LLM-Microscope: Uncovering the Hidden Role of Punctuation in Context Memory of Transformers

TL;DR

This paper reveals that minor tokens like punctuation and stopwords play a crucial role in LLMs' contextual memory, and introduces LLM-Microscope, a toolkit for analyzing token-level contributions and model representations.

Contribution

It uncovers the hidden importance of filler tokens in LLMs' context encoding and provides an open-source toolkit for detailed analysis of token and layer contributions.

Findings

Removing minor tokens degrades model performance.

High correlation between contextualization and linearity in embeddings.

Toolkit enables visualization and measurement of token and layer contributions.

Abstract

We introduce methods to quantify how Large Language Models (LLMs) encode and store contextual information, revealing that tokens often seen as minor (e.g., determiners, punctuation) carry surprisingly high context. Notably, removing these tokens -- especially stopwords, articles, and commas -- consistently degrades performance on MMLU and BABILong-4k, even if removing only irrelevant tokens. Our analysis also shows a strong correlation between contextualization and linearity, where linearity measures how closely the transformation from one layer's embeddings to the next can be approximated by a single linear mapping. These findings underscore the hidden importance of filler tokens in maintaining context. For further exploration, we present LLM-Microscope, an open-source toolkit that assesses token-level nonlinearity, evaluates contextual memory, visualizes intermediate layer contributions (via an adapted Logit Lens), and measures the intrinsic dimensionality of representations. This toolkit illuminates how seemingly trivial tokens can be critical for long-range understanding.