FINAL-Bench/World-Model

World Model Bench（WM Bench）是首个用于评估世界模型和具身AI系统认知能力的基准数据集。它包含100个场景，通过三个支柱（感知、认知、具身）和10个类别进行综合评分，总分1000分，旨在衡量模型在场景理解、预测推理、威胁类型区分响应、自主情感升级、上下文记忆利用、后威胁自适应恢复、运动情感表达、实时认知-行动性能和身体交换可扩展性等方面的能力。数据集通过JSON格式输入，无需3D环境，仅通过文本I/O进行评估，适用于API、性能指标和实时演示等多种参与轨道。

WM Bench is the worlds first benchmark for evaluating the cognitive capabilities of World Models and Embodied AI systems. It includes 100 scenarios structured around three pillars (Perception, Cognition, Embodiment) and 10 categories, with a total score of 1000. The benchmark assesses capabilities such as scene understanding, prediction-based reasoning, threat-type differentiated response, autonomous emotion escalation, contextual memory utilization, post-threat adaptive recovery, motion-emotion expression, real-time cognitive-action performance, and body-swap extensibility. Evaluation is conducted via text I/O using JSON input, without requiring 3D environments, and supports multiple participation tracks including API-only, text with performance metrics, and live demonstrations.