Enhancing Refusal Capability in Large Vision-Language Models
Researchers have developed a framework to improve the ability of Large Vision-Language Models (VLMs) to refuse answering questions that exceed their parametric knowledge, reducing factual hallucinations. The method involves creating a model-specific 'Visual-Idk' dataset using multi-sample consistency probing to distinguish known from unknown facts. Models are then aligned via supervised fine-tuning and preference-aware optimization (DPO, ORPO). On the Visual-Idk dataset, the Truthful Rate improved from 57.9% to 67.3%. Internal probing indicates the model genuinely recognizes its knowledge boundaries.
Key facts
- VLMs are prone to factual hallucinations, especially in long-tail or specialized domains.
- Current models have weak refusal capability for queries exceeding their knowledge.
- A systematic framework enhances refusal capability for unknown questions.
- A model-specific 'Visual-Idk' dataset is curated using multi-sample consistency probing.
- Supervised fine-tuning and preference-aware optimization (DPO, ORPO) are used for alignment.
- Truthful Rate improved from 57.9% to 67.3% on the Visual-Idk dataset.
- Internal probing shows the model genuinely recognizes its boundaries.
- The research is published on arXiv with ID 2604.26419.
Entities
—