Current AI video generators are stuck in a loop of self-deception. They optimize for isolated metrics like audio clarity or visual sharpness, but often produce content that feels disjointed or semantically nonsensical to human viewers. This disconnect happens because traditional reward signals fail to capture the holistic coherence that makes a video feel right, leading to models that game the system rather than create art.

Researchers have introduced VA-Judger, a new reward model designed to align AI generation with actual human preference. Instead of judging audio and video separately, this system evaluates the joint experience, ensuring the text prompt, visuals, and sound work together seamlessly. To build this, they created VAPref-10K, a large dataset of human-preference comparisons comprising 9K prompts and 10.3K fine-grained paired comparisons, and developed a chain-of-thought model that learns to distinguish quality gaps and explain its reasoning.

The conflict lies in how we measure success. Existing methods encourage reward hacking, where models generate technically perfect but emotionally hollow content. VA-Judger breaks this cycle by decomposing feedback into specific dimensions while maintaining an overall sense of coherence. This approach provides denser, more reliable reward signals that guide the model toward genuinely pleasing outputs rather than just high-scoring ones.

The results show that VA-Judger significantly outperforms traditional metric baselines in predicting human preferences, both on familiar and new data. When used to post-train generation models, it leads to substantial improvements in quality. This marks a shift from optimizing for numbers to optimizing for human perception, proving that true AI creativity requires understanding the whole, not just the parts.