Video large language models (Video LLMs) often produce outputs that are not grounded in the actual video content, a problem known as hallucination. Contrastive decoding (CD) has emerged as a way to reduce such errors, but its effectiveness depends heavily on how the negative branch—the signal used to contrast against the model's default behavior—is constructed. A weak or poorly targeted negative branch limits the benefit.
The paper introduces an attention-guided representation perturbation method to address this limitation. Rather than relying on generic perturbations, the approach uses the model's own attention maps to identify which representations to perturb, creating a more effective negative branch for contrastive decoding. This targeted strategy is designed to suppress hallucinated content while preserving accurate video understanding.
Because this is a single arXiv preprint, the results have not yet been peer-reviewed or independently replicated. The abstract does not provide quantitative outcomes, so the claimed improvement should be treated as preliminary until full experimental details and evaluations are available.