OVFact: Measuring and Improving Open-Vocabulary Factuality for Long Caption Models

2025-07-25

AI Summary

Large vision-language models (VLMs) frequently generate factually inaccurate long descriptions, yet existing hallucination evaluation methods rely heavily on human annotation and struggle to scale to lengthy, heterogeneous outputs. To address this, we propose OV-Fact: a reference-free framework for assessing factual consistency in long visual descriptions, leveraging open-vocabulary visual grounding and tool-augmented verification. Its core contributions are threefold: (1) the first integration of open-vocabulary localization with structured tool-based validation to jointly quantify richness and factual accuracy; (2) a reference-free metric and an LLM-driven factuality filtering strategy enabling efficient data curation and model training; and (3) empirical validation showing that training on only 20–40% of pretraining data selected by OV-Fact—reducing noise by 2.5–5×—significantly improves factual precision while preserving descriptive diversity across multiple long-description benchmarks.

Technology Category

• Computer Vision: Large Vision Models
• Natural Language Processing: Fact-Checking / Misinformation Detection (NLP Focus)
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Large vision-language models (VLMs) often struggle to generate long and factual captions. However, traditional measures for hallucination and factuality are not well suited for evaluating longer, more diverse captions and in settings where ground-truth human-annotated captions are unavailable. We introduce OV-Fact, a novel method for measuring caption factuality of long captions that leverages open-vocabulary visual grounding and tool-based verification without depending on human annotations. Our method improves agreement with human judgments and captures both caption descriptiveness (recall) and factual precision in the same metric. Furthermore, unlike previous metrics, our reference-free method design enables new applications towards factuality-based data filtering. We observe models trained on an OVFact-filtered (2.5-5x less) subset of a large-scale, noisy (VLM-generated) pretraining set meaningfully improve factuality precision without sacrificing caption descriptiveness across a range of downstream long caption benchmarks.

Problem

• Measure factuality in long captions without human annotations
• Improve caption descriptiveness and factual precision together
• Filter noisy training data to enhance model factuality

Innovation

• Open-vocabulary visual grounding for factuality measurement
• Tool-based verification without human annotations
• Reference-free metric enables factuality-based data filtering