To produce accurate predictions, language models (LMs) must balance between
generalization and memorization. Yet, little is known about the mechanism by
which transformer LMs employ their memorization capacity. When does a model
decide to output a memorized phrase, and how is this phrase then retrieved from
memory? In this work, we offer the first methodological framework for probing
and characterizing recall of memorized sequences in transformer LMs. First, we
lay out criteria for detecting model inputs that trigger memory recall, and
propose idioms as inputs that typically fulfill these criteria. Next, we
construct a dataset of English idioms and use it to compare model behavior on
memorized vs. non-memorized inputs. Specifically, we analyze the internal
prediction construction process by interpreting the model's hidden
representations as a gradual refinement of the output probability distribution.
We find that across different model sizes and architectures, memorized
predictions are a two-step process: early layers promote the predicted token to
the top of the output distribution, and upper layers increase model confidence.
This suggests that memorized information is stored and retrieved in the early
layers of the network. Last, we demonstrate the utility of our methodology
beyond idioms in memorized factual statements. Overall, our work makes a first
step towards understanding memory recall, and provides a methodological basis
for future studies of transformer memorization.

本文提供了第一种探测 transformer LMs 中记忆序列召回的方法，并分析了记忆预测是一个两步过程，其存储和检索存储在网络的早期层中，这为理解记忆召回提供了第一步，并为未来的 transformer 记忆研究提供了方法基础。