Thinking before answering
A reasoning model, at a conceptual level, is a model trained or prompted to produce extended intermediate reasoning steps before it commits to a final answer, rather than jumping straight from question to answer. Instead of predicting the answer token by token with no visible intermediate work, it first generates a sequence of reasoning tokens, working through the problem the way a person might work through a proof or a word problem on scratch paper, and only then produces its final response.
This isn't a different kind of neural network architecture so much as a different behavior pattern layered on top of the same underlying transformer-style models covered elsewhere in this platform. What changes is how the model is trained or guided to use its output sequence, spending part of it on working through the problem rather than only on stating a conclusion.