Can Perplexity Reflect Large Language Model's Ability in Long Text Understanding?

TL;DR

This work questions the assumption that perplexity (PPL) on long texts reflects an LLM's true long-text understanding. By evaluating three long-context window variants on downstream long-text tasks (QA, summarization) and a retrieval benchmark, the authors show a lack of consistent correlation between PPL and downstream performance, suggesting PPL mainly captures local information modeling. They also demonstrate that a model with low PPL can still underperform on long-text understanding, and attribute this to local-information bias exemplified by methods like ALiBi. The paper advocates using diversified evaluation metrics beyond PPL to assess long-text processing capabilities and cautions against over-reliance on PPL for claiming long-text understanding.

Abstract

Recent studies have shown that Large Language Models (LLMs) have the potential to process extremely long text. Many works only evaluate LLMs' long-text processing ability on the language modeling task, with perplexity (PPL) as the evaluation metric. However, in our study, we find that there is no correlation between PPL and LLMs' long-text understanding ability. Besides, PPL may only reflect the model's ability to model local information instead of catching long-range dependency. Therefore, only using PPL to prove the model could process long text is inappropriate. The local focus feature of PPL could also explain some existing phenomena, such as the great extrapolation ability of the position method ALiBi. When evaluating a model's ability in long text, we might pay more attention to PPL's limitation and avoid overly relying on it.