← Back to the feed
AI & softwareResearch findings

Primitive representations enable few-shot transfer in new tasks

Original published 2026.05.29 09:37 KSTAdded to Blackdeck 2026.11.22 20:38 KST

What has changed?

Models trained with primitive awareness achieve task transfer with only three demonstrations

Key points

  1. Models with primitive-aware training can perform new tasks with only three demonstrations
  2. Primitive-trained models reach 78% of fine-tuned performance with just three demonstrations, while flat-trained models require ten
  3. Ablation of primitive-decodable subspaces reduces transfer performance by 32 percentage points

Connect it to your work

Task transfer is possible with minimal demonstrations

What remains unproven

This research note is based on an abstract; check the original for test conditions and application scope.

Originals used for this note

Links open the original publisher in a new tab.

  1. arxivPrimitive Subspaces Mediate Few-Shot Transfer in VLAs