Through the Magnifying Glass: Adaptive Perception Magnification for Hallucination-Free VLM Decoding