Shorten the SA-CUA privileged-model system prompt

Determine whether the long system prompt used by the SA-CUA privileged large language model can be shortened through fine-tuning on task trajectories that teach the model to operate with masked screenshots.

Background

SA-CUA relies on a privileged LLM (P-LLM) that operates on screenshots in which untrusted regions are masked and uses specialized query, endorsement, and policy-gating tools. The framework currently requires a long system prompt to explain these interaction conventions and to guide the P-LLM in handling masked visual information.

The paper identifies reducing this prompt through fine-tuning on SA-CUA trajectories as unresolved. A shorter prompt could reduce token cost and potentially make the framework more practical, while preserving the P-LLM's ability to work safely with masked screenshots and endorsed values.

References

Further, it is an open question whether the long system prompt of the P-LLM could be shortened, by fine-tuning on trajectories to teach the model to work with masked screenshots.

— Towards Safe Autonomy for Computer Use  (2610.18508 - Seip et al., 13 Oct 2026) in Section 5, Conclusion and Discussion, Limitations