ChenShawn/DeepEyes-7B
DeepEyes-7B by ChenShawn is a 7 billion parameter vision-language model that learns to "think with images" through end-to-end reinforcement learning, directly guided by outcome reward signals without supervised fine-tuning. It integrates visual information into its reasoning chain, showing improved grounding IoU and tool-calling accuracy. This model excels in high-resolution benchmarks, visual grounding, hallucination mitigation, and mathematical problem-solving tasks by enabling emergent visual search and comparison patterns.
Loading preview...
DeepEyes-7B: Vision-Language Model with Reinforcement Learning for "Thinking with Images"
DeepEyes-7B is a 7 billion parameter vision-language model developed by ChenShawn, designed to integrate visual information directly into its reasoning process. Unlike traditional methods, DeepEyes learns this capability through end-to-end reinforcement learning (RL), guided by outcome reward signals without requiring cold-start or supervised fine-tuning.
Key Capabilities & Innovations
- Reinforcement Learning for Visual Reasoning: DeepEyes' ability to "think with images" is learned directly from outcome rewards, eliminating the need for extensive supervised intermediate step training.
- Improved Visual Grounding and Tool-Calling: Despite no direct supervision for intermediate steps, RL training significantly increases both grounding Intersection over Union (IoU) and tool-calling accuracy.
- Enhanced Performance on High-Resolution Tasks: The end-to-end RL training leads to substantial performance gains on high-resolution benchmarks.
- Strong Generalization: Demonstrates robust generalization across various tasks, including visual grounding, hallucination mitigation, and mathematical problem-solving.
- Emergent Thinking Patterns: During RL training, the model develops sophisticated visual reasoning patterns, such as visual search for small objects, visual comparisons, and using
image_zoom_in_toolsfor answer verification.
Training Details
DeepEyes-7B uses Qwen-2.5-VL-7B-Instruct as its foundation model for RL training. The training framework is a general agentic RL system based on VeRL, supporting efficient agent RL training with dynamic multi-modal inputs and hybrid training for agentic and non-agentic data. It supports algorithms like PPO, GRPO, and reinforce++.
Use Cases
DeepEyes-7B is particularly well-suited for applications requiring advanced visual reasoning, precise object grounding, and complex problem-solving that benefits from integrating visual data directly into the thought process. Its ability to mitigate hallucinations and solve math problems with visual context makes it valuable for tasks demanding high accuracy in visual understanding.