Reproduce “Mitigating Hallucination in VideoLLMs via Temporal-Aware Activation Engineering” 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
Jianfeng Cai, Wengang Zhou, Zongmeng Zhang, Jiale Hong, Nianji Zhan, Houqiang Li
USTC researchers develop a temporal-aware activation engineering framework that reduces hallucination in VideoLLMs by identifying and manipulating hallucination-sensitive model components based on video temporal characteristics, achieving up to 24.21% improvement on EventHallusion benchmarks without requiring model retraining.
Key insights
- Hallucination sensitivity correlates strongly with temporal variation rather than task type
- Frame reduction reliably induces hallucination for testing purposes
- Module selection based on temporal characteristics is crucial for effectiveness
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.