Context Poisoning
A phenomenon observed in long-context scenarios where a large language model's ability to locate and utilize key evidence degrades significantly when the input is mixed with large quantities of distractors that share similar formatting and semantics. The underlying cause is attributed to extreme-value interference within the attention mechanism.
Timeline (last 90 days)
'Context Poisoning' refers to erroneous, outdated or poorly structured data entering the context window; the Agent reasons based on this data and errs without any malicious behavior being required.
Context poisoning is usually a precision problem—the relevant records may all be there, but they are drowned out by a large number of irrelevant records, making it hard for the model to weight them correctly.
A paper published on arXiv coined the term "context poisoning" to describe the phenomenon in which long-context models exhibit degraded ability to locate key evidence when distractor content is mixed in.
The paper points out that the root cause of long-context degradation lies not entirely in context length, but in the number of effective distractors it contains.
All Facts (4)
'Context Poisoning' refers to erroneous, outdated or poorly structured data entering the context window; the Agent reasons based on this data and errs without any malicious behavior being required.
50%UnverifiedContext poisoning is usually a precision problem—the relevant records may all be there, but they are drowned out by a large number of irrelevant records, making it hard for the model to weight them correctly.
50%UnverifiedA paper published on arXiv coined the term "context poisoning" to describe the phenomenon in which long-context models exhibit degraded ability to locate key evidence when distractor content is mixed in.
50%UnverifiedThe paper points out that the root cause of long-context degradation lies not entirely in context length, but in the number of effective distractors it contains.
50%