Introduction
The landscape of Large Language Models (LLMs) is undergoing a fundamental shift from "System 1" thinking—fast, intuitive, next-token prediction—to "System 2" thinking, characterized by deliberate, multi-step reasoning. Models such as OpenAI’s o1 and DeepSeek-R1, known as Large Reasoning Models (LRMs), utilize extended Chain-of-Thought (CoT) processing to solve complex logical and mathematical problems. However, this increased capability comes at a significant computational cost. This survey explores the emerging field of efficient reasoning, categorizing the methods designed to maximize "intelligence per token" across the model lifecycle.
The Efficiency Crisis in Large Reasoning Models
While LRMs represent a leap in problem-solving ability, they are notoriously inefficient. The authors identify a "token economy" crisis where models often consume thousands of tokens to solve tasks that their non-reasoning counterparts handle in a fraction of the time. This inefficiency manifests in three primary patterns:
- Redundant Content: Models frequently repeat definitions or rephrase the initial query without advancing the logical state.
- Overthinking Simple Tasks: A reasoning model might generate hundreds of tokens to answer a trivial question like "What is the square of 3?"
- Suboptimal Paths: Models often "meander," falling into circular logic loops or switching between reasoning directions prematurely.
As shown in the figure above, the QwQ-32B model can consume over 500 tokens for a simple arithmetic query that a standard instruct model resolves in 30. Statistically, the distribution of generation lengths for LRMs shows a massive shift toward higher token counts, illustrating the overhead currently required for "deliberation."
A Taxonomy of Efficiency Methods
To address these bottlenecks, researchers are developing strategies across four key stages of the model lifecycle:
1. Inference-Time Strategies
Inference-time methods aim to control the reasoning process dynamically. Length Budgeting (e.g., "Chain-of-Draft") forces models to use concise notations rather than full sentences. System Switching involves a controller that decides whether a query requires a "fast" System 1 response or a "slow" System 2 deliberation. Additionally, Parallel Search allows for generating multiple reasoning paths concurrently, pruning low-quality branches early to save total compute.
2. Supervised Fine-Tuning (SFT)
SFT focuses on training models on higher-quality, more concise data. Reasoning Chain Compression involves distilling verbose traces into dense, logical steps. A more radical approach is Latent-Space SFT, such as the Coconut framework, which moves reasoning from visible text tokens into hidden continuous representations. This allows the model to "think" internally without the latency of generating text tokens.
3. Reinforcement Learning (RL)
Standard RL often inadvertently rewards "length hacking," where longer chains are more likely to reach the correct answer by chance. To counter this, researchers are implementing Length-Aware Rewards that penalize excessive verbosity while rewarding logical correctness and brevity.
4. Pre-training and Architecture
At the architectural level, the goal is to overcome the quadratic cost of standard Transformers. This includes using Subquadratic Attention (like FlashAttention variants) or Linearization techniques that convert Transformers into more efficient structures like Mamba or State Space Models (SSMs).
Key Insights and Benchmarks
Efficiency is not merely about shortening responses; it is about reaching the "aha moment"—the core logical breakthrough—with minimal diversion. The survey highlights that "thinking in latent space" is one of the most promising frontiers, potentially offering orders-of-magnitude speedups for complex tasks.
To evaluate these methods, the authors point to specialized benchmarks like DNA Bench, which is specifically designed to expose the tendency of LRMs to produce excessively verbose reasoning chains. Traditional benchmarks like GSM8K and MATH-500 remain standard for measuring accuracy, but they are increasingly paired with efficiency metrics to assess the cost of that accuracy.
Future Directions and Limitations
Despite progress, significant challenges remain. Multimodal Inefficiency is a major hurdle; current LRMs often struggle with visual reasoning, leading to even more convoluted and circular paths than text-only tasks. Furthermore, there is no standard metric for the "utility" of an individual reasoning step, making it difficult to determine which tokens are truly necessary.
As we move toward "System 1.x" models—which intelligently allocate a thinking budget based on query difficulty—the focus will shift from simply scaling tokens to scaling the density of intelligence within those tokens.
Summary updated