AI Papers Reader

Personalized digests of latest AI research

View on GitHub

Chunked Video Judging Offers AI “World Models” a More Reliable Reward

Camera-conditioned “world models” generate interactive video from a starting image and user camera controls, but judging whether they actually obey is deceptively hard: a model can move in the requested direction while its images drift, blur, or deform, or can look visually polished while ignoring the command. Existing evaluation systems tend to split those problems—geometry trackers check the camera path, image scorers check frame quality—so they miss failures that require seeing action and visual outcome together. In a preprint, researchers from Fudan University, Tencent Hunyuan, and associated labs describe WorldReward, a vision-language reward model designed to judge both command execution and visual quality from the same local evidence.

WorldReward works in paired comparisons: two videos are generated from the same source image, caption, and camera trajectory, and the model predicts which one human prefer. Rather than feeding the entire long video and action list to a vision-language model, it divides the trajectory into chunks of four consecutive actions. Each chunk becomes a structured visual input containing the source image, a frame-grid overview showing the start, middle, and end of each action in both videos, and action-level panels that isolate the first and last frame of each action. The model reasons over that evidence, then casts separate chunk-level votes for action consistency and visual quality. Those votes are aggregated into a video-level winner or tie.

The design is meant to catch short-lived failures. In one annotated example, two candidate videos show a Toyota SUV parked in desert terrain under commands that include yaw-left rotation and forward movement. Both candidates execute the commands, but WorldReward judges Video A the action and visual winner: its rotation retains natural 3D depth and parallax, while Video B’s motion looks flat, “like a 2D pan and zoom,” and the scene undergoes severe source drift—inexplicably changing the time of day to sunset and altering the background landscape. The action-level panels make that local transition visible instead of letting it be diluted by the rest of the clip.

On WorldReward-Bench, a new human-annotated benchmark of 760 paired generations, WorldReward achieved the highest agreement with human preferences on all three measured dimensions. Its action-consistency agreement was 77.63%, surpassing GPT-5.5 by 3.42 percentage points; appearance agreement was 81.32%, a 1.45-point gain; and motion agreement was 73.03%, a 3.56-point gain. Notably, the model was trained in part from judgments distilled from GPT-5.5 and Gemini 3.1 Pro, yet it outperformed both after a multi-stage cleanup involving tool-based agent audits and targeted human review.

The authors also used WorldReward as the reward signal in reinforcement-learning post-training of the open world model HY-WorldPlay 1.5. Compared with the prior WorldCompass pipeline, which uses separate geometric and image rewards, WorldReward improved combined-action accuracy by 1.58 to 2.78 percentage points and basic-action accuracy by 2.28 to 5.81 points across short-, mid-, and long-horizon generation, as estimated by the DepthAnything3 trajectory tracker. Visual quality, measured by the HPSv3 preference score, rose in all six settings, with gains of 0.18 to 0.29 for combined actions and 0.15 to 0.24 for basic actions.

The results do not imply perfect judgment. WorldReward’s strongest benchmark score was still 77.63%, and its pair-level agreement with human annotators on the post-trained model’s videos was 70.3% for action and 72.9% for visual quality, leaving substantial room for ambiguity. The authors also note that action-only training eroded visual quality, while visual-only training gave limited control gains—signals had to be combined. Still, the work shows that a single model can ground both action fidelity and visual plausibility in localized video evidence, a step toward more trustworthy automated evaluation of interactive generation.