single mention · not yet filed
Continuous action-generation inference path
A continuous action-generation inference path generates a whole action chunk with a small number of fixed flow-matching steps in an action expert, rather than autoregressively decoding action tokens with the VLM.1 It avoids the large VLM decoding overhead of the discrete path, which is why the model uses it as the default deployment option.2
Trade-off It runs faster than the discrete action path, so the authors make it the default deployment option.3
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
the action expert emits the entire chunk in a small fixed number of flow matching steps through a smaller cross-attention head, with the VLM cache reused across steps
— MolmoAct2 Action Reasoning Models for Real-World DeploymentThis gap comes from the large VLM decoding overhead
— MolmoAct2 Action Reasoning Models for Real-World DeploymentWe therefore use the continuous path as the default deployment option
— MolmoAct2 Action Reasoning Models for Real-World Deployment
Also called continuous path.