Large Language Models in Space—and Beyond

“Our aim for the test was to improve quality within the LLM Retrieval Augmented Generation (RAG) system without increasing size or power draw. This would enable disconnected users to trust its responses and act on them. That’s a tall order for any AI, especially one so small,” says Booz Allen Principal Dan Wald, the AI solution architect who led the research effort.

The patch, uploaded on Halloween, delivered a treat: It installed instantly and successfully introduced a significant boost in quality improvements to the LLM RAG—improvements that HPE Chief Scientist Mark Fernandez, Ph.D., who leads Spaceborne Computer-2 research, assessed to be both quantitative and qualitative.

“We on the Spaceborne Computer team repeatedly demonstrated that the download of results from space to Earth can go from months to minutes using edge-of-the-edge Spaceborne Computer capabilities,” Mark says. “Similarly, with these modern hardware and software technologies, the upload of improved software to Spaceborne Computer offering increased capabilities can go from days to seconds.”

Booz Allen Chief Engineer Collin Paran, who leads LLM development for products like Booz Allen i2S2, described the gains: “The LLM was able to summarize information through dynamic scraping rather than simply finding and displaying a sentence.”

Dan summarizes the achievement: “Through collective ingenuity and a 22kb patch update, our model has eliminated false responses and hallucinations [fabricated or misleading information] and now reports true responses that include document references for further exploration by users.”

This patch, executed in real time, holds significant promise for multiple missions. In the cockpit, for example, our models could one day assist pilots by providing real-time insights, guiding decision making, and ultimately reducing the risk of accidents in a fast-moving, data-rich environment.

LLMs can also work together, Collin points out. He gives the example of satellite constellations. “An LLM could be on one satellite, another on a separate satellite, and a third on Earth, each with different expertise,” he says. “Think of the power of that collaboration.”

And our vision extends beyond aerospace. “From defense applications to emergency response in remote areas, our small, high-performance LLMs will one day be ready to support those who operate in some of the world’s—actually, our solar system’s—most challenging conditions,” says Dan. “Imagine AI-driven insights assisting medical triage for disaster response, supporting logistics in a far-off military outpost, or providing new insights for climate science.”