Study finds preference optimization trims vision-language hallucination but also cuts object coverage
The arXiv paper reports that four tuned LLaVA-1.5-7B checkpoints name fewer object categories, and shows a training-free fix that restores most of the lost recall.
In plain words
Vision-language models that describe images sometimes mention objects that are not there. A common fix, called preference optimization, reduces these errors but also makes the models mention fewer real objects. The authors show a simple, training-free method that adds the missing object names back.
What was announced
The authors examined four publicly available preference-optimized checkpoints of LLaVA-1.5-7B and found that each one lowered object coverage alongside its hallucination reduction. The tuned models named fewer object categories, and the lower hallucination rates did not translate into significant gains in set-level object F1 on the CHAIR benchmark. Testing whether the base model could match this by simply stopping earlier, they found a positive end-of-sequence logit bias let it reach each checkpoint's recall, with early stopping alone reproducing 58 to 86 percent of the CHAIRi gain of three checkpoints on two disjoint COCO sets.
Writing longer captions did not restore the lost categories: under greedy decoding, a negative EOS bias recovered only 2 to 9 percent of the lost CHAIR recall, and 94 to 96 percent of added object mentions repeated categories already named. The paper's proposed remedy, base-guided object-category restoration, appends high-confidence categories from the base model to the preference-optimized caption without any training. With thresholds fixed on a development set, it brought three checkpoints within one point of base recall on a held-out COCO set while keeping a lower distinct-object hallucination rate than the base model. The authors state they will release the recall-controlled evaluation code.
What it does not show
- The evidence is the paper's abstract only; full methods, figures and peer review status are not available for verification.
- Results cover LLaVA-1.5-7B checkpoints and COCO/CHAIR evaluation; no other model families or benchmarks are reported.
- The restoration gain against the recall-matched base stopping curve held only for RLAIF-V near base recall and for RLAIF-V and CHAIR-DPO at an intermediate recall, per the abstract.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.