Open Models Almanac
Methods

single mention · not yet filed

Embodiment-specific soft prompts

A soft-prompt mechanism that associates each hardware configuration with learnable prompt embeddings prepended to the model's input.1,2 The prompts absorb embodiment-specific variability so a shared backbone can be pretrained stably on heterogeneous data without overwriting its pretrained representations.3,4

Trade-off Adapts each embodiment with only a small set of extra parameters instead of finetuning the whole model.5,6

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. By associating each hardware configuration with learnable prompt embeddings — X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model
  2. a small set of learnable embeddings are prepended to the input sequence — X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model
  3. explicitly absorbs embodiment-specific variability while preserving a shared backbone for general reasoning — X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model
  4. without overwriting its pretrained representations — X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model
  5. a parameter-efficient alternative to full model finetuning — X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model
  6. with minimal additional parameters — X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model

Also called soft-prompt mechanism.