No. AI learns to predict reasons, and doing so as it predicts the answer improves its accuracy at predicting the answer.
In summary, even though they are called "reasoning" models, they are still based on prediction and pattern matching, not true logical reasoning. The improvement in accuracy is likely due to better leveraging of the model's statistical knowledge, rather than any deeper understanding of the problem's logic. And the reasons you see it output have nothing to do with the actual reasons it used to determine the answer.
In fact, R1.Zero hints that, it might be even better to let the AI follow a chain of thought that doesn't actually make logical sense or is understandable, and that doing so could even further improve its ability to accurately predict solutions to code, math and logic problems.
Yes, that's what OpenAI o1 does, and DeepSeek R1. Also Google Gemini 2.0 Thinking models. It's a way to significantly improve benchmark scores, especially in math.
It's funny to watch too. I played with Gemini 2.0 on Google AI Studio and asked it to "come up with your favorite song as you take a long walk to really think this through".
The reasoning can then be shown, and it talked to itself, saying things like "since I'm an AI, I can't take walks, but with a request like this, the user seems to imply that I should choose something that's introspective and meaningful", and went on with how it picked candidates.
I just tried that prompt with gemini-2.0-flash-thinking-exp-01-21
In the reasoning process it concludes on:
From the brainstormed genres/artists, select a specific song. It's better to be concrete than vague. For this request, "Nuvole Bianche" by Ludovico Einaudi emerges as a strong candidate.
Craft the Explanation and Scenario: Now, build the response around "Nuvole Bianche."
Then in the actual answer it proposes:
"Holocene" by Bon Iver.
