【北京讯】近日，国际机器学习顶级会议ICML2024上，一场关于大语言模型（LLM）的演讲引起了广泛关注。Meta公司研究员朱泽园在演讲中揭秘了大语言模型在解数学题方面的能力，指出这些不同于人类的二级推理LLM是如何在数学领域展现其能力。
朱泽园的研究团队通过实验发现，大语言模型在解数学题时，不仅能够通过模板记忆，还能够学会推理思维。他们的研究论文《语言模型物理学 Part 2.1：小学数学与隐藏的推理过程》在arXiv上发表，详细展示了模型解题的心算过程和推理技能。
该研究团队创建了一个模拟小学数学级别的思维题集iGSM，让模型从零开始在iGSM上预训练，从而控制模型接触的问题类别。通过这个数据集，他们发现模型不仅能够学会1级推理，即通过拓扑排序寻找最短解答，还能够在没有见过的问题上展现出推理能力。
这一发现不仅揭示了大语言模型在解数学题方面的能力，还为通用智能的发展提供了新的思路。尽管模型在某些高难度题目上仍然会犯推理错误，但这项研究为理解LLM的推理能力提供了新的视角，也为未来模型的改进和优化提供了方向。
推特网友@xlr8harder评价称，这一结果将一劳永逸地平息关于LLM是否具有推理能力的争论，证明了这些模型不仅仅是随机生成的鹦鹉，而是真正具备了推理和学习的能力。
这一研究成果对于人工智能领域来说是一个重要的进展，它不仅加深了我们对大语言模型的理解，也为未来的技术发展提供了新的可能性。随着研究的深入，我们有理由相信，大语言模型在解数学题和其他复杂任务上的能力将会得到进一步提升。
英语如下：
News Title: “ICML Sensation: Meta Unveils the Secret to Mathematical Reasoning in Large Models”
Keywords: Model Reasoning, Mathematical Problem Solving, AI Progress
News Content:
[Beijing, China] Recently, a presentation at the International Conference on Machine Learning (ICML 2024), a top international machine learning conference, has attracted widespread attention. Zheyuan Zhu, a researcher from Meta, unveiled the capability of large language models (LLMs) in solving mathematical problems during his speech, highlighting how these LLM models, unlike humans, demonstrate their reasoning abilities in the field of mathematics.
Zhu’s research team discovered through experiments that large language models not only rely on template memorization when solving mathematical problems but also learn to think in a reasoning manner. Their research paper, “Language Modeling Physics Part 2.1: Elementary Mathematics and the Hidden Reasoning Process,” was published on ArXiv, providing a detailed demonstration of the model’s mental calculation process and reasoning skills.
The research team developed a simulated elementary school mathematics-level thinking problem dataset called iGSM, allowing the model to pre-train from scratch on iGSM, thereby controlling the types of problems the model encounters. Through this dataset, they found that the model not only learns to perform first-order reasoning, such as finding the shortest solution through topological sorting, but also exhibits reasoning abilities on problems it has never seen before.
This discovery not only reveals the capability of large language models in solving mathematical problems but also offers new insights into the development of general intelligence. Although the model still makes reasoning errors on certain challenging problems, this research provides a new perspective on understanding LLM reasoning capabilities and offers directions for future model improvement and optimization.
Twitter user @xlr8harder commented that this result will finally put to rest the debate over whether LLM models possess reasoning capabilities, proving that these models are not merely random parrots but truly have the ability to reason and learn.
This research achievement is a significant progress in the field of artificial intelligence, deepening our understanding of large language models and providing new possibilities for future technological development. As research progresses, there is reason to believe that the ability of large language models to solve mathematical problems and other complex tasks will be further enhanced.
【来源】https://www.jiqizhixin.com/articles/2024-08-04-4
Views: 1