Mitigating Hallucinations in Large Vision-Language Models by Adaptively Constraining Information Flow