This repository is a curated survey of research papers focused on improving the efficiency of reasoning in Large Language Models (LLMs). It targets researchers and practitioners in NLP and AI who are looking to understand and implement techniques that reduce computational cost and latency for LLM reasoning tasks. The primary benefit is a structured overview of the field, enabling quicker identification of relevant methods and papers.

How It Works

The survey categorizes efficient LLM reasoning techniques into six main areas: Model-based, Output-based, Input Prompt-based, Reasoning Abilities with Efficient Data and Small Language Models, and Evaluation and Benchmarks. It systematically organizes papers within these categories, providing a taxonomy of approaches that range from reinforcement learning with length rewards to prompt-guided reasoning and model compression. This structured approach helps to demystify the complex landscape of efficient reasoning research.

Quick Start & Requirements

This is a survey repository, not a runnable codebase. It primarily serves as a reference and index of research papers. No installation or execution is required to use this resource.

Highlighted Details

• Comprehensive taxonomy of efficient LLM reasoning techniques.
• Extensive list of research papers with links to their publications.
• Covers a wide range of methods including RL, SFT, latent representation compression, dynamic inference, prompt engineering, and data/model efficiency.
• Includes a dedicated section on evaluation benchmarks for efficient reasoning.

Maintenance & Community

The project was updated on April 22, 2025. Contributions are welcomed via pull requests to add recent papers or suggest improvements. The primary author is Yang Sui.

Licensing & Compatibility

The repository itself does not specify a license. The content is a collection of links to research papers, each with its own licensing and usage terms. Compatibility for commercial use depends on the licenses of the individual papers cited.

Limitations & Caveats

As a survey, this repository does not provide implementations or code for the discussed techniques. Users must refer to the individual papers for code and experimental details. The field is rapidly evolving, and the survey may not yet include the very latest research published after its last update.