Too Long, Didn't Model: Decomposing LLM Long-Context Understanding With Novels

TL;DR

This paper introduces the TLDM benchmark to evaluate large language models' ability to understand complex, long-range narrative structures in novels over 128k tokens, revealing current models' limitations beyond 64k tokens.

Contribution

The paper presents the TLDM benchmark for assessing long-context understanding in LLMs and demonstrates that existing models struggle with stable comprehension beyond 64k tokens.

Findings

None of the tested models maintain stable understanding beyond 64k tokens.

Long-range semantic dependencies in novels are challenging for current LLMs.

The benchmark and reference code are released for future research.

Abstract

Although the context length of large language models (LLMs) has increased to millions of tokens, evaluating their effectiveness beyond needle-in-a-haystack approaches has proven difficult. We argue that novels provide a case study of subtle, complicated structure and long-range semantic dependencies often over 128k tokens in length. Inspired by work on computational novel analysis, we release the Too Long, Didn't Model (TLDM) benchmark, which tests a model's ability to report plot summary, storyworld configuration, and elapsed narrative time. We find that none of seven tested frontier LLMs retain stable understanding beyond 64k tokens. Our results suggest language model developers must look beyond "lost in the middle" benchmarks when evaluating model performance in complex long-context scenarios. To aid in further development we release the TLDM benchmark together with reference code and data.