arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2605.08142cs.LGcs.CLcs.CV
收藏

推理源自大型语言模型中受限推断流形

Reasoning emerges from constrained inference manifolds in large language models

Renmin University of China(中国人民大学);Tsinghua University(清华大学);Xiaomi EV(小米电动车);Xiamen University(厦门大学)

Yanbiao Ma, Fei Luo, Linfeng Zhang, Chuangxin Zhao, Mingxuan Wang, Yinan Wu, Zhe Qian, Yang Lu, Long Chen, Zhao Cao, Xiaoshuai Hao, Ji-Rong Wen, Jungong Han

AI总结:

研究通过分析推理过程中的内部表示演变,发现推理动态在高维空间中自组织为低维流形,但需满足表示能力、流形压缩和非退化信息体积等条件才能实现稳定推理。

摘要:

大型语言模型中的推理主要通过标注基准评估,将任务表现与内部推断质量混淆。本文通过研究推理作为内在动态过程,分析推理过程中内部表示的演变。发现推理时间动态一致自组织为嵌入在高维表示空间中的低维流形。尽管这种几何压缩普遍存在,但不足以实现稳定或可靠的推理。有效的推理动态出现在由三个条件定义的受限结构域内:足够的表示表达性、自发的流形压缩以及压缩子空间内非退化的信息体积。域外模型表现出特征性的病态推理动态。基于这些见解,我们引入了一个统一的、无标注的诊断方法,仅从内部动态计算。这些发现表明,LLM中的推理本质上由几何和信息约束决定,为基于基准的评估提供了补充框架。

英文摘要:

Reasoning in large language models is predominantly evaluated through labeled benchmarks, conflating task performance with the quality of internal inference. Here we study reasoning as an intrinsic dynamical process by examining the evolution of internal representations during inference. We find that inference-time dynamics consistently self-organize into low-dimensional manifolds embedded within high-dimensional representation spaces. we find that such geometric compression, although pervasive, is not sufficient for stable or reliable reasoning. Instead, effective reasoning dynamics emerge within a constrained structural regime characterized by three conditions: adequate representational expressivity, spontaneous manifold compression, and preservation of non-degenerate information volume within the compressed subspace. Models outside this regime exhibit characteristic pathological inference dynamics. Based on these insights, we introduce a unified, label-free diagnostic computed solely from internal dynamics. These findings suggest that reasoning in LLMs is fundamentally governed by geometric and informational constraints, offering a complementary framework to benchmark-centric assessment.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