chapter one

1 Understanding reasoning models

 

This chapter covers

  • What "reasoning" means for a Large Language Model
  • Reviewing the conventional pre-training and post-training stages of LLMs
  • Introducing key approaches to improving reasoning abilities in LLMs
  • How reasoning differs from pattern matching
  • Why we should build reasoning models from scratch

Welcome to the next stage of large language models (LLMs): reasoning. LLMs have transformed how we process and generate text, but their success has been largely driven by statistical pattern recognition. New advances in reasoning methods now enable LLMs to tackle more complex tasks, such as solving logical puzzles and multi-step math problems.

Moreover, reasoning is an essential technique for making AI agents practical, for example when an agent has to break a task into steps, use tools, and recover from mistakes. Reasoning LLMs are already used in agent applications such as OpenClaw.

In Build a Reasoning Model (From Scratch), you will learn the inner workings of LLM reasoning methods through a hands-on, code-first approach. We will start from a pre-trained LLM and extend it step by step with reasoning capabilities. We implement these reasoning components ourselves, from scratch, to see how these methods work in practice.

1.1 Defining reasoning in the context of LLMs

1.2 Understanding the standard LLM training pipeline

1.3 Improving LLM reasoning with training and inference techniques

1.4 Pattern matching versus logical reasoning

1.5 Simulating reasoning without explicit rules

1.6 Why build reasoning models from scratch?

1.7 A roadmap to reasoning models from scratch

1.8 Summary