A groundbreaking study by the University of Oxford, Google Cloud, and Radboud University—published in Nature Astronomy—has revealed how a Large Language Model (LLM) can be transformed into an advanced astronomy assistant with minimal training. Using Google’s Gemini, researchers demonstrated that this general-purpose AI could accurately identify and explain real cosmic phenomena such as exploding stars, stellar flares, and asteroids with roughly 93% accuracy, using only 15 example images and a few concise instructions. This achievement shows that complex, data-heavy model training is no longer essential for powerful scientific applications, marking a turning point toward transparent and accessible AI-driven discovery in astrophysics and beyond.
The innovation lies in Gemini’s ability to distinguish genuine celestial events from false signals—like satellite trails or cosmic ray artifacts—while clearly explaining its reasoning. Traditionally, astronomers relied on machine learning systems that acted as “black boxes,” offering no transparency into how decisions were made. The LLM overcomes this by combining visual and textual reasoning in a multimodal framework, allowing scientists to understand not just the outcome, but also the logic behind it. This interpretability builds trust in AI-generated findings and lowers barriers for non-specialists, empowering anyone with curiosity to engage in meaningful scientific exploration.
Researchers tested Gemini on three major sky surveys—ATLAS, MeerLICHT, and Pan-STARRS—supplying it with a small set of labeled examples. Each example included a live sky image, a reference frame, and a “difference” image showing the change, along with a short expert note. Guided by these few-shot examples, the LLM classified thousands of new alerts, assigning each a real/bogus label, confidence score, and plain-language explanation. A panel of 12 astronomers evaluated the outputs, finding the AI’s descriptions to be highly coherent and scientifically useful.
Crucially, the study introduced a “human-in-the-loop” feedback mechanism. Gemini was able to review its own answers, rating the coherence of its explanations and using that as a proxy for confidence. When the LLM detected low-coherence responses, it flagged those cases for human review—helping scientists focus on uncertain or ambiguous results. This self-evaluation process allowed the system to refine its classifications, improving performance from about 93.4% to 96.7%. The approach demonstrates how humans and AI can work synergistically, with the model learning and adapting through transparent interaction.
Professor Stephen Smartt of Oxford noted that this capability could revolutionize how astronomers process data from the next generation of telescopes, such as the Vera C. Rubin Observatory, which will generate 20 terabytes of data daily. Instead of opaque algorithms, scientists now have an intelligent partner that can reason and explain.
Looking ahead, the team envisions “agentic assistants” capable of integrating multiple data streams, self-assessing confidence, and autonomously coordinating telescope follow-ups. Since the approach requires only a small number of examples and plain-language guidance, it can be rapidly adapted to other scientific fields. As co-lead author Turan Bulmus from Google Cloud concluded, this research marks the beginning of a new era in which AI systems learn with us, not for us—transparent, collaborative, and deeply empowering for scientific discovery.
https://phys.org/news/2025-10-ai-advance-astronomers-cosmic-events.html