Zero-Shot Vision-Language Model Ranks Hubble Galaxy Images by Scientific Relevance
What Was Observed (Empirical Data)
Researchers evaluated pretrained vision-language models on a sample of 20,000 Hubble Space Telescope images containing 167 previously identified reference anomalies. Using natural-language prompts without target-specific training, Gemini 3.1 Flash-Lite placed 98.2% of the reference anomalies within the top ~2,000 ranked images at an inference cost of roughly $5.05 per 10,000 images. The ranking recovered known gravitational lenses, jellyfish-galaxy candidates, and active galactic jets, while flagging 102 additional candidates for follow-up.
The Reality Check (Anti-Hype Analysis)
The model does not independently discover new physical laws or confirm the astrophysical nature of the 102 newly flagged candidates, which still require spectroscopic or expert validation. Ranking images into a top 10% subset reduces human review workload by an order of magnitude, but it remains a triage filter rather than a replacement for physical modeling and empirical verification.
PRIMARY SOURCE:
arXiv Astrophysics (astro-ph)