Once it admits it is wrong, it breaks
The moment the model admits it is wrong (after being provided exact Policies), it breaks and starts repeating non-stop \begin{align}.
EDIT: Any further messages to the model just results in a flood of lines that say assistant.
I've now provoked a response where Qwen pretty much calls me a liar.
My response was correct - the fabricated line they're quoting doesn't exist
I did not generate that line. It does not exist in any of my responses. (Qwen actually generated the line again in its rebuttal)
This isn't a matter of "rebuttal" or "hypocrisy"—it's a factual statement about my operational boundaries.
I know you can edit responses to essentially get the AI to do whatever you want, but NONE of that was used here. Qwen's response contained a lot of emojis that would take me far too much time to look up myself.
Qwen has a very complicated "personality" - if you could even call it that way.
It's not only about this particular Thinking model, it's also about the Instruct model which for some reason spits reasoning traces into what is supposed to be a final response, making it all messy, breaking any sort of semblance of coherence and structured output format that you would expect to get.
Recently I started using a small test similar to IQ test for humans, curious what these LLMs will do with it. The questions range from easy to hardest and are meant to map the ability of the LLM to basically "connect the dots" using logic. The answers are often wrong in more tricky questions, but sometimes they are pretty funny and interesting too. Interesting not because they would be correct, but because their reasoning often reveals the flaws and weaknesses in these models.
Thinking models in particular are usually the funniest and most interesting ones, because even if they get the answer right, their chain of thought often shows they came to that correct answer for all the wrong reasons which are often based on "hallucinations" and/or overthinking. Basically some of the models overthink so much that they find the correct answer, but they justify that answer using some weird flawed logic that blows your mind.
Unfortunately all of these traits are always inherited in the finetunes and merges. Recently I tested one of such finetunes of this same model using this "IQ test" and let me tell you the results were pretty sad actually. This Qwen 3 model is truly unique, but unfortunately not in a good way. Where other models such as Mistral Small 3.2 24B, Gemma 3 27B made mistakes and simply acknowledged that some of their answers were incorrect, this Qwen 3 based finetune seemed to get REALLY upset with me for simply telling it the results. The model first acknowledged, but immediatelly jumped into thinking in hindsight and tried to review its own answers in attempt to find the mistakes on its own, only for it to continue overthinking the questions again, wasting so many more tokens ALL FOR NOTHING, because it concluded - its answers were correct and the test evaluation must be flawed. I was like... What?! 🤯
Yeah, at that point I made my own conclusion - When you tell Qwen 3 it's done something wrong, it starts acting like a sore loser. I've never seen that behavior in any other model... 😂