ComAct: Reframing Professional Software Manipulation via COM-as-Action Paradigm

The paper introduces the COM-as-Action paradigm, which reframes professional software manipulation through the Component Object Model (COM) for deterministic program synthesis, addressing limitations in GUI and API-based agents. It presents ComCADBench, a benchmark for evaluating agents in industrial CAD environments, revealing that COM-based execution significantly outperforms GUI methods. The authors develop ComActor, a self-correcting agent that achieves state-of-the-art performance on ComCADBench and demonstrates resilience in long-horizon tasks, highlighting its potential for improving software interaction in complex environments.