Reproduce “Visually Dehallucinative Instruction Generation: Know What You Don't Know” 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
Sungguk Cha, Jusung Lee, Younghyun Lee, Cheoljong Yang
Researchers at NCSOFT Corporation developed IDK-Instructions and the VQAv2-IDK benchmark to train large multi-modal models to express uncertainty when faced with unanswerable visual questions. This approach improved IK hallucination accuracy from baselines like 3.73% to 83.67% for LLaVA1.5 while preserving other visual task performances.
Key insights
- Existing open-source LMMs (e.g., BLIP-2, LLaVA1.5) exhibit very limited capability to express uncertainty, with IDK keyword inclusion rates below 4% on hallucination-inducing questions.
- Carefully designed few-shot examples are critical for generating high-quality, diverse, and expressive 'I don't know' instructions, moving beyond generic uncertainty statements.
- Training LMMs with specific dehallucinative instructions effectively mitigates IK hallucination without compromising or often enhancing general visual recognition and VQA capabilities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.