Through the development of AutoRace, a fully automated reasoning chain evaluation tool, and LLM Reasoners, a library for diverse reasoning algorithms, this research provides comprehensive insights and methodologies to enhance step-by-step reasoning in LLMs.

Key Points:

• Introduction of AutoRace for automatic reasoning evaluation.
• Development of LLM Reasoners for standardized reasoning implementation.
• Extensive studies on search, reward, and world model components in reasoning.

Significance: This paper advances our understanding of LLMs’ operational capabilities in systematic reasoning, promising significant impact on AI’s problem-solving potentials and ethical considerations.