I'm not saying it applies to the new architecture, I'm saying that's a big issue I've observed in existing models and that so far we have no info on whether it's solved in the new one (i.e. accurate recall doesn't imply much in that regard).
Ah, apologies for the misunderstanding. What tests would you suggest to evaluate "muddiness"?
What comes to my mind: run the usual gamut of tests, but with the excess context window saturated with irrelevant(?) data. Measure test answer accuracy/verbosity as a function of context saturation percentage. If there's little correlation between these two variables (e.g. 9% saturation is just as accurate/succinct as 99% saturation), then "muddiness" isn't an issue.
Manual testing on complex documents. A big legal contract for example. An issue can be referred to in 7 different places in a 100 page document. Does it give a coherent answer?
A handful of examples show whether it can do it. For example, GPT-4 turbo is downright awful at something like that.
You need to use relevant data. The question isn't random sorting/pruning, but being able to apply large numbers of related hints/references/definitions in a meaningful way. To me this would be the entire point of a large context window. For entirely different topics you can always just start a new instance.
