Hey folks,

Has anyone tried running Qwen3 (both 8B and 235B-A22B) evaluation on LiveCodeBench? I’m trying to reproduce the reported results, but it seems like Qwen3 tends to overthink indefinitely and never outputs the closing </think> token.

Do we need to manually insert </think> after a certain thinking budget, or is there a recommended workaround for this behavior?

Thanks!