跳到正文
arXiv cs.AI· Matteo He, William F. Shen, Alex Iacob, Andrej Jovanovic, Xinchi Qiu, Nicholas D. Lane·· 4 小时前AI 评分48

语言模型预训练中的 Unembedding 动力学:行为出现时间可能晚于能力习得

Learning to Read Out: Unembedding Dynamics in Language Model Pretraining

AI 导读

一项 NeurIPS 2026 研究提出参数轨迹 crosscoding 方法,通过对每个 checkpoint 的 readout 行拟合稀疏字典,直接追踪语言模型预训练中 unembedding 矩阵的演化。

正文 · 原文

View PDF HTML (experimental)

Abstract:When does a language model acquire a capability? Developmental studies usually answer this by tracking when models start producing the correct token. But that prediction passes through the unembedding matrix, or readout, which is itself still being learned during pretraining. We show that a model's hidden states can encode task information before its own readout expresses it, so timelines read from outputs conflate what the model encodes with what its readout expresses. To track the readout directly, we introduce parameter-trajectory crosscoding, which fits one sparse dictionary to every token's readout row at every checkpoint. Unlike crosscoders on activations, it needs only checkpoint weights. We use it to locate a narrow early window, 0.4-1.1% of the way through Pythia's pretraining, in which the readout reorganizes and WordNet word categories become decodable from its rows. By then, linear probes on the hidden states already decode task information that the model's own outputs do not yet show. To test whether such gaps depend on how fast the readout learns, we pretrain small models from scratch that differ only in the readout's learning rate. We find that raising it makes the readout develop earlier and the gap close sooner, with little effect on final loss. Together, these results imply that early in training, where many developmental transitions are dated, a behavioral date can reflect when the readout learns to express task information and not when the model learns it.
Comments: NeurIPS 2026. 10 pages main text, 51 pages total. Code: this https URL
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
ACM classes: I.2.7; I.2.6
Cite as: arXiv:2610.21025 [cs.CL]
  (or arXiv:2610.21025v2 [cs.CL] for this version)
  https://doi.org/10.48550/arXiv.2610.21025

arXiv-issued DOI via DataCite

Submission history

From: Matteo He [view email]
[v1] Wed, 14 Oct 2026 23:45:26 UTC (1,892 KB)
[v2] Fri, 13 Nov 2026 14:15:42 UTC (1,892 KB)

来源:arXiv cs.AI · arxiv.org