Context Rot
Context Rot（上下文衰减）是指在大型语言模型或AI Agent处理任务时，随着上下文长度持续增加，模型对早期或关键信息的注意力和利用能力逐渐下降的现象。其主要特征包括：模型在长对话或长文档中对重要信息的遗忘或忽略、推理准确性随上下文累积而降低，以及精简上下文往往能带来更优的输出质量。该概念揭示了长上下文处理中注意力机制的局限性。
Core Facts
Timeline (last 90 days)
LLMs exhibit Context Rot: the more tokens loaded, the lower the recall precision; the official recommendation is that a Skill exceeding 500 lines should be split, while practitioners tend to keep it within 200 lines.
Context Rot refers to the phenomenon where, after the context window is filled, the quality of an AI Agent's responses progressively deteriorates.
IBM points out that more context does not mean better performance, and that performance may degrade as context expands, a phenomenon referred to as Context Rot
All Facts (6)
Context Rot refers to the phenomenon where, as the number of conversation turns increases, irrelevant information accumulates and the attention budget becomes filled with low-value content, causing reasoning quality to continuously decline.
65%Unverified上下文腐烂（Context Rot）包含两种不同的故障模式：Distraction（注意力稀释）和Poisoning（因果链污染）
85%UnverifiedLLMs exhibit Context Rot: the more tokens loaded, the lower the recall precision; the official recommendation is that a Skill exceeding 500 lines should be split, while practitioners tend to keep it within 200 lines.
50%UnverifiedContext Rot refers to the phenomenon where, after the context window is filled, the quality of an AI Agent's responses progressively deteriorates.
50%UnverifiedIBM points out that more context does not mean better performance, and that performance may degrade as context expands, a phenomenon referred to as Context Rot
50%Unverified尽管上下文窗口很大，上下文腐化（Context Rot）依然是巨大问题，模型对上下文中不同位置信息的注意力分布不均匀（Lost in the Middle效应）
50%