A Survey of Efficient Reasoning for Large Reasoning Models: Language, Multimodality, and Beyond

---

🔔 News

• [2025-05] We include LIMOPro for Efficient and Effective Reasoning in Test-time Scaling.
• [2025-05] We update more papers on adaptive reasoning, which describe how a system/model autonomously switches between long and short reasoning chains based on problem complexity.
• [2025-05] Welcome to our latest paper "Scaling Reasoning, Losing Control", which shows that the longer the reasoning chain, the poorer its instruction-following ability. Therefore, efficient reasoning may also be important for instruction following in LRMs.
• [2025-04] We include AgentPrune, where efficient reasoning is important for agent systems.
• [2025-04] We include benchmarks for Efficient Reasoning: MME-CoT, S1-Bench, DUMB500.
• [2025-04] We add Mamba Reasoning models (e.g M1) and Hybrid models (e.g Mamba-Transformer) in Efficient Reasoning during Pre-training. It is naturally efficient to infer.
• [2025-04] We add a new "Model Merge" category in Efficient Reasoning during Inference. It is feasible to be a promising direction.
• [2025-04] 📢 Our work is reported by both Synced (机器之心) and Zhuanzhi (专知).
• [2025-03] 📢 Our work is reported by both Deep Learning and NLP (深度学习自然语言处理) and Machine Learning and NLP (机器学习算法与自然语言处理).
• [2025-03] We released our survey "A Survey of Efficient Reasoning for Large Reasoning Models: Language, Multimodality, and Beyond". This is the first survey for efficient reasoning of Large Reasoning Models, covering language, multimodality, agent, and applications. We provide several promising future directions in our survey.
• [2025-03] We created this repository to maintain a paper list on Awesome-Efficient-LRM-Reasoning.

---

Citation

If you find our survey useful for your research, please consider citing:

@article{qu2025survey,
  title={A Survey of Efficient Reasoning for Large Reasoning Models: Language, Multimodality, and Beyond},
  author={Qu, Xiaoye and Li, Yafu and Su, Zhaochen and Sun, Weigao and Yan, Jianhao and Liu, Dongrui and Cui, Ganqu and Liu, Daizong and Liang, Shuxian and He, Junxian and others},
  journal={arXiv preprint arXiv:2503.21614},
  year={2025}
}

---

🔥 Table of Contents

• Awesome-Efficient-LRM-Reasoning
  • 👀 Introduction
  • 💭 Efficient Reasoning during Inference
    • Length Budgeting
    • System Switch
    • Model Switch
    • Model Merge
    • Parallel Search
  • 💫 Efficient Reasoning with SFT
    • Reasoning Chain Compression
    • Latent-Space SFT
  • 🧩 Efficient Reasoning with Reinforcement Learning
    • Efficient Reinforcement Learning with Length Reward
    • Efficient Reinforcement Learning without Length Reward
  • 💬 Efficient Reasoning during Pre-training
    • Pretraining with Latent Space
    • Subquadratic Attention
    • Linearization
    • Efficient Reasoning with Subquadratic Attention
  • 🔖 Future Directions
    • Efficient Multimodal Reasoning and Video Reasoning
    • Efficient Test-time Scaling and Infinity Thinking
    • Efficient and Trustworthy Reasoning
    • Building Efficient Reasoning Applications
    • Evaluation and Benchmark

---

📜Content

👀 Introduction

In the age of LRMs, we propose that "Efficiency is the essence of intelligence." Just as a wise human knows when to stop thinking and start deciding, a wise model should know when to halt unnecessary deliberation. An intelligent model should manipulate the token economy, i.e., allocating tokens purposefully, skipping redundancy, and optimizing the path to a solution. Rather than naively traversing every possible reasoning path, it should emulate a master strategist, balancing cost and performance with elegant precision.

To summarize, this survey makes the following key contributions to the literature:

• Instead of offering a general overview of LRMs, we focus on the emerging and critical topic of efficient reasoning in LRMs, providing an in-depth and targeted analysis.
• We identify and characterize common patterns of reasoning inefficiency, and outline the current challenges that are unique to improving reasoning efficiency in large models.
• We provide a comprehensive review of recent advancements aimed at enhancing reasoning efficiency, structured across the end-to-end LRM development pipeline, from pretraining and supervised fine-tuning to reinforcement learning and inference.

