VidLBEval: Benchmarking and Mitigating Language Bias in Video-Involved LVLMs