alt Hacker NewsI was watching Qwen3.6-35B-A3B (locally) doing the same dance yesterday. It eventually finished and had a reasonable answer, but it sure went back and forth on a bunch of things I had explicitly said not to do before coming to a conclusion. At least said conclusion was not any of the things I'd said not to do.
That is essentially what the reasoning reinforcement training does. It is getting the model to say things that are more likely to result in the correct final answer. Everything it does in between doesn't necessarily need to be valid argument to produce the answer. You can think of it as filling the context with whatever is needed to make the right answer come out next. Valid arguments obviously help. but so might expressions of incorrect things that are not obviously untrue to the model until it sees them written out. The What's The Magic Word paper shows how far that could go. If the policy model managed to learn enough magic words it would be theoretically possible to end up with an LLM that spouts utter gibberish until delivering the correct answer seemingly out of the blue.