Reproduce “Tracing and Mitigating Hallucinations in Multimodal LLMs via Dynamic Attention Localization” 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

Tiancheng Yang, Lin Zhang, Jiaye Lin, Guimin Hu, Di Wang, Lijie Hu

Researchers at MBZUAI and KAUST developed D-LEAF, a training-free method to dynamically localize and mitigate MLLM hallucinations by selectively correcting attention heads based on Layer Image Attention Entropy and Image Attention Focus. The method reduced sentence-level hallucination by 53% on MiniGPT-4 and improved VQA F1-scores by 4% on POPE, adding only 8% inference latency.

Key insights

  • Empirical analysis shows that hallucination-prone attention heads are not uniformly distributed across MLLM layers but tend to cluster in specific layers, indicating the need for localized intervention.
  • Indiscriminate suppression of attention heads, while reducing hallucinations, can lead to shorter outputs and decreased overall F1 scores, underscoring the importance of targeted and dynamic corrections.
  • The D-LEAF framework, through its dynamic attention modulation, provides a training-free approximation of the Direct Preference Optimization (DPO) objective, offering a theoretical link between attention control and preference alignment.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.