PhyDrawGen: Physically Grounded Diagram Generation from Natural Language

PhyDrawGen is a neuro-symbolic AI system that generates physics diagrams from natural language text while maintaining strict physical accuracy. By combining large language models, deterministic solvers, and vision-language models in a pipeline, it overcomes the hallucination problems of current generative models and outperforms GPT-4, Gemini 2.5, and Gemini 3 Pro on physics problems spanning mechanics, optics, and electromagnetism.

PhyDrawGen addresses a critical limitation in current generative AI: the inability to enforce hard constraints from physics when producing visual outputs. While models like GPT-4 and Gemini can generate superficially plausible diagrams, they systematically violate conservation laws, ignore force balance, and misrepresent geometric relationships. This research demonstrates that decoupling semantic understanding from constraint satisfaction significantly improves accuracy on domain-specific tasks requiring rigorous rule compliance.

The neuro-symbolic approach represents a broader shift in AI architecture away from end-to-end learning toward hybrid systems that combine neural networks with symbolic reasoning. Rather than expecting a single model to simultaneously understand language, physics, and visual composition, PhyDrawGen distributes responsibilities: an LLM handles semantic parsing into a typed scene graph, a deterministic solver enforces exact physical constraints, and a fine-tuned vision model iteratively verifies outputs. This design pattern mirrors emerging best practices in autonomous reasoning and scientific computing where accuracy matters more than end-to-end simplicity.

For academic and educational technology markets, PhyDrawGen enables reliable diagram generation for physics education platforms, potentially reducing manual authoring costs. For AI developers, the work validates that foundation models perform better when combined with domain-specific solvers rather than asked to simultaneously learn and enforce domain constraints. The benchmark evaluation on 1,449 problems provides concrete evidence of this approach's superiority, suggesting similar hybrid architectures could improve performance in other constraint-heavy domains like molecular design, circuit layout, and architectural planning.

• →Neuro-symbolic pipelines outperform pure neural approaches on physics-constrained diagram generation by decoupling semantic understanding from physical constraint enforcement.
• →PhyDrawGen successfully handles unusual objects and edge cases better than state-of-the-art commercial models, indicating robust generalization beyond training data.
• →The deterministic solver component converting scene graphs to Planar Straight-Line Graphs provides an interpretable, verifiable alternative to purely learned constraint satisfaction.
• →This architecture pattern could extend to other domain-specific generation tasks requiring strict adherence to mathematical or physical rules beyond current model capabilities.
• →Educational technology and scientific publishing could benefit significantly from reliable automated diagram generation that maintains physical accuracy.