Prompt Tuning / Soft Prompt

Freezing the whole model and training only a short learnable vector prepended to the input to adapt it to a task.

Prompt tuning is a parameter-efficient fine-tuning method proposed by Google's Lester and colleagues in 2021. A hand-written text prompt is a “hard prompt”; a soft prompt, by contrast, is a sequence of vectors learned directly in embedding space that does not correspond to any actual words. During training, the model's parameters stay entirely frozen, and only this short vector sequence is updated by gradient descent, so each task needs only a tiny soft prompt stored for it, and one model can serve many tasks. The paper found that as models get larger, performance gets closer to full fine-tuning, essentially matching it at billions of parameters. Prefix tuning, which prepends a learnable prefix before every attention layer, and visual prompt tuning are related ideas. In robotics, soft prompts have also been used to mark different embodiments or data sources.

ExampleX-VLA gives each data source — different robot embodiments and collection setups — its own learnable soft-prompt embedding on top of a standard Transformer backbone, letting a 0.9-billion-parameter model train across embodiments and get tested on multiple simulated and real-robot platforms.

Also called

Soft Prompt, Soft Prompt Tuning

Related

Parameter-Efficient Fine-Tuning · Prompt / Prompt Engineering · LoRA · Adapter · X-VLA · Embedding

Sources

Lester, Al-Rfou, Constant 2021: The Power of Scale for Parameter-Efficient Prompt Tuning
Zheng et al. 2025: X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment VLA Model

As of

2025-10