single mention · model architecture
Sequential depth, trajectory, and action token generation
A prediction-head procedure in which the model autoregressively generates three token sequences in a fixed order: depth perception tokens, then a visual reasoning trace, then action tokens.1 Each stage is conditioned on the tokens generated before it, so the final actions are grounded in the inferred depth and the planned motion sketch.2
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
the model autoregressively generates three token sequences in order: (i) depth perception tokens ; (ii) visual reasoning trace ; (iii) action tokens
— MolmoAct Action Reasoning Models that can Reason in SpaceBy conditioning each stage on the preceding depth and trajectory tokens, we ensure that the final actions are spatially grounded in both the inferred depth and the planned motion sketch
— MolmoAct Action Reasoning Models that can Reason in Space
Also called Action Reasoning Procedure.
Filed under model architecture › prediction head