---

🚀 Papers

💭 Efficient Reasoning during Inference

Length Budgeting

• Done Is Better than Perfect: Unlocking Efficient Reasoning by Structured Multi-Turn Decomposition
• Concise Reasoning, Big Gains: Pruning Long Reasoning Trace with Difficulty-Aware Prompting
• AdaCtrl: Towards Adaptive and Controllable Reasoning via Difficulty-Aware Budgeting
• Think or Not? Exploring Thinking Efficiency in Large Reasoning Models via an Information-Theoretic Lens
• Dynamic Early Exit in Reasoning Models
• Thought Manipulation: External Thought Can Be Efficient for Large Reasoning Models
• Reasoning Models Can Be Effective Without Thinking
• How Well do LLMs Compress Their Own Chain-of-Thought? A Token Complexity Approach
• Sketch-of-Thought: Efficient LLM Reasoning with Adaptive Cognitive-Inspired Sketching
• Chain of Draft: Thinking Faster by Writing Less
• SafeChain: Safety of Language Models with Long Chain-of-Thought Reasoning Capabilities
• s1: Simple test-time scaling
• Token-budget-aware llm reasoning
• Efficiently Serving LLM Reasoning Programs with Certaindex
• Make every penny count: Difficulty-adaptive self-consistency for cost-efficient reasoning
• Scaling llm test-time compute optimally can be more effective than scaling model parameters
• Concise thoughts: Impact of output length on llm reasoning and cost
• The impact of reasoning step length on large language models
• The benefits of a concise chain of thought on problem-solving in large language models
• Guiding language model reasoning with planning tokens

System Switch

• DynamicMind: A Tri-Mode Thinking System for Large Language Models
• Fast-Slow-Thinking: Complex Task Solving with Large Language Models
• Think More, Hallucinate Less: Mitigating Hallucinations via Dual Process of Fast and Slow Thinking
• Dualformer: Controllable Fast and Slow Thinking by Learning with Randomized Reasoning Traces
• Visual Agents as Fast and Slow Thinkers
• System-1.x: Learning to Balance Fast and Slow Planning with Language Models
• DynaThink: Fast or slow? A dynamic decision-making framework for large language models

Model Switch

• Accelerated Test-Time Scaling with Model-Free Speculative Sampling
• Learning Adaptive Parallel Reasoning with Language Models
• SplitReason: Learning To Offload Reasoning
• SpecReason: Fast and Accurate Inference-Time Compute via Speculative Reasoning
• MixLLM: Dynamic Routing in Mixed Large Language Models
• Closer Look at Efficient Inference Methods: A Survey of Speculative Decoding
• EAGLE-2: Faster Inference of Language Models with Dynamic Draft Trees
• RouteLLM: Learning to Route LLMs with Preference Data
• LayerSkip: Enabling Early Exit Inference and Self-Speculative Decoding
• EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty
• Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads
• Routing to the Expert: Efficient Reward-guided Ensemble of Large Language Models
• Speculative Decoding with Big Little Decoder

Model Merge

• Unlocking efficient long-to-short llm reasoning with model merging

Parallel Search

• Sampling-Efficient Test-Time Scaling: Self-Estimating the Best-of-N Sampling in Early Decoding
• Efficient Test-Time Scaling via Self-Calibration
• Scalable Best-of-N Selection for Large Language Models via Self-Certainty
• Meta-Reasoner: Dynamic Guidance for Optimized Inference-time Reasoning in Large Language Models
• Test-Time Preference Optimization: On-the-Fly Alignment via Iterative Textual Feedback
• Fast Best-of-N Decoding via Speculative Rejection
• TreeBoN: Enhancing Inference-Time Alignment with Speculative Tree-Search and Best-of-N Sampling
• Scaling llm test-time compute optimally can be more effective than scaling model parameters

💫 Efficient Reasoning with SFT

Reasoning Chain Compression

• DRP: Distilled Reasoning Pruning with Skill-aware Step Decomposition for Efficient Large Reasoning Models
• Long-Short Chain-of-Thought Mixture Supervised Fine-Tuning Eliciting Efficient Reasoning in Large Language Models
• Z1: Efficient Test-time Scaling with Code
• Self-Training Elicits Concise Reasoning in Large Language Models
• TokenSkip: Controllable Chain-of-Thought Compression in LLMs
• Stepwise Perplexity-Guided Refinement for Efficient Chain-of-Thought Reasoning in Large Language Models
• C3oT: Generating Shorter Chain-of-Thought without Compromising Effectiveness
• Can Language Models Learn to Skip Steps?
• Distilling System 2 into System 1

Latent-Space SFT

• Beyond Chains of Thought: Benchmarking Latent-Space Reasoning Abilities in Large Language Models
• From Explicit CoT to Implicit CoT: Learning to Internalize CoT Step by Step
• CODI: Compressing Chain-of-Thought into Continuous Space via Self-Distillation
• Token Assorted: Mixing Latent and Text Tokens for Improved Language Model Reasoning
• SoftCoT: Soft Chain-of-Thought for Efficient Reasoning with LLMs
• LightThinker: Thinking Step-by-Step Compression
• Efficient Reasoning with Hidden Thinking
• Training Large Language Models to Reason in a Continuous Latent Space
• Compressed Chain of Thought: Efficient Reasoning Through Dense Representations

