From Routing Signals to Selective Review: Visual regrounding in MoE VLMs

Imagine you're building a vision-language model that needs to answer questions about objects in images, like their color or location. A common issue arises when the model confidently provides answers about objects that aren't even present in the image, leading to what’s known as a target-absence grounding failure. Current methods to address this often rely on analyzing the model's generated responses or its uncertainty, which can miss critical cues about the actual visual content. This is where the new approach comes in: by tapping into the internal routing decisions of Mixture-of-Experts (MoE) models, it detects when a target object is absent before generating an answer, allowing for a more informed response. The authors extract routing probabilities from specific models and train a linear detector to identify when to invoke a review prompt for better accuracy. This method not only improves the model's performance significantly on benchmark datasets but does so without altering the model's weights, making it a low-cost solution for enhancing reliability in real-world applications. Overall, this approach provides a way to utilize existing computation in MoE models to improve visual grounding, which is crucial for building more dependable AI systems.

The approach leverages internal routing decisions in MoE models, which is a novel method for detecting target absence.

The results show high ROC-AUC scores and significant improvements in accuracy, indicating solid reliability.