Foundations and Base Model Selection for Reinforcement Learning-Based Reasoning Enhancement: A Systematic Analysis

The emergence of DeepSeek-R1 (DeepSeek-AI, 2025) has demonstrated that large language models (LLMs) can develop sophisticated reasoning capabilities through pure reinforcement learning (RL), without reliance on human-annotated reasoning trajectories. This finding fundamentally alters the economics and accessibility of frontier AI development, reducing the barrier from billions of dollars to hundreds of thousands. However, the practical implementation of these techniques remains poorly documented outside of the original research teams. This paper series presents a comprehensive, reproducible laboratory manual for constructing reasoning-enhanced language models using the methodology described in the DeepSeek-R1 technical report. In this first paper, we establish the theoretical foundations, analyze the critical factors in base model selection, and present empirical criteria for evaluating candidate models prior to RL training. We demonstrate that model capacity, pre-training data composition, and architectural properties are deterministic factors in RL success, and provide a decision framework applicable to resource-constrained settings. Our analysis suggests that organizations with budgets as low as $500-$50,000 USD can produce reasoning models competitive with or superior to commercial offerings in targeted domains.