Preference optimization reduces hallucinations but drops object coverage

Say Less, Hallucinate Less? Diagnosing and Restoring Object Coverage in Preference-Optimized Vision-Language Models

Computer Vision and Pattern RecognitionArtificial Intelligence

Summary

Vision-language models sometimes make mistakes by naming objects that aren't really in a picture, known as hallucination. The authors found that a popular way to reduce these hallucinations, called preference optimization, also makes the model mention fewer objects overall. They discovered that stopping the model's captioning earlier can regain many of the missed objects without training. Finally, they propose a simple fix to add back high-confidence objects missed by preference optimization, improving object coverage while keeping hallucinations low.

What this means in practice

  • For computer vision engineers: Improve image captioning systems by balancing hallucination reduction with better object coverage using early stopping and object-category restoration methods.
  • For machine learning platform teams: Add a post-processing step that supplements model outputs with high-confidence object mentions to enhance coverage without retraining vision-language models.

Authors

Hansang Lee

Abstract

Preference optimization (PO) is the standard way to reduce object hallucination in large vision-language models. We find that four public PO checkpoints of LLaVA-1.5-7B also lower object coverage. They name fewer object categories, and their lower hallucination rates bring no significant gain in set-level object F1 on CHAIR. We therefore ask how much of the PO gain the base model can reproduce by stopping earlier. A positive end-of-sequence (EOS) logit bias lets the base model reach the recall of each PO model, and this early stopping alone reproduces 58 to 86 percent of the CHAIRi gain of three checkpoints on two disjoint COCO sets. Writing longer does not bring the lost categories back. Under greedy decoding, a negative EOS bias recovers only 2 to 9 percent of the lost CHAIR recall of these checkpoints, and 94 to 96 percent of the added object mentions repeat categories already named. On CHAIR, the base caption of the same image still names about half of the missed object categories. We therefore propose base-guided object-category restoration, which appends high-confidence base-only categories to the PO caption without training. With thresholds fixed on a development set, it brings three checkpoints within one point of the base recall on a held-out COCO set with a lower distinct-object hallucination rate than the base model. Against the recall-matched base stopping curve, its gain holds for RLAIF-V near the base recall and for RLAIF-V and CHAIR-DPO at an intermediate recall. We will release the code of the recall-controlled evaluation.