在科技与学术的交汇处，一场引人注目的科研突破正在悄然发生。人工智能（AI）作为当前科技领域的前沿技术，正不断挑战并拓展着人类的认知边界。最新研究表明，通过公理训练，大型预训练模型如LLM（Language Model）能够学会因果推理，这一发现不仅在AI领域内激起了波澜，也为数学家、科学家在研究探索过程中引入AI工具提供了新的可能性。
#### 基于公理的因果推理
因果推理，作为人类思维的重要组成部分，是理解世界、进行科学发现和构建理论体系的基础。通过学习特定的公理，模型能够学会如何基于因果关系进行推断。这项研究中，科研人员通过展示一系列因果链给LLM模型，使其在小规模图谱上学习因果传递性公理。令人惊喜的是，经过这样的训练，模型能够将学到的因果推理能力泛化应用到更大、更复杂的图谱上。这意味着，如果让模型学会执行简单的因果推理任务，便有可能将其拓展至更复杂的推理场景。
#### AI在科学中的新角色
这一突破性进展不仅展现了AI在理解复杂系统和进行高级推理方面的潜力，也为AI在科学研究中的应用开辟了新的方向。著名数学家陶哲轩等专家的亲身体验表明，AI工具如GPT等在数学研究、理论验证乃至新发现的探索中扮演着越来越重要的角色。通过与AI的合作，科学家们能够更高效地处理数据、模拟实验、验证假设，甚至在某些情况下，AI能够提出人类未曾想到的理论框架。
#### 公理训练框架：学习因果推理的新范式
此次研究中提出的公理训练框架，是一种基于被动数据的学习方法，旨在通过演示足够数量的因果链，使模型学会公理。这一框架的独特之处在于，它能够使模型学习任意公理，而不受数据集大小的限制。这不仅简化了模型训练的过程，也极大地扩展了AI在科学领域应用的范围和深度。
### 结语
随着AI技术的不断演进，我们有理由相信，未来AI在科学探索、理论构建乃至实践应用中的角色将更加重要。通过公理训练学会因果推理，LLM等AI模型不仅为科学界带来了新的研究工具，也为人类对世界本质的理解提供了新的视角。这一研究成果不仅预示着AI在科学领域中的巨大潜力，也为跨学科合作提供了新的可能，标志着AI与人类智慧融合的新篇章即将开启。
英语如下：
### Groundbreaking Research: LLM Learns Causal Inference Through Axiomatic Training, AI’s Role in Scientific Exploration Reinforced
At the confluence of technology and academia, a notable scientific breakthrough is quietly unfolding. Artificial Intelligence (AI), a cutting-edge technology in the realm of contemporary science, is continuously pushing the boundaries of human cognition. The latest studies indicate that through axiomatic training, large pre-trained models such as LLM (Language Model) can learn causal inference. This discovery not only stirs waves within the AI domain but also opens new possibilities for the introduction of AI tools into the arsenal of mathematicians and scientists in their research endeavors.
#### Causal Inference Through Axioms
Causal inference, a fundamental aspect of human thought, is essential for understanding the world, making scientific discoveries, and constructing theoretical frameworks. By learning specific axioms, models are able to comprehend how to infer based on causal relationships. In this study, researchers trained the LLM model on a series of causal chains, enabling it to learn causal transitivity axioms on small-scale graphs. What’s remarkable is that after such training, the model can generalize its causal inference abilities to larger and more complex graphs. This means that if the model is taught to perform simple causal inference tasks, it could potentially be expanded to more complex reasoning scenarios.
#### AI’s New Role in Science
This breakthrough not only showcases AI’s potential in understanding complex systems and performing advanced reasoning but also opens new avenues for AI’s application in scientific research. Notable mathematician Terence Tao and other experts have firsthand experience of AI tools such as GPT in mathematics research, theory verification, and exploration of new discoveries. Through collaboration with AI, scientists can handle data more efficiently, simulate experiments, validate hypotheses, and, in some cases, AI can propose theoretical frameworks that humans might not have conceived.
#### Axiomatic Training Framework: A New Paradigm for Learning Causal Inference
The axiomatic training framework proposed in this study is a data-driven learning approach that aims to teach models by demonstrating a sufficient number of causal chains. Its uniqueness lies in its ability to enable models to learn any axiom without being constrained by the size of the dataset. This not only simplifies the training process for models but also greatly expands the range and depth of AI’s applications in the scientific domain.
### Conclusion
As AI technology continues to evolve, there is reason to believe that the role of AI in scientific exploration, theory development, and practical applications will become increasingly significant. Through axiomatic training to learn causal inference, AI models like LLM not only provide new research tools for the scientific community but also offer new perspectives on understanding the essence of the world. This research result not only hints at the immense potential of AI in the scientific field but also opens up new possibilities for interdisciplinary collaboration, marking the dawn of a new chapter in the fusion of AI with human wisdom.
【来源】https://www.jiqizhixin.com/articles/2024-07-16-4
Views: 1