Source-linked AI summary
A Survey of Efficient Reasoning for Large Reasoning Models: Language, Multimodality, and Beyond
Xiaoye Qu, Yafu Li, Zhao-Chen Su, Weigao Sun, Jianhao Yan, Dongrui Liu, Ganqu Cui, Daizong Liu, Shuxian Liang, Junxian He, Peng Li, Wei Wei, Jing Shao, Chaochao Lu, Yue Zhang, Xian-Sheng Hua, Bowen Zhou, Yu Cheng
TL;DR
LRMs gain capability through explicit, often lengthy reasoning, but their traces can waste tokens through redundancy, over-analysis, and shallow path exploration. This survey synthesizes efficient-reasoning methods and challenges across the LRM lifecycle, concluding that efficiency requires balancing solution quality against computational cost while addressing unresolved evaluation and scalability issues.
Problem
LRMs often generate excessive, redundant, or meandering reasoning, creating a need to improve reasoning efficiency rather than merely accelerate general inference.
Method
The survey reviews efficient-reasoning approaches across inference, supervised fine-tuning, reinforcement learning, and pretraining, while organizing inefficiency patterns and challenges.
Results
The survey defines reasoning efficiency, characterizes recurring inefficiency patterns, and synthesizes methods and future directions across the LRM lifecycle.
Takeaways & Limitations
Reasoning efficiency is framed as a trade-off between solution quality and computational cost, including generated tokens and inference latency.
Takeaways & Limitations
Evaluating the utility of individual reasoning steps remains difficult, and simultaneously sampled ultra-long traces impose substantial computational overhead.
Abstract
from arXiv · showhide
Recent Large Reasoning Models (LRMs), such as DeepSeek-R1 and OpenAI o1, have demonstrated strong performance gains by scaling up the length of Chain-of-Thought (CoT) reasoning during inference. However, a growing concern lies in their tendency to produce excessively long reasoning traces, which are often filled with redundant content (e.g., repeated definitions), over-analysis of simple problems, and superficial exploration of multiple reasoning paths for harder tasks. This inefficiency introduces significant challenges for training, inference, and real-world deployment (e.g., in agent-based systems), where token economy is critical. In this survey, we provide a comprehensive overview of recent efforts aimed at improving reasoning efficiency in LRMs, with a particular focus on the unique challenges that arise in this new paradigm. We identify common patterns of inefficiency, examine methods proposed across the LRM lifecycle, i.e., from pretraining to inference, and discuss promising future directions for research. To support ongoing development, we also maintain a real-time GitHub repository tracking recent progress in the field. We hope this survey serves as a foundation for further exploration and inspires innovation in this rapidly evolving area.
1 Introduction
Large Reasoning Models improve difficult-task reasoning by generating explicit chain-of-thought, but their deliberate behavior can produce excessive, costly reasoning even for simple problems. This survey focuses specifically on reasoning efficiency, reviews methods across the LRM lifecycle, and identifies inefficiency patterns and open challenges.
- Motivation: 1248 tokens versus 30 tokens: QwQ-32B uses far more output than Qwen2.5-32B-Instruct on an elementary-school math problem.Figure 1 also compares generation-length distributions on mixed GSM8K, MATH-500, and AIME 2024 problems.
- Motivation: LRMs such as DeepSeek R1, OpenAI-o1/o3, and QwQ generate intermediate chain-of-thought steps to support deliberate reasoning on complex tasks.They differ from rapid, heuristic-driven LLMs by exhibiting more analytical System 2-like reasoning.
- Motivation: Reasoning inefficiency includes redundant content, over-analysis of simple problems, and superficial exploration of multiple reasoning paths.These patterns can make reasoning slower and more verbose despite the effectiveness of long chains on difficult problems.
- Position and Contribution: Existing surveys cover LRM development, long chain-of-thought, and test-time scaling, but none specifically addresses reasoning efficiency in LRMs.The survey frames efficiency as an emerging issue for deployment, scalability, and practical application.
- Position and Contribution: The survey provides a targeted review of efficient reasoning across pretraining, supervised fine-tuning, reinforcement learning, and inference.It also characterizes common inefficiency patterns and outlines challenges and future directions across the end-to-end development pipeline.
2 Reasoning Efficiency: Definition, Patterns, and Challenges
The survey defines reasoning efficiency as solution quality relative to computational cost across a task distribution, then characterizes redundancy, overthinking, and incoherent reasoning as persistent LRM inefficiencies. It identifies challenges involving evaluation, length control, cross-task adaptation, and architectures for long reasoning traces.
