To address the limitations of separate metrics for audio quality, visual fidelity, and synchronization in joint video-audio generation, researchers introduced VA-Judger, a chain-of-thought omni-reward model designed to align with human preferences. Traditional metric-based optimization often leads to reward hacking, where models produce high-scoring but semantically incoherent content. To combat this, the authors constructed VAPref-10K, a large-scale dataset containing 9K prompts and 10.3K fine-grained paired comparisons from open-source models. The VA-Judger training pipeline employs a three-stage approach: first, it learns structured output and coarse discrimination from pairs with significant quality gaps; second, it utilizes rejection sampling verified against human annotations to distill reliable preference explanations for subtle, near-quality comparisons; and third, it implements dimension-wise reinforcement learning to decompose human feedback into specific quality dimensions, providing denser reward signals than binary labels. Evaluation via the newly introduced VA-Judger-Bench benchmark demonstrates that the model outperforms existing metric baselines in predicting human preferences across both in-domain and out-of-domain scenarios. Furthermore, employing VA-Judger's human-aligned rewards for post-training significantly enhances the generation quality of joint video-audio models, capturing the complex semantic and temporal coherence required for high-fidelity multimodal content.