This repository serves as a comprehensive, curated collection of state-of-the-art methods for achieving "long-to-short" reasoning in Large Reasoning Models (LRMs). It targets researchers, engineers, and power users seeking to optimize LRM efficiency by reducing reasoning output length without compromising accuracy. The primary benefit is a centralized, structured overview of novel techniques, facilitating rapid assessment and adoption of methods for more concise and cost-effective LRM inference.

How It Works

The repository categorizes various strategies designed to enable Large Reasoning Models (LRMs) to produce concise outputs. These methods aim to reduce the length of reasoning chains while maintaining or improving accuracy. Categories include Prompt Guidance (using explicit instructions in prompts), Reward Guidance (reinforcement learning for length optimization), Length-Agnostic Optimization, Latent Space Compression (replacing reasoning tokens with compressed representations), Routing Strategies (task-specific reasoning paths), and Model Distillation/Merge (training smaller models or combining parameters). This structured approach highlights diverse paradigms for achieving efficient LRM reasoning.

Quick Start & Requirements

This repository is a curated list of research papers and code, not a runnable software project. Therefore, no installation commands or specific prerequisites are listed for the repository itself.

Highlighted Details

• Comprehensive Taxonomy: Organizes "long-to-short" reasoning methods into distinct categories like Prompt Guidance, Reward Guidance, Latent Space Compression, Routing Strategies, and Model Distillation, providing a structured overview of the research landscape.
• Active Research Focus: Features a significant number of recent publications (2024-2025), reflecting the dynamic and evolving nature of efficient LRM reasoning research.
• Code Accessibility: Provides direct links to code repositories for many listed papers, enabling practical implementation and empirical validation of presented methods.
• Diverse Methodologies: Encompasses a wide array of techniques, from prompt engineering and reinforcement learning to latent space manipulation and model merging, offering a broad perspective on efficiency optimization.

Maintenance & Community

No specific information regarding maintenance status, contributors, or community channels (like Discord/Slack) is present in the provided README.

Licensing & Compatibility

No licensing information is provided in the README. Compatibility for commercial use or closed-source linking cannot be determined from the given text.

Limitations & Caveats

This repository serves as a curated list of research papers and associated resources, not a deployable software project. Users must independently evaluate and implement individual methods. Performance claims and practical effectiveness are specific to each cited paper and are not aggregated or benchmarked at the repository level. No licensing information or community support details are provided within the README.