Reproduce “Verb Mirage: Unveiling and Assessing Verb Concept Hallucinations in Multimodal Large Language Models” on OpenResearch

About this paper

Zehao Wang, Xinpeng Liu, Yudonglin Zhang, Xiaoqian Wu, Zhou Fang, Yifan Fang, Junfu Pu, Cewu Lu and 1 more

Researchers at Shanghai Jiao Tong University conducted the first systematic investigation into verb hallucination in Multimodal Large Language Models (MLLMs), developing a dedicated benchmark and a baseline mitigation method. Their approach successfully improved MLLM verb understanding without compromising general performance on other tasks.

Key insights

• MLLMs exhibit severe verb hallucination, often showing low F1 scores for verb-related questions despite performing well on object-centric benchmarks.
• Verb understanding in MLLMs heavily depends on explicit object references, leading to significant performance degradation when object cues are absent, and models tend to over-generate "Yes" answers.
• Existing object-centric hallucination mitigation methods are largely ineffective or detrimental for verb hallucination, indicating a need for verb-specific approaches.
• MLLMs may possess reasonable underlying verb knowledge, but suffer from mis-calibration of output token probabilities, suggesting that low accuracy might not always mean a complete lack of understanding.

Summary by alphaXiv.