Reproduce “Quantity Matters: Towards Assessing and Mitigating Number Hallucination in Large Vision-Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Huixuan Zhang, Junzhe Zhang, Xiaojun Wan
Researchers at Peking University introduced and quantitatively assessed "number hallucination" in Large Vision-Language Models, identifying pervasive inaccuracies in object counting. They developed a consistency training approach that improved counting accuracy by an average of 8% Macro-F1 across various models, addressing a gap in comprehensive quantitative understanding.
Key insights
- All tested open-source LVLMs exhibited substantial "number hallucination" with an average MAE of approximately 2, and model scaling alone did not significantly mitigate this issue.
- LVLMs demonstrated high levels of "inner" and "outer" inconsistency across related numerical tasks, indicating a fundamental lack of robust numerical understanding rather than just simple errors.
- Improving a model's consistency through training on a diverse set of numerical questions (binary, comparative, and direct counting) leads to a reduction in number hallucination, outperforming direct finetuning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.