Semi-supervised semantic segmentation (SSSS) balances annotation efficiency and segmentation performance by leveraging both labeled and unlabeled data. Although recent work focuses on improving pseudo-label reliability, they often suffer from confirmation bias caused by relying on the predictions of a single model, leading to erroneous probability maps and suboptimal pseudo-labels. To address this, we propose MPIF, a novel framework integrating probabilistic fusion, multi-model collaboration, and adaptive balancing to mitigate prediction bias and enhance pseudo-label quality. Specifically, we first introduce a probabilistic information fusion strategy that aggregates prediction maps from multiple training stages to reduce single-model bias. Then, a multi-model collaborative prediction mechanism selects high-performing historical models from a queue to generate diverse and accurate predictions. Finally, an adaptive balancing strategy adjusts the learning weights of labeled and unlabeled data based on pseudo-label uncertainty. MPIF effectively suppresses noisy supervision, stabilizes training, and achieves state-of-the-art performance on PASCAL VOC 2012, COCO, and ADE20K datasets.