🧪 Test?View on arXiv
Targeting the Attention Heads Behind Object Hallucination in LLaVA
Author1, Author2, Author3, Author4, Author5
multimodalinterpretabilityobject detectionattention mechanisms
2608.24966
Builder Relevance
3h ago80%
Abstract
This paper investigates the interpretability of object hallucination in vision-language models and proposes a targeted intervention to reduce hallucinated object mentions.
Reality Card
Core Claim
The proposed method reduces the fraction of captions with hallucinated objects from 0.370 to 0.230 and hallucinated object mentions from 0.156 to 0.096.
Method / Result
The combined method lowers hallucination rates significantly on 400 held-out COCO images.
Limitations
The method may lower object recall from 0.78 to 0.70, indicating a trade-off between hallucination reduction and recall.
Paper to code
Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.
No verified implementation link has been attached yet. AIBuzzHub will keep this panel separate from unverified search results.