Reproduce “Beyond Language Priors: Diagnosing and Fixing Visual-Origin Hallucinations in Multimodal LLM” 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

Peiyang Xu, Xiaopei Zhu, Jun Zhu, Xiaolin Hu

Researchers at Tsinghua University identified "visual-origin hallucination" as a key source of MLLM errors, stemming from faulty visual feature extraction and image-text misalignment. Their Adversarial Contrastive Fine-Tuning (ACFT) method, using adversarially generated image pairs, achieved state-of-the-art hallucination mitigation on benchmarks like POPE, improving LLaVA's accuracy by up to 3.3% without degrading general visual comprehension.

Key insights

  • MLLM hallucinations can originate from the visual modality due to incorrect feature extraction and misalignment between visual and textual representations, not solely from language priors.
  • Hallucinating models exhibit lower image-text embedding similarity and inverted visual attention, misdirecting focus on absent objects and dispersing it for present ones.
  • Minimal adversarial perturbations can flip an MLLM's object hallucination status, indicating the inherent fragility and proximity to decision boundaries of current visual representations.

Summary by alphaXiv.

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