首次从数学上证明了Transformer训练中语法到语义的分阶段学习机制。

Transformers may exhibit two-stage training dynamics during the real-world training process. For instance, when training GPT-2 on the Counterfact dataset, the answers progress from syntactically incorrect to syntactically correct to semantically correct. However, existing theoretical analyses hardly account for this feature-level two-stage phenomenon, which could be conceptually attributed to disentangled two-type features like syntax and semantics. In this paper, we theoretically demonstrate how the two-stage training dynamics potentially occur in transformers. Specifically, we analyze the feature learning dynamics induced by the aforementioned disentangled two-type feature structure, grounding our analysis in a simplified yet illustrative setting that comprises normalized ReLU self-attention and structured data. Such disentanglement of feature structure is general in practice, e.g., natural languages contain syntax and semantics, and proteins contain primary and secondary structures. To our best knowledge, this is the first rigorous result regarding a feature-level two-stage optimization process in transformers within this theoretical framework. A corollary further indicates that such a two-stage process is closely related to the spectral properties of attention weights.