IC-106Logit lens on LLaVA and InstructBLIP image representations shows higher internal confidence for objects present in the image than for hallucinated objects
Nicholas Jiang, Anish Kachinthaya, Suzanne Petryk, Yossi Gandelsman
The paper applies the logit lens to intermediate image patch embeddings in LLaVA 1.5 and InstructBLIP, projecting them through the language model's unembedding matrix to obtain per-patch, per-layer probability distributions over the vocabulary. For 5000 COCO 2014 images, the maximum probability of each of the 80 COCO objects across all patches and layers (the internal confidence) is consistently higher for objects actually present than for hallucinated ones. Using this internal confidence as a classifier for object presence achieves mAP 0.78 (InstructBLIP) and 0.60 (LLaVA), outperforming the baseline of maximum output token probability (0.53 and 0.49). The same pattern holds for LLaVA-Next and Cambrian-1 in the appendix.
Evidence
correlational
Key metric
mAP 0.78 (InstructBLIP) vs 0.53 (baseline), 0.60 (LLaVA) vs 0.49 (baseline); ROC AUC 0.83 vs 0.55, 0.68 vs 0.47; 47.17% mAP improvement (InstructBLIP), 22.45% (LLaVA); LLaVA-Next mAP 0.95 vs 0.93, Cambrian-1 mAP 0.97 vs 0.94
Caveat
Multi-token objects are handled by taking the max probability over constituent tokens, which the authors note can introduce noise if tokens are shared across different words.