---
license: apache-2.0
base_model:
  - Qwen/Qwen2.5-VL-7B-Instruct
pipeline_tag: image-text-to-text
library_name: transformers
tags:
  - multimodal
  - vision-language
  - reinforcement-learning
  - grpo
  - context-aware
  - visual-reasoning
---
# ContextRL-Qwen2.5-VL-7B
This is the **multimodal model** released with the paper
**Context-Aware RL for Agentic and Multimodal LLMs**.
It is fine-tuned from **[Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct)**
using **ContextRL**, a context-aware reinforcement learning method that augments
standard GRPO with an auxiliary *context-selection* objective to improve fine-grained
visual grounding.
## Results
Across 12 diverse multimodal benchmarks, ContextRL improves over the standard GRPO
baseline by **+2.0 points** on average, while improving every individual benchmark.
| Benchmark | Base | RL (GRPO) | ContextRL (Ours) |
|---|---|---|---|
| MathVista | 68.2 | 72.5 | **73.6** |
| MathVerse | 43.9 | 45.3 | **49.1** |
| MathVision | 22.8 | 25.5 | **26.8** |
| MMMU-Pro | 36.6 | 41.3 | **42.8** |
| MMMU | 50.7 | 53.3 | **54.6** |
| V* | 70.1 | 70.7 | **73.3** |
| MMStar | 62.6 | 64.1 | **65.1** |
| BLINK | 55.3 | 56.5 | **58.9** |
| ScienceQA | 88.2 | 91.0 | **95.4** |
| PhyX | 25.4 | 48.7 | **50.0** |
| OlympiadBench Phy | 1.5 | 3.1 | **4.6** |
| MME-RealWorld Lite | 38.4 | 45.1 | **46.7** |
| **Overall Avg.** | 47.0 | 51.4 | **53.4** |

The +2.0 average gain over GRPO also exceeds the +0.8 of PAPO, a method purpose-built for multimodal perception (see the paper for the full comparison).
## Usage
This model follows the same interface as Qwen2.5-VL-7B-Instruct and can be loaded with
`transformers`. Training and evaluation code, data construction pipelines, and detailed
configurations are available in the repository:
👉 **https://github.com/xupy2003/ContextAwareRL**
Please refer to the repo's README for environment setup, inference scripts, and
reproduction instructions.