Expediting Astronomical Discovery with AI Agents
Abstract: The expansive, interdisciplinary nature of astronomy, combined with its open-access culture, makes it an ideal testing ground for exploring how Large Language Models (LLMs) can accelerate scientific discovery. In this talk, I will present our recent advances in applying LLMs as autonomous research agents. We demonstrate that AI agents now achieve gold-medal performance on International Astronomy Olympiad problems, and that multi-agent frameworks like Mephisto can conduct end-to-end galaxy spectral fitting—iteratively refining physical models and accumulating knowledge through self-play, approaching human-like intuition and domain reasoning. However, the Moravec paradox manifests clearly: tasks requiring abstract calculation may be easier for AI than seemingly simple perceptual tasks like chart reading and visual reasoning, which remain key bottlenecks. To address the cost barrier at scale, we developed open-source specialized models (AstroSage) that match frontier performance on astronomy Q&A at a fraction of the cost. I will conclude by reflecting on the epistemological implications—what counts as knowledge and understanding when AI agents can reason but not yet perceive the way scientists do, and what this means for the future of astronomical discovery.
For help, please contact Webmaster.