There is a belief that learning to compress well will lead to intelligence.Recently, language modeling has been shown to be equivalent to compression,which offers a compelling rationale for the success of large language models(LLMs): the development of more advanced language models is essentiallyenhancing compression which facilitates intelligence. Despite such appealingdiscussions, little empirical evidence is present for the interplay betweencompression and intelligence. In this work, we examine their relationship inthe context of LLMs, treating LLMs as data compressors. Given the abstractconcept of "intelligence", we adopt the average downstream benchmark scores asa surrogate, specifically targeting intelligence related to knowledge andcommonsense, coding, and mathematical reasoning. Across 12 benchmarks, ourstudy brings together 30 public LLMs that originate from diverse organizations.Remarkably, we find that LLMs' intelligence -- reflected by average benchmarkscores -- almost linearly correlates with their ability to compress externaltext corpora. These results provide concrete evidence supporting the beliefthat superior compression indicates greater intelligence. Furthermore, ourfindings suggest that compression efficiency, as an unsupervised metric derivedfrom raw text corpora, serves as a reliable evaluation measure that is linearlyassociated with the model capabilities. We open-source our compression datasetsas well as our data collection pipelines to facilitate future researchers toassess compression properly.
URL
Affiliations
Abstract
Translation (by gpt-3.5-turbo)
Summary (by gpt-3.5-turbo)