Intrinsic Entropy of Context Length Scaling in LLMs

TL;DR

This paper introduces the concept of intrinsic entropy to analyze how context length affects language model performance, providing theoretical insights and experimental validation on natural and synthetic data.

Contribution

It proposes a novel intrinsic entropy framework to explain context length effects and offers practical guidelines for optimal context length based on dataset size.

Findings

Training dataset size influences optimal context length.

Theoretical bounds on context length scaling are established.

Experimental validation supports the proposed framework.

Abstract

Long Context Language Models have drawn great attention in the past few years. There has been work discussing the impact of long context on Language Model performance: some find that long irrelevant context could harm performance, while some experimentally summarize loss reduction by relevant long context as Scaling Laws. This calls for a more thorough understanding of how long context impacts Language Modeling. In this work, we (1) propose to use `Intrinsic Entropy' for explaining the impact of context length on language modeling; and (2) conduct experiments on natural language and synthetic data, validating our proposed theoretical assumptions and deductions. Our theoretical framework can provide practical insights such as establishing that training dataset size dictates an optimal context length and bounds context length scaling for certain cases. We hope our work may inspire new long context Language Models, as well as future work studying the physics of Language Models.