I tried out a GLM-5.3 model and the thinking trace was full of "wait, what if I'm running in this specific benchmark? Then the expectation is foo"