GUIDE: Guided Updates for In-context Decision Evolution in LLM-Driven Spacecraft Operations

A team from MIT and the University of Minnesota has introduced GUIDE (Guided Updates for In-context Decision Evolution), a novel framework designed to solve a critical flaw in using Large Language Models (LLMs) for autonomous spacecraft operations. Current LLM agents rely on static prompts, meaning they don't learn or adapt from repeated mission executions. GUIDE enables cross-episode adaptation without the computational cost and risk of updating the model's core weights, a process known as non-parametric policy improvement.

The system operates with a dual-model architecture. A lightweight 'acting' LLM handles real-time, closed-loop control of the spacecraft. Separately, an offline 'reflection' process analyzes prior mission trajectories to evolve a structured, state-conditioned library of natural-language decision rules—essentially a dynamic playbook. This playbook is then provided as context to the acting model for future missions. The researchers validated GUIDE using an adversarial orbital interception scenario within the Kerbal Space Program Differential Games simulator, where it demonstrated consistent performance gains over static LLM baselines.

The results indicate that GUIDE effectively performs policy search over interpretable decision rules. This evolution in context allows the AI agent to refine its strategies for complex, real-time spacecraft interactions, such as orbital rendezvous or interception, based on historical success and failure. The work, accepted to the AI4Space workshop at CVPR 2026, represents a significant step toward more adaptive and reliable AI supervisors for critical space operations where traditional retraining is impractical.