arXiv cs.CL· Qing Yao, Sasha Boguraev, Tiago Pimentel, Kyle Mahowald·· 12 小时前AI 评分31
语言模型中句法电路的出现与稀疏化：追踪 pythia 的主谓一致机制
The Emergence and Sparsification of a Syntactic Circuit
AI 导读
研究追踪 pythia-1b 和十个随机种子的 pythia-410m 训练过程，发现主谓一致能力习得较早，初期由弥散分布的神经元承担。行为稳定后机制仍持续重组：pythia-1b 早期由两个注意力头控制，后期改由第三个不同的头负责，且最终电路显著稀疏化。
正文
Abstract:To understand the emergence of grammatical behavior in language models, we track circuits for subject-verb number agreement in pythia-1b and ten seeds of pythia-410m across training. We find that subject-verb agreement is learned early, and is initially handled by a diffuse collection of neurons. However, after agreement behavior has stabilized, the mechanisms responsible for it continue to reorganize. The attention heads which mediate agreement often shift during training: in the case of pythia-1b, behavior is controlled by two heads early in training but by a third (different) head later. Furthermore, the circuit shrinks over training, with pythia-1b's final circuit being strikingly sparse. We take these results to be informative about how linguistic knowledge emerges and is represented in a language model.
| Comments: | 26 pages, 7 figures, EMNLP 2026 |
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.33299 [cs.CL] |
| (or arXiv:2610.33299v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.33299 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Qing Yao [view email]
[v1]
Fri, 23 Oct 2026 21:10:54 UTC (491 KB)
来源：arXiv cs.CL · arxiv.org