Researchers have developed a new training-free framework called Attention-guided Representation Perturbation for Contrastive Decoding (ARP-CD) to address hallucinations in Video Large Language Models (Video LLMs). This method directly constructs contrastive negatives in the representation space by using the attention mass assigned to visual tokens to guide updates to the language model's input visual hidden states. Experiments on four Video LLMs and four benchmarks demonstrate that ARP-CD effectively reduces hallucinations and enhances general video understanding, outperforming existing training-free baselines in most settings. AI
IMPACT This research offers a method to improve the reliability and accuracy of Video LLMs, potentially leading to better video understanding applications.
RANK_REASON The cluster describes a new research paper detailing a novel method for improving Video LLMs. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- Attention-guided Representation Perturbation for Contrastive Decoding
- Hugging Face
- language model
- Video LLMs
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
