Guiding Generative AI Solution for Fundamental Astrodynamics Problems

Large Language Models (LLMs) as generative AIs have seen significant progress and adoption in space engineering. When it comes to methodological innovation, the fundamental question is how to quantify LLM quality; the logical next step is to improve performance further. Moreover, because interpretability is a crucial requirement in space engineering, it is imperative to make LLMs explainable so they can reliably support critical tasks. In addition to introducing the first benchmark dataset for testing LLMs, we experiment with the in-context learning (ICL) technique of in-context vectors (ICV) and fractional reasoning (FR) to enhance typically subpar model performance and to provide a training-free and model-agnostic framework. We also introduce a new way to leverage the dataset’s existing solutions in building novel ICVs to steer reasoning. Finally, we present the first attempt to apply ICL through generative AI to solve fundamental astrodynamics tasks within the benchmark.