"LLM performance can become less reliable as input contexts grow, even when models nominally support long context windows."
Chroma evaluated 18 language models across tasks designed to test their use of increasingly long inputs. The report documents performance deterioration and inconsistent use of context, supporting the concern that expanding an agent's working history does not guarantee reliable comprehension.
View Source