protologue

Context Rot

Also called long-context degradation, context degradation.

Context rot is the decline in a model's reliability as the amount of input in its context grows, even on simple tasks and well within its advertised context window.

Description

A 2025 study of 18 models found performance degrading with input length in ways that depended on the task and on the content around the relevant information, such as similar-looking distractors. It is a practical reason to keep contexts lean, retrieve only what is needed, and compact long histories rather than relying on maximum context length.

Sources

  1. Hong et al. (2025). Context Rot: How Increasing Input Tokens Impacts LLM Performance.

Cite this entry

Protologue. (2026). Context Rot. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.4.0, PTL-0169). https://protologue.com/t/context-rot/

BibTeX
@misc{protologue_context_rot,
  title = {Context Rot},
  author = {{Protologue}},
  year = {2026},
  howpublished = {Protologue: A Taxonomy of Prompting and LLM Techniques, v1.4.0},
  note = {Entry PTL-0169},
  url = {https://protologue.com/t/context-rot/}
}

Markdown JSON