ClipTTT: CLIP-Guided Test-Time Training Helps LVLMs See Better

TL;DR

This paper introduces ClipTTT, a test-time training method guided by CLIP, which adapts large vision-language models to reduce hallucinations caused by visual corruptions in real-time.

Contribution

The paper presents a novel CLIP-guided test-time training approach that enhances LVLM robustness against corrupted visual inputs without modifying the original models.

Findings

ClipTTT significantly reduces hallucinations under visual corruptions.

It improves descriptive faithfulness in degraded visual conditions.

The method is effective across 15 common corruption types.

Abstract

Large vision-language models (LVLMs) tend to hallucinate, especially when visual inputs are corrupted at test time. We show that such corruptions act as additional distribution shifts, significantly amplifying hallucination rates in real-world applications. To address this, we propose CLIP-guided Test-Time Training (ClipTTT), a method to adapt LVLMs under degraded conditions on the fly with a single test sample. Specifically, we leverage the image-text alignment strength of a pre-trained CLIP model as a stable guidance signal to identify reliable self-supervision targets, enabling rapid adaptation without altering the base LVLMs. Extensive experiments on standard hallucination benchmarks, with 15 common corruptions, demonstrate that ClipTTT effectively mitigates hallucinations and improves descriptive faithfulness under visual corruptions.