Reproduce “PAS : Prelim Attention Score for Detecting Object Hallucinations in Large Vision--Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nhat Hoang-Xuan, Minh Vu, My T. Thai, Manish Bhattarai
Researchers from Los Alamos National Laboratory and the University of Florida introduced the Prelim Attention Score (PAS) to detect object hallucinations in Large Vision-Language Models (LVLMs), identifying that high attention to previously generated 'prelim' tokens indicates hallucination. PAS achieves an average AUROC of 85.0% across various LVLMs and datasets, outperforming existing reference-free, training-free baselines.
Key insights
- An LVLM's over-reliance on its own previously generated output tokens, rather than the visual input, is a strong and previously overlooked signal for object hallucination.
- This over-reliance suggests an 'alternative unreliable operating mode' where the model prioritizes self-generated textual context over visual grounding.
- The first attention layer (layer 0) of the Transformer architecture is particularly effective in capturing these preliminary token dependencies for hallucination detection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.