If you track the breathless progress of AI development, you will have heard of "reasoning" models and the breakthroughs they have made.
There is much debate about this, much of it semantic, but the likely answer is no—at least not in the way humans reason.
LLM models contain a huge amount of information. One of the key goals of prompting is trying to help the LLM find the information most relevant to our request. Reasoning models automatically traverse a larger area of the information stored during the model's training, making it more likely to encounter information relevant to our query and thus give better, more accurate answers.
Some of these models give feedback while they are reasoning, sometimes appearing to provide very human-like updates on discoveries and new understanding. This feedback is part of the design of these models and maps the underlying process of AI "reasoning" to terms humans find acceptable and understandable. It's a UX and marketing sleight of hand that causes confusion.
A peek at the reasoning streams coming from more open models like DeepSeek shows the reasoning process in a more raw format, and it's a jumble that generally makes little sense.
The best parallel is with the use of the word "know" with the original LLM models. Although models may contain information about a particular topic and do appear to "know" the topic in a very human way, they don't know the topic the way a human expert does—they just do a very good impersonation. The difference at times seems tenuous, but it is vitally important to understand if you want to get the most from LLMs.
Reasoning or not, the "reasoning" models are a major advance over previous models for certain tasks. Why don't we use them for everything then?
Reasoning models have disadvantages. They can be much slower than previous models—all of that "reasoning" or traversing over their trained knowledge takes time and compute. That extra compute incurs additional costs, creates environmental impact, and most noticeably, makes them feel slow. Slow frustrates users.
In education, reasoning models are best at tricky questions where the methods required to reach an answer are not clearly revealed in the question. Answering these types of questions requires a multistep process of figuring out what methods map to the question, extracting the relevant values, then applying the selected methods in the correct order on the correct values. For more straightforward questions, a reasoning model is overkill and results in a poorer experience for the student. How should you choose then? Tutello does it for you—it will analyze the conversation with a student and decide when it might be best to employ a "reasoning" model.
