WorldReward-Bench
收藏资源简介:
WorldReward-Bench是由复旦大学、腾讯混元等机构联合构建的人类标注基准数据集,旨在评估相机条件世界模型的奖励模型。该数据集包含760对配对的视频生成结果,覆盖多种轨迹类型、视觉风格及世界模型来源,每对视频均附带独立的人类标注偏好,涵盖动作一致性、外观质量和运动质量三个维度。数据集通过多轮工具辅助代理审核与人工校准精心创建,确保标注的可靠性与细粒度。其核心应用在于为统一评估动作执行与视觉质量的奖励模型提供标准测试平台,解决现有奖励模型难以同时衡量这两方面耦合需求的难题。
WorldReward-Bench is a human-annotated benchmark dataset jointly constructed by Fudan University, Tencent Hunyuan and other institutions, aiming to evaluate reward models for camera-conditioned world models. This dataset includes 760 pairs of matched video generation results, covering diverse trajectory types, visual styles and world model sources. Each pair of videos is accompanied by independent human annotation preferences, which span three dimensions: action consistency, appearance quality and motion quality. It was meticulously curated through multi-round tool-assisted agent review and manual calibration, ensuring the reliability and fine-grained nature of the annotations. Its core application is to provide a standardized testbed for the unified evaluation of reward models that assess both action execution and visual quality, addressing the challenge that existing reward models struggle to simultaneously quantify the coupled requirements of these two aspects.
数据集概述
WorldReward 是一个基于视觉语言模型(VLM)的成对偏好奖励模型,专为相机条件世界模型(Camera-Conditioned World Models)设计,用于统一评估生成视频中的动作一致性和视觉质量。
核心思想
WorldReward 将成对视频分解为动作对齐的片段(action-aligned chunks),并将每个片段组织为结构化视觉证据,从而评估单个动作的执行情况以及视觉质量。片段级决策通过投票聚合为全局偏好。
训练数据构建
训练数据采用推理增强的偏好数据(Reasoning-augmented preference data),其构建流程如下:
- 由前沿 VLM 生成结构化判断
- 通过多轮基于工具(tool-based)的智能体审计(agent auditing)进行细化
- 结合针对性的人工审查(human review)进行精修
基准测试
WorldReward-Bench 是一个人工标注的基准数据集,用于衡量奖励模型与人类偏好的对齐程度,涵盖以下三个维度:
- 动作一致性(Action)
- 外观质量(Appearance)
- 运动质量(Motion)
该基准包含来自 9 个世界模型的 760 个人工标注视频对,覆盖多种轨迹、视觉风格和来源模型。
应用效果
实验表明,WorldReward 在动作一致性和视觉质量评估上优于现有的开源奖励模型和专有 VLM 评测器。将其用于 HY-WorldPlay 1.5 的强化学习(RL)后训练时,可有效提升动作执行效果和生成视频的视觉质量。




