Open Models Almanac
Methodsinference & servingdecoding strategy

single mention · inference & serving

Autoregressive discrete action-token inference

Autoregressive discrete action-token inference is a decoding strategy in which the vision-language model autoregressively decodes action tokens for an entire action chunk.1 It runs slower than a continuous action path, whose action expert emits the whole chunk in a small fixed number of flow matching steps with a smaller cross-attention head.2

Trade-off It decodes actions with the full VLM and is slower than the continuous path, which is therefore used as 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.

  1. the VLM autoregressively decodes action tokens for the entire action chunk — MolmoAct2 Action Reasoning Models for Real-World Deployment
  2. the large VLM decoding overhead, while the action expert emits the entire chunk in a small fixed number of flow matching steps through a smaller cross-attention head — MolmoAct2 Action Reasoning Models for Real-World Deployment
  3. slower than the corresponding continuous path. This gap comes from the large VLM decoding overhead — MolmoAct2 Action Reasoning Models for Real-World Deployment

Also called discrete action path.

Filed under inference & serving › decoding strategy