IME all "reasoning" models do is confuse themselves, because the underlying problem of hallucination hasn't been solved. So if the model produces 10K tokens of "reasoning" junk, the context is poisoned, and any further interaction will lead to more junk.
I've had much better results from non-"reasoning" models by judging their output, doing actual reasoning myself, and then feeding new ideas back to them to steer the conversation. This too can go astray, as most LLMs tend to agree with whatever the human says, so this hinges on me being actually right.
I've had much better results from non-"reasoning" models by judging their output, doing actual reasoning myself, and then feeding new ideas back to them to steer the conversation. This too can go astray, as most LLMs tend to agree with whatever the human says, so this hinges on me being actually right.