arXiv cs.CL· Albert Qin, Jonathan D. Cohen, Gautam Reddy·· 11 小时前AI 评分48
大语言模型中的认知地图：LLM 如何表征关系结构
Cognitive Maps in Large Language Models
AI 导读
研究通过关系推理任务与因果“mosaic”实验发现，参数量超过百亿的大语言模型能推断未明示的关系，并泛化到合成实体名称。模型将关系结构抽象表征并与实体身份表征相加组合，关系主要以成对链接存储、回溯式组合，而非前瞻性编码进全局坐标系。稀疏分布的一组键值对承担绑定作用，扰动这些键值对可植入“虚假记忆”并引发可预测的行为变化。
正文 · 原文
Abstract:It has long been proposed that animals construct cognitive maps to support flexible behavior in novel environments. We tested whether large language models form analogous internal representations using relational reasoning tasks and causal "mosaic" experiments. Models beyond the ten-billion-parameter scale inferred unstated relationships and generalized to synthetic entity names. Across models and tasks, models formed abstract representations of relational structure, additively combining them with representations of entity identity. Relationships were primarily stored as pairwise links and composed retrospectively, rather than prospectively encoded in a global coordinate system. A sparse, distributed set of key-value pairs mediated binding; perturbing them implanted "false memories" that produced predictable behavioral changes. Our findings reveal a conserved organization analogous to a cognitive map supporting flexible and systematic reasoning in language models.
| Comments: | Changes in v2: Revised Discussion, minor edits throughout, and added a summary figure (Fig. 6). Results unchanged |
| Subjects: | Computation and Language (cs.CL); Machine Learning (cs.LG) |
| Cite as: | arXiv:2610.25596 [cs.CL] |
| (or arXiv:2610.25596v2 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.25596 arXiv-issued DOI via DataCite |
Submission history
From: Albert Qin [view email]
[v1]
Mon, 19 Oct 2026 00:02:30 UTC (12,475 KB)
[v2]
Fri, 6 Nov 2026 14:33:05 UTC (12,702 KB)
来源：arXiv cs.CL · arxiv.org