HuggingFace Papers
Domain Arithmetic: One-Shot VLA Adaptation under Environmental Shifts
agentic
What happened
This paper introduces Domain Arithmetic, a method for adapting Vision-Language-Action (VLA) models to new environments using a single demonstration. It leverages weight vector arithmetic to isolate and apply domain-specific information via subspace alignment, bypassing the need for heavy fine-tuning.
Why it matters
It demonstrates that complex physical agents can be adapted to new environments using a single demo and simple weight arithmetic rather than expensive retraining.
The take
Weight arithmetic (similar to model merging or task vectors) is proving to be a highly efficient way to patch models. Applying this to VLAs for one-shot environmental adaptation is a clever way to bypass the massive data requirements of robotics, though its utility is currently confined to physical agent developers.
Do this
If you are building VLA or robotics agents, read this paper to understand how to apply weight-space arithmetic for fast domain adaptation.
Don't read this site daily. Get it in your inbox.
The daily brief and Sunday deep dive — distilled, scored, and opinionated. For builders only.