Aeronix-zzz/VERSE-7B
VERSE-7B is a 7 billion parameter vision-language model developed by Aeronix-zzz, built upon Qwen2.5-VL-7B-Instruct with a 32768 token context length. It is uniquely trained using a self-evolve GRPO stage that incorporates an engine-verified VSI object-grounding process reward, enabling it to draw on images and perform spatial reasoning. This model excels at tasks requiring visual grounding and agentic think-draw-observe reasoning, making it suitable for benchmarks like VSI-Bench, SPAR-Bench, and MAZE.
Loading preview...
VERSE-7B Overview
VERSE-7B is a 7 billion parameter vision-language model developed by Aeronix-zzz, originating from the VERSE project. It is built upon the robust Qwen2.5-VL-7B-Instruct base model and distinguishes itself through a unique self-evolutionary training process.
Key Capabilities
- Spatial Reasoning and Object Grounding: VERSE-7B is specifically trained to understand and interact with spatial information within images. It can "draw on images" to map objects and trace paths.
- Self-Evolve GRPO Training: The model learns by proposing spatial questions, having a deterministic engine verify them, and then solving these questions through an agentic think-draw-observe reasoning loop. This process includes an engine-verified VSI object-grounding process reward.
- Agentic Reasoning: It employs a think\u2192draw\u2192observe reasoning paradigm, allowing it to iteratively refine its understanding and actions based on visual feedback.
Good for
- Vision-Language Tasks: Particularly strong in scenarios requiring detailed spatial understanding and interaction with visual elements.
- Research in Agentic VLMs: Ideal for researchers exploring self-evolving agents and reinforcement learning in vision-language models.
- Benchmarking: Designed to perform well on specialized benchmarks such as VSI-Bench, SPAR-Bench, and MAZE, which focus on spatial reasoning and object grounding.