Toward Efficient and Faithful Reasoning in Large Language Models

Large Language Models (LLMs) have demonstrated remarkable performance across a wide range of tasks, from natural language understanding to arithmetic reasoning and code generation. However, enabling these models to reason efficiently—achieving high performance with minimal computational overhead and maximal interpretability—remains an open challenge. This survey presents a comprehensive overview of methodologies for building efficient reasoning models with LLMs. We categorize the landscape into prompt-based methods (e.g., chain-of-thought, self-consistency), architectural and tool-augmented enhancements (e.g., retrieval-augmented generation, program-aided reasoning, memory systems), and training-time techniques (e.g., distillation, curriculum learning). We also review evaluation protocols and benchmark datasets that capture diverse reasoning requirements, from symbolic logic and mathematical problem solving to multi-hop question answering. In addition to characterizing the trade-offs between accuracy and inference cost, we highlight emerging trends in neuro-symbolic integration, adaptive computation, lifelong learning, and interpretable reasoning. We conclude by identifying open challenges and future directions toward general-purpose reasoning agents. This survey aims to serve both as a structured map of recent developments and a call to advance reasoning efficiency as a first-class objective in the next generation of LLM research.