Fudan-Tencent WorldReward tops GPT-5.5 as world-model judge

A new reward model from Fudan University and Tencent Hunyuan beats GPT-5.5 at judging camera-conditioned world-model video, and the margins are not narrow: 3.42 percentage points on action consistency, 1.45 on appearance, 3.56 on motion.

The WorldReward paper targets a specific pain point in world models that respond to commanded camera actions. Existing rewards, the authors write, split the job in two: "geometry-based rewards estimate trajectory execution but cannot judge the visual quality of the executed motion, whereas image-based rewards measure frame quality without capturing action execution or temporal dynamics." WorldReward is a VLM-based pairwise preference model that scores both dimensions at once.

The trick is chunk-level structure. Feeding a whole video and its full action sequence to a VLM produces a context in which, the paper reports, "short-lived local action evidence can be missed or diluted." WorldReward instead decomposes paired videos into action-aligned chunks, organizes each chunk into structured visual evidence, then votes chunk-level decisions into separate video-level preferences on action and visual quality.

Training used roughly 100,000 chunk-level examples derived from 50,000 video pairs, with labels seeded by a frontier VLM and refined via "multi-turn tool-based agent auditing and targeted human review." The companion benchmark, WorldReward-Bench, holds 760 paired comparisons with three independent annotators each.

Wired into RL post-training of Tencent's HY-WorldPlay 1.5, the reward lifted long-horizon combined-action accuracy 1.58 points over the WorldCompass baseline and basic-action accuracy 2.28 points. It lands during an active week of world-model releases we've been tracking, including SolarWM's open weights and Tencent-backed H3-World.