Agentic Design Of Compositional Machines

The task of compositional machine design is illustrated in our BesiegeField environment. The figure shows a high-level sketch of the agentic workflow (w/ Gemini Pro 2.5), along with the resulting machines and their simulated performance. The design objective is to create a machine that throws boulders long distances.

We demonstrate the motion mechanism of a human-designed trebuchet — a powerful medieval type of catapult. Each component works in close coordination, enabling the machine to launch the projectile much farther than those built by LLMs. We also show how omitting even a single part can cause the entire mechanism to fail, highlighting the inherent difficulty of compositional machine design.

🔍 Click to zoom

Example CoT of inspector agents (w/ Gemini 2.5 Pro). Blue text highlights the moderate capability of LLMs in spatial reasoning and imagined physical simulation.

Generated by agentic systems with different LLMs.

Reinforcement learning with verifiable rewards (RLVR) improves Qwen-2.5-14B model in compositional machine design.

@article{zhang2025besiegefield,
  title={Agentic Design of Compositional Machines},
  author={Zhang, Wenqian and Liu, Weiyang and Liu, Zhen},
  journal={arXiv preprint arXiv:2510.14980},
  year={2025},
}