Tool-Augmented Agent for Closed-loop Optimization,Simulation,and Modeling Orchestration
The article discusses the introduction of COSMO-Agent, a tool-augmented reinforcement learning framework aimed at optimizing the CAD-CAE process. This framework addresses the challenges of translating simulation feedback into geometric edits while adhering to various constraints. Experimental results indicate that COSMO-Agent significantly enhances the performance of small open-source LLMs in constraint-driven design tasks.
- ▪COSMO-Agent is designed to bridge the CAD-CAE semantic gap in industrial design-simulation optimization.
- ▪The framework utilizes a multi-constraint reward system to ensure feasibility and robustness in toolchain operations.
- ▪Training with COSMO-Agent has shown to outperform both large open-source and closed-source models in terms of feasibility, efficiency, and stability.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.20190 (cs) [Submitted on 1 Apr 2026] Title:Tool-Augmented Agent for Closed-loop Optimization,Simulation,and Modeling Orchestration Authors:Liyuan Deng, Shujian Deng, Yongkang Chen, Yongkang Dai, Zhihang Zhong, Linyang Li, Xiao Sun, Yilei Shi, Huaxi Huang View a PDF of the paper titled Tool-Augmented Agent for Closed-loop Optimization,Simulation,and Modeling Orchestration, by Liyuan Deng and 8 other authors View PDF HTML (experimental) Abstract:Iterative industrial design-simulation optimization is bottlenecked by the CAD-CAE semantic gap: translating simulation feedback into valid geometric edits under diverse, coupled constraints.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
