Too Long, Didn't Model: Decomposing LLM Long-Context Understanding With Novels
This addresses the need for better evaluation methods for LLMs in long-context scenarios, particularly for developers, but is incremental as it builds on existing computational novel analysis work.
The paper tackles the problem of evaluating large language models (LLMs) on complex long-context understanding beyond simple benchmarks by introducing the TLDM benchmark based on novels, and finds that none of seven tested frontier LLMs retain stable understanding beyond 64k tokens.
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.