This repository is a curated, regularly updated paper list focused on "Efficient Reasoning" in Large Language Models (LLMs). It serves researchers and practitioners by cataloging advancements in techniques that reduce computational cost and improve reasoning performance, covering areas like compression, distillation, sampling, and scaling.

How It Works

The list categorizes papers based on specific efficiency techniques applied to LLM reasoning. It covers broad survey papers, methods for efficient training (e.g., curriculum learning, RL), latent reasoning approaches, compression strategies, step decomposition, distillation for smaller models, collaboration between model sizes, and various test-time scaling and sampling methods. The organization allows users to quickly find relevant research across different efficiency paradigms.

Quick Start & Requirements

This is a paper list, not a software library. No installation or execution is required. Users can browse the categorized links to papers, code repositories, and related resources.

Highlighted Details

• Comprehensive coverage of efficient reasoning techniques, including surveys, training, inference, and analysis.
• Categorization by specific methods like Latent Chain-of-Thought, Compression, Balanced CoT, Reasoning Shortcuts, and Test-Time Scaling.
• Includes links to papers, code, blogs, talks, and other related reading lists for deeper exploration.
• Features recent publications, with many entries dated 2024 and 2025, reflecting the cutting edge of the field.

Maintenance & Community

The list is presented as "regularly updated" and encourages contributions from the community to include missed works. Links to related "Awesome" lists are provided, indicating community engagement.

Licensing & Compatibility

Not applicable, as this is a curated list of research papers and resources.

Limitations & Caveats

The list is a compilation of external research and does not provide any executable code or models itself. The quality and applicability of the listed papers depend on their original sources.