[bensyverson]

If you view the "thinking" traces you can see why; it will go back and forth on potential solutions, writing full implementations in the thinking block then debating them, constantly circling back to points it raised earlier, and starting every other paragraph with "Actually…" or "But wait!"

[nothinkjustai]

I see this with Opus too.

[girvo]

Indeed. And that’s with Anthropic hiding reading traces unlike these other comparisons.

[FlyingSnake]

> "Actually…" or "But wait!"

You’re absolutely right!

Jokes apart, I did notice GLM doing these back and forth loops.

[tonyarkles]

I 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.

[Lerc]

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.

[tonyarkles]

That's pretty cool, thanks for the extra context! (pardon the... not even pun I guess)

Also, thanks for pointing me at that specific paper; I spend a lot more of my life closer to classical control theory than ML theory so it's always neat to see the intersection of them. My unsubstantiated hypothesis is that controls & ML are going to start getting looked at more holistically, and not in the way I normal see it (which is "why worry about classical control theory, just solve the problem with RL"). Control theory is largely about steering dynamic systems along stable trajectories through state space... which is largely what iterative "fill in the next word" LLM models are doing. The intersection, I hope, will be interesting and add significant efficiency.