How Do LLMs Acquire New Knowledge? A Knowledge Circuits Perspective on Continual Pre-Training
Fuente:
arXiv
Guardado en:
| Autores principales: | , , , , , , , |
|---|---|
| Formato: | Preprint |
| Publicado: |
2025
|
| Materias: | |
| Acceso en línea: | |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| _version_ | 1866912407731830784 |
|---|---|
| author | Ou, Yixin Yao, Yunzhi Zhang, Ningyu Jin, Hui Sun, Jiacheng Deng, Shumin Li, Zhenguo Chen, Huajun |
| author_facet | Ou, Yixin Yao, Yunzhi Zhang, Ningyu Jin, Hui Sun, Jiacheng Deng, Shumin Li, Zhenguo Chen, Huajun |
| contents | Despite exceptional capabilities in knowledge-intensive tasks, Large Language Models (LLMs) face a critical gap in understanding how they internalize new knowledge, particularly how to structurally embed acquired knowledge in their neural computations. We address this issue through the lens of knowledge circuit evolution, identifying computational subgraphs that facilitate knowledge storage and processing. Our systematic analysis of circuit evolution throughout continual pre-training reveals several key findings: (1) the acquisition of new knowledge is influenced by its relevance to pre-existing knowledge; (2) the evolution of knowledge circuits exhibits a distinct phase shift from formation to optimization; (3) the evolution of knowledge circuits follows a deep-to-shallow pattern. These insights not only advance our theoretical understanding of the mechanisms of new knowledge acquisition in LLMs, but also provide potential implications for improving continual pre-training strategies to enhance model performance. Code and data will be available at https://github.com/zjunlp/DynamicKnowledgeCircuits. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2502_11196 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | How Do LLMs Acquire New Knowledge? A Knowledge Circuits Perspective on Continual Pre-Training Ou, Yixin Yao, Yunzhi Zhang, Ningyu Jin, Hui Sun, Jiacheng Deng, Shumin Li, Zhenguo Chen, Huajun Machine Learning Artificial Intelligence Computation and Language Computer Vision and Pattern Recognition Human-Computer Interaction Despite exceptional capabilities in knowledge-intensive tasks, Large Language Models (LLMs) face a critical gap in understanding how they internalize new knowledge, particularly how to structurally embed acquired knowledge in their neural computations. We address this issue through the lens of knowledge circuit evolution, identifying computational subgraphs that facilitate knowledge storage and processing. Our systematic analysis of circuit evolution throughout continual pre-training reveals several key findings: (1) the acquisition of new knowledge is influenced by its relevance to pre-existing knowledge; (2) the evolution of knowledge circuits exhibits a distinct phase shift from formation to optimization; (3) the evolution of knowledge circuits follows a deep-to-shallow pattern. These insights not only advance our theoretical understanding of the mechanisms of new knowledge acquisition in LLMs, but also provide potential implications for improving continual pre-training strategies to enhance model performance. Code and data will be available at https://github.com/zjunlp/DynamicKnowledgeCircuits. |
| title | How Do LLMs Acquire New Knowledge? A Knowledge Circuits Perspective on Continual Pre-Training |
| topic | Machine Learning Artificial Intelligence Computation and Language Computer Vision and Pattern Recognition Human-Computer Interaction |
| url | https://arxiv.org/abs/2502.11196 |