Video language models reduce hallucinations using attention based perturbations
Mitigating Video Hallucinations via Attention-Guided Representation Perturbation
Computer Vision and Pattern RecognitionArtificial IntelligenceComputation and Language
Summary
Video large language models sometimes give answers not supported by the video they analyze, which is called hallucination. The authors introduce a way to guide the model's focus away from visual parts it relies on too strongly by slightly changing its internal video representations. This creates a negative example that helps the model contrast good answers with less truthful ones during generation. Their method, tested on several models and benchmarks, reduces hallucinations and improves overall video understanding without needing extra training.
What this means in practice
- •For video app developers: Enhance video question-answering apps to produce more accurate responses by reducing unsupported hallucinations using attention-guided perturbations.
- •For multimedia content platforms: Improve content indexing and retrieval quality by integrating more reliable video understanding models that minimize hallucinated video descriptions.
Authors
Zhixin Cai, Ruoqi Yuan, Xinchen Yang, Zhuofan Chen, Jun Bai, Chuantao Yin, Wenge Rong
Abstract
Video large language models (Video LLMs) often generate responses unsupported by the input video. Contrastive decoding (CD) can mitigate these hallucinations, but its effectiveness depends on how the negative branch weakens visual evidence. Existing methods rely on predefined perturbations or masking guided by question reconstruction, offering limited control over the model's visual reliance during answer generation. We introduce Attention-guided Representation Perturbation for Contrastive Decoding (ARP-CD), a training-free framework for constructing contrastive negatives directly in representation space. ARP-CD uses the attention mass assigned to visual tokens at the answer-prediction position to guide bounded gradient updates of the language model's input visual hidden states. These updates reduce visual attention and produce a negative input, whose predictions are contrasted with those of the original branch throughout generation. Experiments across four Video LLMs and four benchmarks show that ARP-CD outperforms representative training-free baselines in most model--benchmark settings, reducing hallucinations while improving general video understanding. Further analyses show that attention-guided perturbation directions improve negative quality and that the resulting negatives provide effective contrastive references for correcting the original predictions.