Bridging Inference and Adaptation in Vision-Language Models via Optimal Transport
August 20, 2026
A new Test-Time Adaptation (TTA) method for VLMs uses optimal transport to bridge the gap between inference and adaptation objectives. This approach avoids the noise amplification common in existing methods that rely on unreliable pseudo-labels from raw embedding similarities.
HOW THIS AFFECTS YOU
●
builderYou can improve VLM robustness to real-world distribution shifts during inference without noisy pseudo-labels.
●
researcherThe optimal transport formulation provides a more principled way to model sample-level cross-modal relationships.