AI Learns to Identify Exploding Stars with Just 15 Examples (www.universetoday.com)

An international team tested Google’s LLM Gemini as a few-shot classifier for astronomical transients, showing that with just 15 labeled examples and task prompts it can triage night-sky images from three survey datasets (ATLAS, MeerLICHT, Pan-STARRS). Using prompt instructions to label candidates as “No interest”, “Low interest” (variable stars) or “High interest” (explosive events/real transients), and sharing their prompts and examples publicly (github.com/turanbulmus/spacehack), the researchers re-ran the experiment after a six-month Gemini update and reported accuracies of 91.9% (ATLAS), 93.4% (MeerLICHT) and 94.1% (Pan-STARRS). The result is significant because it demonstrates that a generalist LLM, with minimal task-specific training, can perform image-based transient triage at near–state-of-the-art levels, potentially cutting the human time and custom-model engineering needed to sift through false positives in large sky surveys. Technical implications include using LLMs for rapid, low-cost pre-screening of alerts (few-shot prompting instead of full supervised retraining), easier reproducibility via shared prompt repos, and broader democratization of survey analysis for citizen scientists and smaller teams. Scaling and integration with specialized pipelines, robustness to domain shifts, and systematic benchmarking against purpose-built CNNs remain key next steps.