AINeutralarXiv – CS AI · 10h ago6/10
🧠
Beyond Templates: Revisiting Zero-Shot Remote Sensing through Meta-Prompting
Researchers analyze how vision-language models perform zero-shot remote sensing tasks across multiple datasets and find that textual design choices critically impact performance. The study reveals that semantically rich LLM-generated descriptions don't consistently outperform simpler template-based descriptions due to noise in text embeddings, but lightweight query embedding calibration effectively improves results.