🧩 Efficient Reasoning with Reinforcement Learning

Efficient Reinforcement Learning with Length Reward

• How Far Are We from Optimal Reasoning Efficiency?
• ConciseRL: Conciseness-Guided Reinforcement Learning for Efficient Reasoning Models
• When to Continue Thinking: Adaptive Thinking Mode Switching for Efficient Reasoning
• Learn to Reason Efficiently with Adaptive Length-based Reward Shaping
• Incentivizing Dual Process Thinking for Efficient Large Language Model Reasoning
• ARM: Adaptive Reasoning Model
• ShorterBetter: Guiding Reasoning Models to Find Optimal Inference Length for Efficient Reasoning
• HAWKEYE: Efficient Reasoning with Model Collaboration
• ThinkPrune: Pruning Long Chain-of-Thought of LLMs via Reinforcement Learning
• Think When You Need: Self-Adaptive Chain-of-Thought Learning
• DAST: Difficulty-Adaptive Slow-Thinking for Large Reasoning Models
• L1: Controlling How Long A Reasoning Model Thinks With Reinforcement Learning
• Demystifying Long Chain-of-Thought Reasoning in LLMs
• Training Language Models to Reason Efficiently
• O1-Pruner: Length-Harmonizing Fine-Tuning for O1-Like Reasoning Pruning
• Kimi k1.5: Scaling Reinforcement Learning with LLMs

Efficient Reinforcement Learning without Length Reward

• Concise Reasoning via Reinforcement Learning
• Optimizing Test-Time Compute via Meta Reinforcement Fine-Tuning
• Think Smarter not Harder: Adaptive Reasoning with Inference Aware Optimization
• Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMs

💬 Efficient Reasoning during Pre-training

Pretraining with Latent Space

• LLM Pretraining with Continuous Concepts
• Scalable Language Models with Posterior Inference of Latent Thought Vectors
• Byte latent transformer: Patches scale better than tokens
• Large Concept Models: Language Modeling in a Sentence Representation Space

Subquadratic Attention

• RWKV-7 "Goose" with Expressive Dynamic State Evolution
• LASP-2: Rethinking Sequence Parallelism for Linear Attention and Its Hybrid
• Native sparse attention: Hardware-aligned and natively trainable sparse attention
• MoBA: Mixture of Block Attention for Long-Context LLMs
• MoM: Linear Sequence Modeling with Mixture-of-Memories
• Gated Delta Networks: Improving Mamba2 with Delta Rule
• Transformers are SSMs: Generalized models and efficient algorithms through structured state space duality
• Various Lengths, Constant Speed: Efficient Language Modeling with Lightning Attention
• Gated linear attention transformers with hardware-efficient training

Linearization

• Liger: Linearizing Large Language Models to Gated Recurrent Structures
• Llamba: Scaling Distilled Recurrent Models for Efficient Language Processing
• LoLCATs: On Low-Rank Linearizing of Large Language Models
• The Mamba in the Llama: Distilling and Accelerating Hybrid Models

Efficient Reasoning with Subquadratic Attention

• M1: Towards Scalable Test-Time Compute with Mamba Reasoning Models
• Nemotron-H: A Family of Accurate and Efficient Hybrid Mamba-Transformer Models
• Compositional Reasoning with Transformers, RNNs, and Chain of Thought
• Cosmos-Reason1: From Physical Common Sense To Embodied Reasoning
• Thinking Slow, Fast: Scaling Inference Compute with Distilled Reasoners

🔖 Future Directions

Efficient Multimodal Reasoning and Video Reasoning

• Think or Not Think: A Study of Explicit Thinking in Rule-Based Visual Reinforcement Fine-Tunin
• Think or Not? Selective Reasoning via Reinforcement Learning for Vision-Language Models
• Fast-Slow Thinking for Large Vision-Language Model Reasoning
• Can Atomic Step Decomposition Enhance the Self-structured Reasoning of Multimodal Large Models?
• Skywork R1V: Pioneering Multimodal Reasoning with Chain-of-Thought

Efficient Test-time Scaling and Infinity Thinking

• Value-Guided Search for Efficient Chain-of-Thought Reasoning
• LIMOPro: Reasoning Refinement for Efficient and Effective Test-time Scaling
• Efficient Test-Time Scaling via Self-Calibration
• Dynamic self-consistency: Leveraging reasoning paths for efficient llm sampling

Efficient and Trustworthy Reasoning

• X-Boundary: Establishing Exact Safety Boundary to Shield LLMs from Multi-Turn Jailbreaks without Compromising Usability
• Deliberative alignment: Reasoning enables safer language models

Building Efficient Reasoning Applications (RAG, Tool, Agent)

• The Danger of Overthinking: Examining the Reasoning-Action Dilemma in Agentic Tasks
• Chain-of-Retrieval Augmented Generation
• Cut the Crap: An Economical Communication Pipeline for LLM-based Multi-Agent Systems

Evaluation and Benchmark

• THOUGHTTERMINATOR: Benchmarking, Calibrating, and Mitigating Overthinking in Reasoning Models
• S1-Bench: A Simple Benchmark for Evaluating System 1 Thinking Capability of Large Reasoning Models
• DNA Bench: When Silence is Smarter -- Benchmarking Over-Reasoning in Reasoning LLMs
• MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency
• Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMs

---

🎉 Contribution

Contributing to this paper list

⭐" Join us in improving this repository! If you know of any important works we've missed, please contribute. Your efforts are highly valued! "

Contributors

---