BUILDING REASONING LLMS FROM SCRATCH: A Practical Guide to Transformers, Reinforcement Learning, GRPO, Test-Time Scaling, and Training AI Models to Reason - Softcover

TRENT WILDER, LOGAN

9798173210197: BUILDING REASONING LLMS FROM SCRATCH: A Practical Guide to Transformers, Reinforcement Learning, GRPO, Test-Time Scaling, and Training AI Models to Reason

Softcover

ISBN 13: 9798173210197

Publisher: Independently published, 2026

View all copies of this ISBN edition

0 Used

3 New

From US$ 36.56

BUILDING REASONING LLMS FROM SCRATCH

A Practical Guide to Transformers, Reinforcement Learning, GRPO, Test-Time Scaling, and Training AI Models to Reason

How do modern AI models learn to reason, solve complex problems, verify their answers, and improve their performance with additional computation?

BUILDING REASONING LLMS FROM SCRATCH takes you inside the technology behind modern reasoning Large Language Models (LLMs) and provides a practical path from the fundamentals of language modeling to advanced reasoning techniques.

This book goes beyond simply using existing AI models. You will learn how reasoning LLMs are designed, trained, evaluated, optimized, and prepared for real-world applications.

Inside this practical guide, you will discover:

• How tokens, embeddings, neural networks, attention, and Transformers work together to power modern LLMs.

• How to build and train a Transformer-based language model from the ground up.

• How pretraining, supervised fine-tuning, and instruction tuning transform a language model into a more capable AI assistant.

• How chain-of-thought reasoning, synthetic reasoning data, self-verification, and structured problem solving can improve reasoning capabilities.

• How reinforcement learning can be applied to language models through RLHF and Reinforcement Learning with Verifiable Rewards (RLVR).

• How reward functions and reward models work—and why reward hacking can cause unexpected or undesirable behavior.

• How Group Relative Policy Optimization (GRPO) works and why it has become an important technique for training reasoning models.

• How probability ratios, advantage estimation, KL divergence, and policy optimization contribute to reinforcement learning for LLMs.

• How test-time scaling, Best-of-N sampling, self-consistency, search, and verification can improve inference-time reasoning.

• How to balance accuracy, reasoning quality, compute, latency, throughput, and cost.

• How KV caching, quantization, memory optimization, reasoning distillation, and model compression can make reasoning systems more efficient.

• How to evaluate reasoning LLMs using accuracy, reliability, benchmarks, calibration, and failure analysis.

• How to design a complete reasoning LLM pipeline from data preparation and training to evaluation, optimization, deployment, and monitoring.

WHO THIS BOOK IS FOR

This book is designed for AI engineers, machine learning practitioners, LLM developers, software developers, students, researchers, and anyone who wants to understand the technology behind modern reasoning AI.

Whether you are beginning your journey into LLM development or already have experience with machine learning, this book provides a structured progression from foundational concepts to advanced reasoning techniques.

FROM FUNDAMENTALS TO ADVANCED REASONING

You will progress through the complete reasoning LLM development lifecycle:

Transformers → Pretraining → Instruction Tuning → Reasoning Data → RLVR → GRPO → Test-Time Scaling → Evaluation → Optimization → Deployment

By the end of this book, you will have a deeper understanding of how reasoning language models work and the key techniques used to build, train, evaluate, and optimize them.

If you want to move beyond simply using AI and start understanding how reasoning LLMs are actually built, this book gives you the knowledge and practical foundation to begin.

"synopsis" may belong to another edition of this title.