PuzzleVQA
收藏arXiv2025-09-30 收录
下载链接：
https://github.com/declare-lab/LLM-PuzzleTest
下载链接
链接失效反馈官方服务：
资源简介：
该数据集名为PuzzleVQA，旨在评估大型多模态模型在抽象推理能力方面的表现。在多项选择题的子集上，人类的性能达到了91.4%的得分，而模型o1在此数据集上取得了模型中的最高分，达到了79.2%。该任务的目标是评估多模态模型在抽象推理能力方面的表现。
This dataset, named PuzzleVQA, is designed to evaluate the abstract reasoning capabilities of large multimodal models. On its multiple-choice subset, human performance attained an accuracy score of 91.4%, while model o1 achieved the highest score across all evaluated models on this dataset, reaching 79.2%. The objective of this task is to assess the abstract reasoning performance of multimodal models.
提供机构：
OpenAI