- 2.1 Definition of Reasoning Efficiency: Reasoning efficiency η(M) evaluates an LRM across tasks by relating solution quality Q(M, D) to computational cost C(M, D).Quality may include accuracy, exact match, or creativity; cost may include FLOPs, generated tokens, or inference latency.
- 2.2 Patterns of Reasoning Inefficiency: LRMs produce redundant reasoning, including repetitive question rephrasing and verbose explanations that increase computational cost.One example uses 138 tokens merely to rephrase and interpret the question.
- 2.2 Patterns of Reasoning Inefficiency: LRMs often overthink simple questions by generating redundant reasoning rounds and unnecessary verification even after reaching the correct answer.Table 1 illustrates repetitive phrasing and multiple reasoning rounds on a straightforward math problem.
- 2.2 Patterns of Reasoning Inefficiency: Underthinking causes premature direction changes, shallow reasoning, and fragmented solution paths, reducing quality or increasing computational cost.The phenomenon is especially reported for complex mathematical tasks.
- 2.2 Patterns of Reasoning Inefficiency: Excessively long CoT sequences destabilize RL optimization, increase memory use, and raise autoregressive inference latency, especially beyond 10,000 tokens.These costs are particularly consequential in multi-agent systems requiring timely plans and responses.
- 2.3 Unique Challenges for Efficient Reasoning in the Era of LRMs: Evaluating each reasoning step’s utility remains difficult, making precise pruning and the brevity–correctness trade-off unresolved.The survey also identifies semantic length control, cross-task adaptation, and architectures beyond quadratic-attention Transformers as open challenges.
3 Efficient Reasoning during Inference
Inference-time efficiency methods allocate or reduce reasoning computation through length budgeting, switching between systems or models, speculative decoding, routing, and parallel search. These approaches seek better efficiency–accuracy trade-offs, but forced shortening can degrade accuracy.
- 3 Efficient Reasoning during Inference: Inference methods explicitly budget computation by controlling sequential reasoning length or the number of parallel candidate generations.Parallel search provides test-time scaling by generating multiple candidates concurrently.
- 3.1 Length Budgeting: Length budgeting uses prompts, adaptive concise reasoning, end-of-thinking delimiters, or zero-/short-thought decoding to limit intermediate tokens.TALE estimates a token budget, while S1 directly forces termination with an end-of-thinking delimiter.
- 3.1 Length Budgeting: Forced reasoning-length constraints may reduce accuracy, motivating studies of the trade-off between response length and performance.A systematic study reports a universal trade-off curve across compression instructions.
- 3.1 Length Budgeting: Difficulty-adaptive methods allocate more inference resources to hard or uncertain queries and penalize premature switching between thoughts.DSC uses query difficulty, Dynasor uses model certainty, and TSP discourages superficial thought transitions.
- 3.2 System Switch and Model Switch: System switching alternates fast intuitive and slow deliberate reasoning, while model switching routes queries across models to balance performance and cost.Dyna-Think dynamically selects slow reasoning using generation consistency and thought complexity; many model-switch techniques remain insufficiently applied to LRMs.
- 3.3 Speculative Decoding: Speculative decoding uses draft models or early exits to propose tokens before full-model verification, balancing speed and quality through fallback or rollback.BiLD exemplifies this pattern with a small fast model followed by larger-model correction.
- 3.4 Model Routing: Routing modules select models for prompts according to difficulty, using learned rewards, query embeddings, or lightweight quality-and-cost predictors.Zooter, RouteLLM, and MixLLM implement distinct reward-guided or meta-decision routing strategies.
- 3.5 Parallel Search: Parallel search improves efficiency by evaluating partial responses, terminating weak candidates early, pruning search trees, or recursively revising samples.SBoN aims for comparable performance with substantially lower computational resource demands.
4 Efficient Reasoning with SFT
Efficient SFT methods teach models to shorten explicit reasoning chains or replace token-by-token CoT with latent representations. The surveyed approaches use compression, skipping, controllable reasoning lengths, and continuous hidden states to reduce token and computational overhead.
- SFT methods mainly compress reasoning chains or train models with latent-space tokens instead of explicit sequences.
- Reasoning Chain Compression: Concise-chain methods construct short targets through token budgets, best-of-N sampling, few-shot conditioning, compression, or supervised step-skipping.
- Reasoning Chain Compression: CoT-Valve learns a task vector controlling reasoning length, enabling extrapolation to chains longer or shorter than those seen during training.
- Latent-Space SFT: Latent-space methods progressively replace explicit CoT with continuous hidden states, including feedback of hidden representations for subsequent reasoning steps.
- Latent-Space SFT: CODI aligns teacher explicit-CoT and student implicit-CoT hidden activations, while LightThinker and Heima compress intermediate reasoning to reduce memory, computation, and token usage.
