随着人工智能技术的迅速发展，多模态大模型（LMMs）在处理跨领域信息的能力上取得了显著进步。这些模型，如GPT-4o，能够整合视觉、文本等不同模态的信息，展现出一定的推理和理解能力，在视觉问答、图像生成、跨模态检索等任务中表现出色。这一领域的研究热点在于如何严谨地评估人工智能模型的推理能力，特别是在解决数学问题时，其表现与人类的差异尤为显著。
近期，一篇由来自北京邮电大学、腾讯微信、华中科技大学、北京理工大学的研究人员共同撰写的论文在AIxiv专栏上发布，揭示了大模型在解数学题时与人类的显著差异。研究团队通过构建一个包含6.5千个多模态小学数学问题的数据集，以及一个多层级的知识架构，旨在探索大模型的解题机制与人类认知过程的异同。
论文指出，大模型在解决复杂问题时展现出较强的推理能力，能够回答出较为复杂的问题。然而，在面对一些简单问题时，模型往往表现出知识欠缺，显得捉襟见肘。这一现象促使研究者对大模型的解题过程进行深入分析，发现模型的解题策略与人类存在明显差异。
研究团队提出了一种基于人类解题思维模式的模型构建方法，首先通过构建一个包含67个原子知识点的多层级知识体系，为模型提供必要的知识提示。接着，他们将复杂问题拆解为多个原子知识点对应的子问题，以此评估模型的作答机制。
通过这一研究，我们得以一窥大模型在数学解题方面的表现与人类认知的差异，以及其在知识泛化和推理能力上的局限。GPT-4o作为研究中表现最佳的模型，展示了在多模态信息整合和复杂问题解决方面的潜力。然而，其在简单问题上的知识欠缺问题，也揭示了当前大模型在知识深度和广度上的挑战。
这一研究不仅为人工智能领域提供了宝贵的数据和理论支持，也为未来提升大模型在数学解题能力方面提供了方向。通过深入理解人类解题过程，未来的研究可以进一步优化大模型的训练和设计，使其在数学以及其他复杂问题解决方面取得更大进展。
英语如下：
### Large Models Tackle Math: Performance vs. Human Reasoning, GPT-4o Shines
As artificial intelligence (AI) technology rapidly advances, multi-modal large models (LMMs) have made significant strides in their ability to handle cross-domain information. These models, such as GPT-4o, are capable of integrating information from various modalities like visuals and text, demonstrating a degree of reasoning and comprehension. They excel in tasks like visual question answering, image generation, and cross-modal retrieval. A key area of research interest is the rigorous evaluation of AI models’ reasoning capabilities, particularly in solving math problems, where their performance contrasts markedly with that of humans.
A recent paper, authored by researchers from Beijing University of Posts and Telecommunications, Tencent WeChat, Huazhong University of Science and Technology, and Beijing Institute of Technology, and published on the AIxiv column, sheds light on the discrepancies between large models and human performance in math problem-solving. The research team constructed a dataset comprising 6,500 multi-modal elementary math problems, along with a hierarchical knowledge structure, aiming to explore the differences between the problem-solving mechanisms of large models and human cognitive processes.
The paper highlights that large models exhibit strong reasoning capabilities when tackling complex problems, capable of formulating answers to intricate questions. However, they often display a lack of knowledge in answering simple questions, appearing to struggle. This phenomenon prompts a deeper analysis of the model’s problem-solving process, revealing significant differences from human approaches.
The researchers proposed a method for building models based on human problem-solving thinking patterns. This involved constructing a multi-level knowledge system containing 67 atomic knowledge points to provide necessary hints to the model. They then broke down complex problems into sub-problems corresponding to atomic knowledge points to evaluate the model’s answering mechanism.
Through this study, we gain insight into the differences between large models and human cognition in math problem-solving, as well as their limitations in knowledge generalization and reasoning ability. GPT-4o, as the model that performed best in the study, showcases potential in multi-modal information integration and complex problem-solving. However, its shortcomings in knowledge deficiencies for simple problems reveal the current challenges faced by large models in terms of knowledge depth and breadth.
This research not only provides valuable data and theoretical support for the AI field but also paves the way for future advancements in large models’ math problem-solving abilities. By delving into the human problem-solving process, future research can further optimize the training and design of large models, enabling them to make greater strides in math and other complex problem-solving tasks.
【来源】https://www.jiqizhixin.com/articles/2024-07-23-3
Views: 1