Hubble: Without any training, Gemini 3.1 Flash-Lite class 98.2% of the remarkable galaxies known in the 10% most promising images

Rather than looking for the rarest object, researchers ask a vision-language model what is scientifically interesting. Cost: about $5 per 10,000 images.

Anomaly detection algorithms identify statistically unusual objects in large astronomical surveys, but rarity does not always rhyme with scientific interest. So a team tested another approach: describe in natural language what deserves an astronomer's attention, and let a generalist vision-language model rate each image, without labeled examples.

Out of a sample of 20,000 Hubble Space Telescope images containing 167 reference anomalies, already identified by machine learning and human verification, Gemini 3.1 Flash-Lite places 98.2% of them among the 2,000 or so highest-rated images. Sorting costs about $5.05 for 10,000 images at current prices. The model includes gravitational lenses, "jellyfish galaxies" candidates and galaxies with visible jets.

The authors publish 102 new, highly-rated candidates for complementary observations, along with their code, and see it as a convenient way to prioritize huge sets of images based on their relevance rather than just their singularity.