IC-107Linear orthogonalization of LLaVA and InstructBLIP image features against text embeddings removes hallucinated objects at 83-86% individual rate versus 7-16% for correctly detected objects
Nicholas Jiang, Anish Kachinthaya, Suzanne Petryk, Yossi Gandelsman
The paper introduces ProjectAway, which subtracts a scaled text embedding from each image patch embedding to orthogonalize them with respect to a target object. Applied to 5000 COCO 2014 images, this linear edit removes hallucinated objects at individual rates of 83.3% (InstructBLIP) and 86.0% (LLaVA), while correctly detected objects are removed at only 16.2% and 6.9% respectively. In the full hallucination reduction pipeline (thresholding internal confidence to detect hallucinations, then mass-removing), the raw count of hallucinated objects drops by 25.7% for InstructBLIP and 23.8% for LLaVA compared to beam search, while recall is preserved. The differential removal rate suggests hallucinated objects are more weakly encoded in the image representations.
Evidence
interventional
Key metric
Individual RR: 83.3% (InstructBLIP hallucinations), 86.0% (LLaVA hallucinations), 16.2% (InstructBLIP CD), 6.9% (LLaVA CD); Mass RR: 74.3%, 72.8%, 15.0%, 8.3%; Raw hallucinated object count reduction: 25.7% (InstructBLIP, 564 to 419), 23.8% (LLaVA, 583 to 444); CHAIR_s: 14.6 to 12.5 (InstructBLIP), 15.0 to 12.2 (LLaVA)
Caveat
The method simplifies multi-token objects by using the last token's embedding or averaging constituent tokens, which the authors note can introduce approximation. At high weight factors, captions become nonsensical with complete loss of both CD and hallucinated objects.