Unmasking LVLM Hallucinations

Large Vision-Language Models (LVLMs) exhibit impressive capabilities but remain susceptible to generating outputs not grounded in visual input. A critical question has been the relative contribution of vision backbone limitations versus language dominance to this hallucination problem. New research from Khayatan et al. introduces the HalluScope benchmark, a novel tool designed to dissect the factors inducing these visual grounding failures.

Excessive Textual Priors Fuel Hallucinations

The analysis conducted using the HalluScope benchmark reveals a significant finding: LVLM hallucinations largely stem from an over-reliance on textual priors and background knowledge. This is particularly evident when information is introduced through textual instructions, suggesting that the language component's learned associations can override or misinterpret visual cues. This insight challenges previous assumptions and provides a clearer target for mitigation strategies.

HalluVL-DPO: Grounding Language with Preference Optimization

To address the identified issue of instruction-induced hallucinations, the researchers propose HalluVL-DPO. This framework employs preference optimization, fine-tuning existing LVLMs using a carefully curated dataset. The training process guides the model to favor visually grounded responses over those prone to hallucination. Initial results demonstrate that HalluVL-DPO effectively reduces targeted hallucination failure modes while maintaining or even improving performance on other benchmarks and core visual capabilities.