smgjch/MeowBench

MeowBench是一个高保真、经过专家验证的四模态基准测试，旨在评估多模态大型语言模型（MLLMs）在猫科动物意图解码方面的能力。它通过将外部观察数据（视频/音频）与内部生物标记（ECG/EEG/IMU）相关联，解决动物行为中的“语义别名”问题，测试模型是否能够进行真正的潜在状态推理。数据集结构为多项选择题（MCQ），每个样本包括视频、音频和时间序列数据的同步（或意图匹配）三元组，以及一个自然语言提示和四个选项。数据集的构建和验证过程包括下一步行为预测（NBP）逻辑、意图匹配合成和由八位专业猫科动物行为学家进行的手动审核。

MeowBench is a high-fidelity, expert-verified quad-modal benchmark designed to evaluate Multimodal Large Language Models (MLLMs) on feline intention decoding. It provides a rigorous testing ground for models to determine if they can move beyond superficial pattern matching to genuine latent state reasoning by correlating external observational data (video/audio) with internal biological markers (ECG/EEG/IMU). Each sample in MeowBench is structured as a Multiple Choice Question (MCQ), including a synchronized (or intent-matched) triplet of Video, Audio, and Time-Series data, a natural language prompt, and four options. The datasets construction and verification process includes Next-Behaviour Prediction (NBP) Logic, Intent-Matched Synthesis, and manual review by eight Professional Feline Ethologists.