Reproduce “Attention Hijackers: Detect and Disentangle Attention Hijacking in LVLMs for Hallucination Mitigation” 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
Beitao Chen, Xinyu Lyu, Lianli Gao, Jingkuan Song, Heng Tao Shen
Researchers from the University of Electronic Science and Technology of China and Southwestern University of Finance and Economics identified "Attention Hijacking" by instruction tokens as a cause of hallucinations in Large Vision-Language Models (LVLMs) and proposed AID, a training-free strategy to detect and disentangle this phenomenon. AID improved F1-scores on the POPE benchmark, reduced hallucination rates (e.g., CHAIR_s by 8.2% and CHAIR_i by 2.4% on LLaVA-1.5), and maintained general performance across various LVLM architectures.
Key insights
- "Attention Hijacking" by instruction tokens is a distinct and quantifiable source of hallucinations, accounting for up to 46.7% of total hallucinations across different LVLMs.
- The autoregressive nature of LVLMs, reinforced by the KV-cache, causes persistent misdirection of attention from hijackers to irrelevant visual regions.
- Selectively disentangling the influence of these hijacking instruction tokens can effectively mitigate hallucinations without requiring additional training or compromising overall model performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.