Beyond Visual Enhancement: Adaptive Multi-Context Steering to Mitigate LVLM Hallucinations.

arXiv:2610.11907v1 Announce Type: cross Abstract: Hallucination remains a significant challenge in Large Vision-Language Models (LVLMs). Existing training-free methods generally mitigate hallucinations through contrastive decoding or visual enhancement, often