Mastering Syntax, Unlocking Semantics: A Mathematically Provable Two-stage Learning Process in Transformers

26 Sept 2024 (modified: 09 Dec 2024)ICLR 2025 Conference Withdrawn SubmissionEveryoneRevisionsBibTeXCC BY 4.0
Keywords: Two-stage learning, Optimization dynamics, Feature learning theory
TL;DR: Rigorously prove that transformers follow a syntax-then-semantics learning paradigm
Abstract: Transformers have emerged as a cornerstone across various fields with extensive applications. However, the training dynamics of transformers remain relatively underexplored. In this work, we present a novel perspective on how transformers acquire knowledge during the training dynamics, inspired by the feature learning theory. To this end, we conceptualize each token as embodying two types of knowledge: elementary knowledge represented by syntactic information, and specialized knowledge represented by semantic information. Building on this data structure, we rigorously prove that transformers follow a syntax-then-semantics learning paradigm, i.e., first mastering syntax in the Elementary Stage and then unlocking semantics in the subsequent Specialized Stage. The results are derived from the training dynamics analysis and finite-time convergence within the in-context learning framework for supervised classification. To our best knowledge, this is the \textbf{\emph{first}} rigorous result of a two-stage optimization process in transformers from a feature learning perspective. Empirical findings on real-world language datasets support the theoretical results of the two-stage learning process. Moreover, the spectral properties of attention weights, derived from our theoretical framework, align with the experimental observations, providing further validation.
Primary Area: learning theory
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 7086
Loading

OpenReview is a long-term project to advance science through improved peer review with legal nonprofit status. We gratefully acknowledge the support of the OpenReview Sponsors. © 2025 OpenReview