Out-of-distribution generalization via composition: A lens through induction heads in Transformers

• 成果类型: Article
• 署名作者: Song, Jiajun; Xu, Zhuoyan; Zhong, Yiqiao
• 署名单位: University of Wisconsin System; University of Wisconsin Madison
• 刊物名称: PROCEEDINGS OF THE NATIONAL ACADEMY OF SCIENCES OF THE UNITED STATES OF AMERICA
• ISSN/ISSBN: 0027-14331
• DOI: 10.1073/pnas.2417182122
• 发表日期: 2025-02-11

摘要

Large language models (LLMs) such as GPT-4 sometimes appear to be creative, solving novel tasks often with a few demonstrations in the prompt. These tasks require the models to generalize on distributions different from those from training data-which is known as out-of-distribution (OOD) generalization. Despite the tremendous success of LLMs, how they approach OOD generalization remains an open and underexplored question. We examine OOD generalization in settings where instances are generated according to hidden rules, including in-context learning with symbolic reasoning. Models are required to infer the hidden rules behind input prompts without any finetuning. We empirically examined the training dynamics of Transformers on a synthetic example and conducted extensive experiments on a variety of pretrained LLMs, focusing on a type of component known as induction heads. We found that OOD generalization and composition are tied together-models can learn rules by composing two self- attention layers, thereby achieving OOD generalization. Furthermore, a shared latent subspace in the embedding (or feature) space acts as a bridge for composition by aligning early layers and later layers, which we refer to as the common bridge representation hypothesis.