Mechanism and emergence of stacked attention heads in multi-layer transformers

ICLR 2025 Conference Submission427 Authors

Keywords: mechanistic interpretability, large language models, transformers, emergent abilities, curriculum learning

Abstract: In this paper, we introduce the retrieval problem, a simple reasoning task that can be solved only by transformers with a minimum number of layers. The task has an adjustable difficulty that can further increase the required number of layers to any arbitrary value. We demonstrate that large language models can solve the task under different prompting formulations without any fine-tuning. To understand how transformers solve the retrieval problem, we train several transformers on a minimal formulation. We find that successful learning occurs only under the presence of an implicit curriculum. We uncover the learned mechanisms by studying the attention maps in the trained transformers. We also study the training process, uncovering that attention heads always emerge in a specific sequence.

Primary Area: interpretability and explainable AI

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: 427