Giving models reasoning ability is difficult and costly. Here is a quick summary of some common techniques to infuse reasoning capability in models.
a) Pretraining gives the model the basic ability to work through a problem step by step because it sees a lot of math and code and solved examples. The base model can reason but not reliably.
b) Fine tuning models on long reasoning examples gives the ability to reason first and then answer. This step teaches the format of reasoning.
c) Reinforcement learning is where the real reasoning ability comes from. For each problem the model makes many attempts and the final correct answer is rewarded.
For example if the problem is a coding task the model writes many solutions and each one runs against unit tests and only the passing attempts get a reward. The model learns that checking its own work and going back to try another path gets more reward.
d) The length of reasoning is an important criteria to for models to learn. Meta adds a penalty on reasoning length during training and the model first reasons longer and then learns to solve the same problems with much less tokens. Muse Spark Open these levels map to 16 and 64 and 256 and 512.
e) Models like Kimi K3 and Nemotron 3 ultra train different models for each domain and then merge into one final model with multi-teacher distillation.
f) At inference the reasoning tokens and final answer tokens go as separate fields in server response. The reasoning attempts can also be done in-parallel the way Muse Spark Open does it in Contemplating mode.