PhyDrawGen: AI Generates Physics Diagrams from Text
A group of researchers has developed PhyDrawGen, a new framework that converts natural language into physics diagrams while following the laws of physics. This model stands out because it doesn’t confuse force directions or violate conservation laws like some others do. First, it uses a large language model to create a typed scene graph from the input text. Then, a specific solver changes this into a Planar Straight-Line Graph (PSLG) that accurately depicts forces, light paths, and field structures. Lastly, the Qwen-VL model fine-tunes the output through a visual feedback loop to fix any rule violations. PhyDrawGen was tested on 1,449 physics problems and outperformed both GPT-5-image and Gemini 2.5 Flash.
Key facts
- PhyDrawGen is a neuro-symbolic pipeline for generating physics diagrams from text.
- It decouples semantic scene understanding from physical constraint satisfaction.
- A large language model extracts a typed scene graph from the problem text.
- A deterministic solver converts the graph into a Planar Straight-Line Graph (PSLG).
- PSLG encodes force balance, optical paths, and field topologies as geometric primitives.
- A fine-tuned Qwen-VL model implements a propose-verify loop to correct violations.
- Evaluated on 1,449 problems from mechanics, optics, and electromagnetism.
- Outperforms GPT-5-image and Gemini 2.5 Flash.
- The paper is available on arXiv with ID 2605.30512.
- The approach addresses hallucination of force vectors and violation of conservation laws.
Entities
Institutions
- arXiv