MMReason
收藏资源简介:
MMReason 是一个开放式的多模态多步推理基准数据集,旨在推动多模态大型语言模型(MLLMs)向通用人工智能(AGI)发展。该数据集包含来自不同学科(如数学、商业、科学、工程、社会科学和健康)的各种难度级别的挑战性问题,这些问题需要多步推理和批判性思维。为了确保推理的可靠性,数据集中的问题已被重新制定为开放式问题,并通过多模态投票技术进行筛选,以消除与猜测和记忆相关的捷径情况。此外,MMReason 还为每个问题提供了详细的分步解决方案,并设计了一个基于参考的三元评分机制,以可靠地评估 MLLMs 的中间推理步骤。
MMReason is an open-ended multimodal multi-step reasoning benchmark dataset designed to advance Multimodal Large Language Models (MLLMs) toward Artificial General Intelligence (AGI). This dataset comprises challenging questions spanning diverse difficulty levels across various disciplines, including mathematics, business, science, engineering, social sciences, and health, all of which require multi-step reasoning and critical thinking. To ensure the reliability of reasoning, the questions in this dataset have been reformulated into open-ended questions and screened via multimodal voting techniques to eliminate shortcut cases associated with guessing and memorization. Furthermore, MMReason provides detailed step-by-step solutions for each question, and has designed a reference-based tripartite scoring mechanism to reliably evaluate the intermediate reasoning steps of MLLMs.
MMReason: 开放型多模态多步推理基准数据集
数据集概述
- 名称:MMReason
- 全称:An Open-Ended Multi-Modal Multi-Step Reasoning Benchmark for MLLMs Toward AGI
- 类型:多模态多步推理基准数据集
- 目标:面向通用人工智能(AGI)的多模态大语言模型(MLLMs)评估
关键特征
- 开放型问题设计
- 多模态输入支持
- 多步推理能力评估
- 面向AGI的基准测试
相关资源
- 论文发布日期:2025年7月1日
- 论文链接:https://arxiv.org/abs/2506.23563

- 1MMReason: An Open-Ended Multi-Modal Multi-Step Reasoning Benchmark for MLLMs Toward AGI南洋理工大学, 清华大学, 百度公司, 加利福尼亚大学, 中国科学技术大学 · 2025年



