Mitigating Fine-Grained Hallucination by Fine-Tuning Large Vision-Language Models with Caption Rewrites

1 Introduction

• Existing evaluation methods such as CHAIR and POPE focus on coarse object hallucination, not inaccurate attributes and behaviors.The paper calls captions containing inaccurate attributes or behaviors fine-grained object hallucination.Source: §1 · p. 2
• MiniGPT-4’s example caption invents an airplane color and incorrectly says the flying airplane is on the runway.The example distinguishes incorrect object attributes from incorrect object behavior.Source: §1 · p. 2
• ReCaption rewrites captions with ChatGPT and fine-tunes instruction-tuned LVLMs on the rewritten captions.Its two-stage prompting extracts verbs, nouns, and adjectives, then uses them to generate diverse captions that retain the original caption’s core content.Source: §1 · p. 2
• FGHE evaluates object hallucination and probes hallucinated object attributes and behaviors through binary Yes-or-No questions.It incorporates POPE-style questions about object presence alongside questions about object attributes and behaviors.Source: §1 · p. 2
• ReCaption effectively reduces fine-grained object hallucination across LVLMs evaluated using POHE and FGHE.The introduction reports this outcome for any LVLM adopting the framework.Source: §1 · p. 3

2.1 Large Vision-Language Models

• LVLM research includes textual visual descriptions connected to closed-source LLMs and models trained on vision-language instructions.Examples include Visual ChatGPT, MM-REACT, and HuggingGPT for the first paradigm, and MultiInstruct, LLaVA, MiniGPT-4, mPLUG-Owl, and InstructBLIP for the second.Source: §2.1 · p. 3

2.2 Hallucnation in Large Vison-Language Mdoels

• Prior work classifies object hallucinations as intrinsic errors involving incorrect or nonexistent objects and external errors involving irrelevant objects.The discussion also notes POPE’s empirical examination of object hallucinations across LVLMs.Source: §2.2 · p. 3
• Prior LVLM hallucination work mainly examines object presence, while this paper targets finer-grained errors such as incorrect object attributes.The paper’s stated focus is to mitigate and evaluate finer-grained hallucinations.Source: §2.2 · p. 3

3 ReCaption Framework

• ReCaption combines ChatGPT caption rewriting with additional training of instruction-tuned LVLMs on rewritten captions.The framework is described as LVLM-agnostic.Source: §3 · p. 3

3.1 Caption Rewriting

• A two-stage prompt extracts verbs, nouns, and adjectives from each original caption, then generates rewritten captions containing those keywords.Repeated generation produces captions with varied details while preserving essential information from the original caption.Source: Figure 2 · p. 4, §3.1 · p. 4

3.2 Additional Tuning

• ReCaption generates R rewritten captions per image-caption pair and uses them for additional LVLM tuning; experiments set R = 5.The rewritten pairs are model-agnostic and can be adapted to multiple LVLM architectures with minimal changes.Source: §3.2 · p. 5
• The authors report that a small number of rewritten image-caption pairs can reduce fine-grained hallucination and improve caption quality.They describe the additional tuning cost as negligible compared with training the original LVLM.Source: §3.2 · p. 5

• The authors omit CHAIR because it suffers from prompt sensitivity and inaccuracies, as reported by Li et al. (2023).Source: §4 · p. 5

POPE dataset and evaluation method

• POPE probes object presence with yes-or-no questions and evaluates answers as binary classifications using Accuracy, Precision, Recall, and F1.Its dataset contains 3,000 questions for captions of 500 images.Source: §4 · p. 5

FGHE dataset and evaluation method

• FGHE uses 200 binary questions on 50 MSCOCO validation images to probe object relationships, attributes, and behaviors.The questions were manually defined, with 100 yes-questions and 100 no-questions.Source: §4 · p. 5

5.1 Experimental Setup

• Experiments compare four open-source instruction-tuned LVLMs; training uses 500 image-caption pairs and five rewritten versions per caption.The models are mPLUG-Owl, LLaVA, Multimodal-GPT, and MiniGPT-4.Source: §5.1 · p. 6

5.2 Main Results

• Recall findings indicate overconfident “Yes” responses in several models; MultiModal-GPT performs best on both hallucination types, while mPLUG-Owl scores 69.89% on POPE and 68.42% on FGHE.Source: §5.2 · p. 6
• MiniGPT-4 with ReCaption improves F1 by 7% on POPE and 10% on FGHE; mPLUG-Owl gains 1.32% on POPE and 3.71% on FGHE.Source: §5.2 · p. 6, Table 2 · p. 7

5.3 Further Analysis

• ReCaption improves F1 across FGHE’s multiple-object, attribute, and behavior categories, with larger reductions for attributes and behaviors than for multiple objects.The authors suggest rewritten captions’ verb and adjective keywords may help reduce attribute and behavior hallucinations.Source: §5.3 · p. 7, Table 3 · p. 7
• Simple prompt-based rewriting yields minor gains, whereas the two-stage ReCaption strategy produces substantial improvements with diverse captions retaining the same keywords.Source: Figure 4 · p. 7, §5.3 · pp. 7–8
• MiniGPT-4’s F1 score rises across training epochs on both POPE and FGHE when trained with ReCaption.The authors interpret this trend as improved image-text fine-grained alignment from further training on the limited rewritten data.Source: Figure 5 · p. 8, §5.3 · p. 8
• Across MiniGPT-4, mPLUG-Owl, and LLaVA, FGHE performance improves as both the number of rewritten images and rewrites per image increase.Source: §5.3 · p. 8
• Case examples show ReCaption improves caption accuracy even for vanilla mPLUG-Owl, whose captions were relatively stronger than those of the other three models.Source: §5.3 · p. 8

• The paper addresses fine-grained object hallucinations in instruction-tuned LVLMs.Source: §6 · p. 8
• ReCaption combines ChatGPT caption rewriting with LVLM fine-tuning on rewritten captions, and FGHE evaluates fine-grained hallucinations.The authors report that ReCaption mitigates these hallucinations across LVLMs while enhancing text generation quality.Source: §6 · p. 8
• Future work should refine rewriting for unnatural images, including invoices and cartoons, and for longer output text.Source: §6 · p. 9

7 Limitations

• Keyword-only rewriting disregards structural information and relationships among keywords, hindering preservation of crucial details from the original caption.Source: §7 · p. 9
• Training any LVLM requires a small corpus of alignment-enriched rewrites, even though training costs are minimal.Source: §7 · p. 9

8 Ethics Statement

• FOHE comes from the MSCOCO validation set, licensed under Creative Commons Attribution 4.0, which permits building upon the evaluation dataset.Source: §8 · p. 9
• The research group annotated the small evaluation dataset under agreed salaries and workloads.Source: §8 · p. 9
• The authors state that FOHE texts contain no private information or social issues.Source: §8 · p. 9

A Appendix: More Examples

• Figure 7 and Figure 11 compare MiniGPT-4 outputs with outputs from MiniGPT-4 with ReCaption.Source: Figure 7 · p. 12, Figure 11 · p. 13
• Figure 8 and Figure 12 compare mPLUG-Owl outputs with outputs from mPLUG-Owl with ReCaption.Source: Figure 8 · p. 12, Figure 12 · p. 13
• Figure 9 and Figure 13 compare MultiModal-GPT outputs with outputs from MultiModal-GPT with ReCaption.Source: Figure 9 · p. 12, Figure 13 · p. 13
• Figure 10 and Figure 14 compare LLaVa outputs with outputs from LLaVa with ReCaption.Source: Figure 10 · p. 12, Figure 14 · p. 13