- Summary and Outlook: Future SFT research may combine explicit and implicit reasoning, support multimodal latent reasoning, adapt curricula, and improve alignment between CoT and latent representations.
5 Efficient Reasoning with Reinforcement Learning
Reinforcement-learning methods improve reasoning efficiency by controlling or implicitly shaping computation along reasoning trajectories. The survey organizes them around length-aware rewards and alternative mechanisms that balance correctness with token consumption.
- Efficient RL research seeks to reduce or control token consumption while maintaining task performance, primarily through reward design.
- Length-Aware Rewards: Length-aware methods treat reasoning length as a trajectory-level resource and formulate efficiency as constrained optimization under soft or target budgets.
- Length-Aware Rewards: Extremely long CoT approaching context limits can reduce accuracy, while length hacking may extend reasoning through repetition rather than problem-solving.
- Length-Aware Rewards: Step-level, redundancy-aware, difficulty-adaptive, and discrete-mode rewards refine length control beyond raw token budgets.
- Implicit Efficiency Signals: Implicit approaches let efficient behavior emerge through meta-reinforcement learning, preference optimization, or training biases rather than direct length rewards.
- Summary and Outlook: Across methods, the central objective is balancing reasoning depth against token efficiency while preserving correctness.
6 Efficient Reasoning during Pretraining
Pretraining-based approaches target efficiency through latent representations, subquadratic sequence models, and linearization of pretrained transformers. These methods seek lower computational cost while preserving reasoning performance, but latent shortcuts may weaken complex stepwise reasoning.
- The survey covers latent-space pretraining, linear models with subquadratic attention, and transformer linearization.
- Pretraining in Latent Space: Latent-space pretraining replaces traditional token-based processing with continuous representations intended to improve reasoning efficiency and model understanding.
- Pretraining in Latent Space: BLT groups raw bytes into entropy-based patches, while LCMs predict sentences in pretrained embedding spaces for modality- and language-agnostic processing.
- Pretraining in Latent Space: CoCoMix interleaves sparse-autoencoder-derived concept vectors with token prediction to improve sample efficiency and higher-order abstraction.
- Summary and Outlook: Implicit reasoning shortcuts may impede true stepwise reasoning on complex, multi-step tasks, motivating work on consistency and accuracy.
- Subquadratic and Linear Models: Linear attention, sparse attention, state-space models, and linear RNNs reduce sequence-processing costs through subquadratic or recurrent computation.
- Transformer Linearization: Linearization methods convert pretrained transformers into linear recurrent structures, including gated recurrent models, linear approximations, and distilled SSMs.
7 Future Directions
Future work should make reasoning adaptive across modalities, inference strategies, trustworthiness requirements, and deployment settings. The survey highlights matching reasoning effort to task difficulty while addressing computational overhead, reliability, and broader application needs.
- Efficient Multimodal Reasoning and Video Reasoning: Efficient multimodal reasoning remains under-explored because image and video inputs contain more noise and can trigger excessive computation.
- Efficient Multimodal Reasoning and Video Reasoning: SCoT decomposes tasks into atomic, semantically meaningful steps to align reasoning processes with problem types and complexity.
- Efficient Multimodal Reasoning and Video Reasoning: Dynamic multimodal strategies should use simpler reasoning for easier inputs and more sophisticated reasoning for challenging ones.
- Efficient Test-time Scaling: Shifting test-time scaling from search depth toward width can reduce latency, but simultaneous ultra-long traces create substantial processing overhead.
- Trustworthy Reasoning: Long CoT raises safety and reliability concerns because harmful or noisy information may accumulate during extended reasoning.
- Applications: RAG and agent systems can adjust sampled retrieval chains or reasoning effort, but extensive iterative reasoning remains costly in time-critical settings.
- Applications: Tool-use efficiency may benefit from hierarchical reasoning, early exits, and parallel execution, while coding, driving, healthcare, and embodied AI remain open applications.
8 Conclusion
The survey reviews efficient reasoning in Large Reasoning Models, defining reasoning efficiency, characterizing inefficiency, and examining methods across the LRM lifecycle. It also identifies future directions and aims to provide foundational knowledge for this emerging field.
- The survey defines reasoning efficiency and presents patterns of reasoning inefficiency in Large Reasoning Models.
- It highlights unique challenges associated with achieving efficient reasoning in the era of Large Reasoning Models.
- The review examines efficient-reasoning methods across inference, supervised fine-tuning, reinforcement learning, and pre-training.
- The survey proposes future directions that may benefit from efficient reasoning.
- It aims to serve as a comprehensive entry point that equips readers with foundational knowledge of this challenging field.