COSMO-Agent teaches LLMs to close the CAD-CAE design loop
Small open-source LLMs beat GPT-4 in industrial design optimization using RL
Deep Dive
Researchers propose COSMO-Agent, a tool-augmented RL framework that trains LLMs to automate CAD-CAE closed-loop optimization. It teaches models to generate geometry, run simulations, parse results, and revise designs until constraints are met. Using a multi-constraint reward and an industry dataset of 25 component categories, small open-source LLMs improved substantially and exceeded large open-source and strong closed-source models in feasibility, efficiency, and stability.
Key Points
- COSMO-Agent uses RL to teach LLMs complete CAD-CAE loop: geometry generation, simulation, result parsing, iterative revision.
- Achieves 2-3x higher constraint satisfaction compared to untrained models, with small LLMs outperforming GPT-4o on industrial benchmarks.
- Includes a new dataset of 25 component categories with curated executable CAD-CAE tasks for realistic training and evaluation.
Why It Matters
Automates time-consuming design-simulation iterations, enabling faster product development and reduced engineering overhead.