context distraction
A long context pulling a model towards repeating its history or attending to similar but irrelevant passages instead of the task.
In more detail
Distractors, passages that look relevant but do not answer, lower accuracy faster than plain length does. Prune and summarise rather than letting history and search results pile up.
Related terms
- context rot — The fall in a model's accuracy as its input grows longer, even on simple tasks and well within the context window.
- context poisoning — A mistake or false claim entering the context, for example from a hallucination or a bad tool result, and being repeated as fact on later turns.
Taught in
- When Context Fails — AI Engineering: LLM Apps, RAG & Agents
- One Leaked Kubeconfig: How Far Does It Go? — Kubernetes and Cloud Native Security Associate (KCSA) Exam Prep
- Eleven Months Of Yes: The API Server — Kubernetes and Cloud Native Security Associate (KCSA) Exam Prep
- Open To Every Interface: Controller Manager And Scheduler — Kubernetes and Cloud Native Security Associate (KCSA) Exam Prep
- Created, But Never Running: Pod Security Admission — Kubernetes and Cloud Native Security Associate (KCSA) Exam Prep
- There Is No Users Table: Authentication — Kubernetes and Cloud Native Security Associate (KCSA) Exam Prep
References
- Drew Breunig: How Long Contexts Fail (2025-06-22)
- Chroma: Context Rot: How Increasing Input Tokens Impacts LLM Performance (2025-07-14)