DeciCAD introduces a domain‑native CAD harness that exposes semantic geometric actions, decision‑oriented observations, environment feedback, and persistent state transitions to the model. By constraining the agent's decision space to professional CAD actions, the learning process can directly exploit geometric constraints and long‑horizon procedures.
The authors build a learning pipeline that first extracts causal decision supervision from interaction trajectories, then improves the policy via supervised fine‑tuning (SFT) and geometry‑grounded online reinforcement learning (RL). SFT raises $F1$ from $0.6883$ to $0.7437$, while online RL increases geometry availability from $66.5\%$ to $73.5\%$ and pushes $F1$ further to $0.7880$.
Across four CAD benchmarks, DeciCAD achieves the highest $F1$ and intersection‑over‑union (IoU) scores, and reduces model output tokens by $45\%\sim88\%$ compared to a general‑purpose coding harness. These findings demonstrate that the decision space induced by a harness materially impacts both execution efficiency and policy learning in CAD agents.
Review