**Ok, So what they are saying tis that they They helped in solving the hallucination in VLM. The problem why hallicination exist in VLM is current approach tryies to fuse vision embedding subspace with text subspace. And there is a problem of misaligning both subspaces. Previous work has tried to aggressively fuse them, costly post decoding strategies Without solving the fundamental issue.
The fundamental issue they identified is the misalignment concentrates in the principal components of “universal data agnostic text subspace”. 

Their method to solve is a non finetuning denoising technique and finetuning 

Limitations:
Current decoder based models are fundamentally flawed. Compressing Visual fine grained space into text manifold is a bottleneck. 

In Future, we should rely on modality specific encoders to preserve semantic integrety.**