This repository provides a comprehensive survey of efficient reasoning techniques for Large Reasoning Models (LRMs), targeting researchers and engineers working with LLMs. It aims to consolidate and categorize recent advancements in making LRMs more efficient, addressing the growing need for optimized performance and resource utilization in complex reasoning tasks across language, multimodality, and agent systems.

How It Works

The survey categorizes efficient reasoning methods across the LRM development pipeline: pre-training, supervised fine-tuning (SFT), reinforcement learning (RL), and inference. It highlights techniques like length budgeting, model switching, model merging, reasoning chain compression, latent-space SFT, and RL with length rewards. The core advantage of this structured approach is its holistic view, enabling a deep understanding of how efficiency can be integrated at various stages, rather than focusing on isolated optimizations.

Quick Start & Requirements

This is a survey repository, not a runnable codebase. It lists and categorizes research papers. No installation or specific requirements are needed to browse the content.

• Official Survey Paper: https://arxiv.org/pdf/2503.21614
• Related Discussion: https://x.com/suzhaochen0110/status/1905461785693749709?s=46

Highlighted Details

• Covers efficient reasoning across language, multimodality, and agent applications.
• Organizes techniques by development stage: pre-training, SFT, RL, and inference.
• Includes recent papers on adaptive reasoning, multimodal reasoning, and agent efficiency.
• Features benchmarks for efficient reasoning such as MME-CoT, S1-Bench, and DUMB500.

Maintenance & Community

The repository is actively updated with new papers, with recent additions in June 2025. It encourages community contributions to expand the paper list.

Licensing & Compatibility

The repository is licensed under the MIT License, permitting commercial use and integration with closed-source projects.

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 practical application. The rapid pace of research means the survey may not yet include the very latest advancements.