Open Models Almanac
Methodsoptimizationtraining stability

single mention · optimization

Two-step prompt-and-head warm-up adaptation

A two-step adaptation procedure that first warms up new learnable prompts while keeping pretrained weights frozen, then jointly optimizes the backbone and warmed-up prompts.1,2 The warm-up projects prompts onto pretrained embodiment-agnostic features, giving a good starting point before full joint training specializes the policy to a new domain.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 prompt is first warmed up while keeping the pretrained weights frozen — X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model
  2. we jointly optimize both the backbone and the warmed-up prompt — X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model
  3. prompts are projected to exploit pretrained embodiment-agnostic features, offering good starts for next-round joint training — X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model

Also called two-step adaptation procedure.

Filed under optimization › training stability