works

The project implements an "R1-style" training approach, primarily using Group Relative Policy Optimization (GRPO) to train models like Qwen2.5-VL and InternVL. Unlike SFT, which mimics specific labels, GRPO allows the model to develop reasoning capabilities through reward-based learning. The system supports full fine-tuning, LoRA, and multi-node training, and can handle both single and multi-image inputs to solve complex grounding and reasoning tasks.

Who it’s for

This project is for AI researchers and developers working on multimodal reasoning, object detection, and visual grounding who want to implement reinforcement learning (specifically GRPO) on vision-language models to achieve better generalization.

Highlights

• State-of-the-Art Performance: Achieves top results on the Open-Compass Math Leaderboard (under 4B parameters) and SOTA performance on OVDEval.
• Versatile Task Support: Specialized models available for Open-Vocabulary Detection (OVD), Multi-Modal Math, Referring Expression Comprehension (REC), and GUI Defect Detection.
• Flexible Training: Supports LoRA, full fine-tuning, and multi-node training across various VLMs including QwenVL and InternVL.
• Hardware Optimization: Optimized for Huawei Ascend Atlas series using vllm-ascend and xllm frameworks for reduced TTFT and increased throughput.