Beyond Visual Enhancement: Adaptive Multi-Context Steering to Mitigate LVLM Hallucinations
AIMS is a lightweight, training-free method for Large Vision-Language Models (LVLMs) to reduce hallucinations during text generation. It adaptively coordinates three context sources—visual, prefilled text, and generated tokens—by creating prototypes and steering the query representation at each decoding step. This allows the model to dynamically balance information from different sources to suppress errors without needing additional training.