Large vision-language models (LVLMs) that combine image and text understanding often produce hallucinations—outputs that describe objects or relationships that do not exist in the image. These errors can take the form of fabricated objects, incorrect attributes, or wrong spatial relations.
To address this, researchers propose PLANE, a plug-and-play method that integrates fine-grained visual experts into the model. The approach is designed to mitigate hallucination without requiring extensive retraining, as suggested by the plug-and-play label.
The work is positioned as a response to previous empirical studies that have indicated likely causes or patterns of hallucination. By leveraging specialized visual knowledge, PLANE aims to correct these errors at inference time.