New framework enables LLMs to improve spacecraft control decisions through learning from past missions without retraining

GUIDE is a non-parametric framework that allows LLM-based spacecraft supervisory agents to adapt and improve across multiple missions without updating model weights

Uses offline reflection to evolve a playbook of natural-language decision rules based on prior trajectories, enabling real-time policy improvement

Tested on adversarial orbital interception tasks in Kerbal Space Program, consistently outperforming static prompting approaches

Demonstrates that in-context learning in LLMs functions as policy search over structured decision rules for closed-loop spacecraft operations