This repository serves as a comprehensive survey of research on efficient reasoning models for Large Language Models (LLMs). It aims to catalog and organize papers, techniques, and benchmarks related to making LLM reasoning more efficient, targeting researchers and practitioners in the field of natural language processing and artificial intelligence.

How It Works

The repository categorizes research into various approaches for achieving efficient reasoning, including fine-tuning methods (SFT-based, RL-based), prompt-driven techniques, latent reasoning, distillation, quantization, pruning, and multimodal reasoning. It provides a structured overview of the landscape, linking to papers and code where available.

Quick Start & Requirements

This repository is a curated list of research papers and does not require installation or execution. It serves as a reference guide.

Highlighted Details

• Extensive Categorization: Organizes efficient reasoning methods into fine-grained categories like "Make Long CoT Short," "Prompt-driven Methods," "Latent Reasoning," and "Efficient Multimodal Reasoning."
• Up-to-date Research: Includes recent advancements and papers, with updates noted for new benchmarks and sections.
• Contribution Focused: Encourages community contributions to add new papers, update details, or suggest improvements via pull requests or email.
• Links to Resources: Provides direct links to arXiv papers and HuggingFace repositories for many listed methods.

Maintenance & Community

The repository is actively maintained, with recent updates and acknowledgments of contributors. It welcomes Pull Requests for additions and corrections.

Licensing & Compatibility

The repository itself is likely under a permissive license (e.g., MIT, Apache 2.0) as it is a collection of links and information. Individual papers and code repositories will have their own licenses.

Limitations & Caveats

As a survey, this repository does not provide executable code or models itself. The effectiveness and implementation details of the listed methods depend on the original research papers and their associated codebases.