الرؤوس التحريضية في المحولات

The canonical circuit composes attention heads across layers. An earlier previous-token head writes information about a position's predecessor into that position's representation. A later induction head can match the current token A to a prior position carrying predecessor-A information and promote the token B at that position. Thus the search locates the following-token position, not merely the earlier A itself. The 2022 study reports a relationship between induction-head formation, training-loss behavior, and increased in-context learning in examined models, including stronger causal evidence in small attention-only settings. It presents broader large-model explanations as a hypothesis rather than proving that this two-head circuit accounts for every task or architecture. Applying a learned circuit to prompt examples does not require updating model weights during that inference.