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

The near-zero frontier model performance on ComCADBench isn't just a capability gap, it's a diagnostic: it suggests that scaling vision-language models trained on web data hits a hard ceiling when the task requires deterministic, structured program synthesis over domain-specific software APIs rather than perceptual pattern matching.

This connects directly to two threads in recent coverage. The MACCO work on visio-linguistic compositionality (covered same day) identified that vision-language models fail at structured relational reasoning, which is precisely the failure mode COM-as-Action is designed to route around by abandoning visual grounding entirely. Meanwhile, SkillCAT's framing of agent skill reuse as a bottleneck is relevant here too: if COM-based agents succeed, the question of how to accumulate and retrieve procedural software knowledge across tasks becomes the next unsolved problem. Neither prior paper addresses professional CAD workflows specifically, but together they sketch the same underlying diagnosis: end-to-end vision pipelines are structurally mismatched to tasks requiring precise, composable action sequences.

Watch whether any frontier lab (Anthropic, OpenAI, Google DeepMind) adopts ComCADBench as an official evaluation target within the next six months. Adoption would signal the benchmark has cleared independent reproducibility review; continued silence would suggest the near-zero scores reflect benchmark construction choices rather than a genuine capability ceiling.