meituan/ViC-Bench

ViC-Bench是一个专门用于评估大型语言模型（MLLMs）在视觉交错的思维链（VI-CoT）方面的能力的数据集。它包括四个代表性任务：迷宫导航、拼图、具身长期规划和复杂计数。每个任务都通过三个阶段进行处理，每个阶段都提供中间视觉状态，以帮助模型进行决策。此外，数据集还包括一个增量提示信息注入（IPII）策略，以探索提示因素对VI-CoT的影响。README文件中还包括每个阶段的数据样本，并提供了数据集的引用。

ViC-Bench is a specialized dataset designed to evaluate the Visual-Interleaved Chain-of-Thought (VI-CoT) capability in MLLMs. It includes four representative tasks: maze navigation, jigsaw puzzle, embodied long-horizon planning, and complex counting. Each task is processed through three stages, with intermediate visual states provided at each stage to assist models in decision-making. Additionally, the dataset includes an Incremental Prompting Information Injection (IPII) strategy to explore the impact of prompting factors on VI-CoT. The README provides data samples for each stage and includes a citation for the dataset.