The workflow ingests New, Reference and Difference images per candidate and outputs a real/bogus decision, a concise text rationale and an interest score for follow-up triage, addressing the data deluge from surveys such as the Vera C. Rubin Observatory expected to produce roughly 20 terabytes per day.

"It's striking that a handful of examples and clear text instructions can deliver such accuracy," said Dr Fiorenzo Stoppa. Co-lead author Turan Bulmus said the approach "demonstrates how general-purpose LLMs can democratise scientific discovery."

A panel of 12 astronomers rated the AI's explanations highly coherent; by using the model's self-assessed coherence to flag uncertain cases for human review and refining the few-shot set, performance on one dataset improved from ~93.4% to ~96.7%. Professor Stephen Smartt noted the LLM's accuracy with minimal task-specific training "was remarkable," adding that scaling could be "a total game changer."

The team envisions agentic assistants that integrate imaging and photometry, check confidence, request robotic follow-up and escalate only the most promising events. Published 8 October in Nature Astronomy, the method can be rapidly adapted to new instruments and domains because it relies on small example sets and plain-language prompts.

Research Report:Textual interpretation of transient image classifications from large language models

Related Links
University of Oxford
Stellar Chemistry, The Universe And All Within It