Evaluating and Enhancing Negation Comprehension in Remote Sensing MLLMs

The article introduces RS-Neg, the first benchmark for evaluating negation comprehension in Multimodal Large Language Models (MLLMs) applied to Remote Sensing tasks. It details an automated data generation pipeline for synthesizing negation queries and a dynamic visual focus module for verification, revealing that existing advanced RS MLLMs struggle with negation, leading to hallucinations and performance drops. To address this, the authors propose NeFo, a test-time learning method that enhances negation understanding using a small set of unlabeled test samples, demonstrating significant improvements and generalization to new tasks.