Those ‘lower quality’ models can be more accurate yhan the expensive chain of thought ones for certain simpler problems. The chain of thought ones are liable to overthink simpler problems.
They don’t think at all. What they’re more likely to do is hallucinate an answer for a question that doesn’t perfectly fit the training model. That’s not “overthinking”. That’s failing.
Those ‘lower quality’ models can be more accurate yhan the expensive chain of thought ones for certain simpler problems. The chain of thought ones are liable to overthink simpler problems.
They don’t think at all. What they’re more likely to do is hallucinate an answer for a question that doesn’t perfectly fit the training model. That’s not “overthinking”. That’s failing.
Well that’s what it’s called in the academic literature.
https://arxiv.org/abs/2412.21187
that’s not academic literature, that’s unreviewed preprint