fscdc/Awesome-Efficient-Reasoning-Models

What it solves

Large reasoning models (LLMs that think step-by-step before answering) are powerful but slow and expensive. They often generate long chains of thought, which costs compute and time. This project collects and organizes the latest research on how to keep the reasoning accuracy while reducing the cost — for example by shortening the chain of thought, using smaller models, or making the decoding process more efficient.

How it works

This repository is a living, paper-based survey. It groups hundreds of research papers into clear categories:

• Make Long CoT Short — methods to compress or shorten the chain-of-thought (CoT) that reasoning models produce, using fine-tuning, reinforcement learning, or prompt tricks.
• Build SLM with Strong Reasoning Ability — approaches to distill powerful reasoning into small language models (SLMs), or to reduce model size via quantization and pruning.
• Let Decoding More Efficient — techniques to speed up the actual token-generation step, like efficient test-time scaling or better decoding strategies.
• Efficient Multimodal Reasoning — reducing the cost when models reason about images or other modalities.
• Efficient Agentic Reasoning — making agent-like multi-step reasoning more efficient.
• Evaluation and Benchmarks — benchmarks and tests that help measure efficiency gains.

Each paper entry includes a figure, a short introduction, links to the GitHub repo and arXiv paper. The taxonomy diagram and a full list of quick links help readers navigate. The maintainers actively update the list and accept pull requests to add new papers.

Who it’s for

Researchers, engineers, and students working on large language models — especially anyone interested in making reasoning models faster, cheaper, or more practical to deploy. It’s also useful for anyone wanting a broad overview of the current state of efficient reasoning research, and for those who want to contribute their own papers to a public, citable collection.

Highlights

• Covers both training-time methods (SFT, RL) and inference-time tricks (decoding, pruning) in one place.
• Structured with clear sections and a visual taxonomy, making it easy to browse by approach.
• Includes many papers with GitHub links and illustrative figures, so you can quickly assess and reproduce.
• Actively maintained — new sections (like efficient agentic reasoning) are added over time, and it welcomes community contributions via pull requests.