Hi, I ran the test and found that many models can only answer 320 questions correctly. Why is that? Why is it 320, and if I enter 330, except for the initial "conetext windows", all subsequent ones fail,

[red-co]

I tested gemini2.0-flash, gemini2.5-flash, gemma3-27b, qwen3_235b-a22b,
among which,2.0-flash, 2.5-flash, qwen235b-a22b, are all exactly 320.

This is my prompt

Here are n five-digit additions in the form of
Qn. xn+yn,

You need to answer in the form of
An. {anwser}
, do not group,

example:
`
A1. 79281
A2. 138779
A3. 139180
...
`

[red-co]

Here, if I don't let it use the prefix of An. , it will crash at the first window size. // Because the rotated embedding lacks anchors to mark the position of the embedding?
For qwen3 235b it is 48,
For gemini2-flash it is 60, and for gemini2.5-flash it is 80,
I call it the window size,
because if you group by a smaller number of this size, you can also get the model to answer correctly.
But the An. prefix is the maximum number that the model can answer correctly.