Large Reasoning Models (LRMs) have revolutionized complex problem-solving, yet they exhibit a pervasive "overthinking", generating unnecessarily long reasoning chains. While current solutions improve token efficiency, they often sacrifice fine-grained control or risk disrupting the logical integrity of the reasoning process. To address this, we introduce Stepwise Adaptive Thinking (SAT), a framework that performs step-level, difficulty-aware pruning while preserving the core reasoning structure. SAT formulates reasoning as a Finite-State Machine (FSM) with distinct thinking modes (Slow, Normal, Fast, Skip). It navigates these states dynamically using a lightweight Process Reward Model (PRM), compressing easy steps while preserving depth for hard ones. Experiments across 9 LRMs and 7 benchmarks show that SAT achieves up to 40% reduction in reasoning tokens while generally maintaining or improving accuracy.
Figures & tables
Figure 1: Baselines vs. SAT. COT spends redundant checks for easy steps, Early-Stopping halts after a high-confidence first-pass answer and fails, while SAT skips redundancy on easy steps but preserves verification on hard steps, achieving the correct answer efficiently.
Figure 2: Overview of the proposed framework that models LRM reasoning as a Finite-State Machine (FSM). The left panel provides a detailed view of the specific state transition dynamics and stability rules of the proposed reasoning FSM. The right panel illustrates the closed-loop control flow: the Pilot perceives the difficulty ( rt ) of the previous step ( yt−1 ), driving the State Transition Logic to update the thinking state ( st ). This state is then mapped to a control tag ( ct ) via In-Context Steering to modulate the next generation step ( yt ).
Figure 3: Lightweight Pilot: A GRU-based framework fusing semantic embeddings and uncertainty features to capture stepwise difficulty trajectories.
Model
Method
GSM8K
MATH500
AIME 2024
AIME 2025
AMC
Acc ( ↑ , %)
Tokens ( ↓ )
Acc ( ↑ , %)
Tokens ( ↓ )
Acc ( ↑ , %)
Tokens ( ↓ )
Acc ( ↑ , %)
Tokens ( ↓ )
Acc ( ↑ , %)
Tokens ( ↓ )
COT
95.1
2136
95.0
4823
60.0
12662
50.0
12720
92.5
7408
DEER
94.5 ( -0.6 )
1250 ( -41% )
92.6 ( -2.4 )
3214 ( -33% )
63.3 ( +3.3 )
9327 ( -26% )
55.0 ( +5.0 )
12084 ( -5% )
87.5 ( -5.0 )
4906 ( -34% )
CGRS
–
–
91.3 ( -3.7 )
2704 ( -44% )
56.7 ( -3.3 )
7893 ( -37% )
–
–
86.7 ( -5.8 )
4351 ( -41% )
Qwen3-4B
SAT
95.1 ( +0.0 )
845 ( -60% )
95.6 ( +0.6 )
2833 ( -41% )
60.0 ( +0.0 )
9462 ( -25% )
53.3 ( +3.3 )
9907 ( -22% )
92.5 ( +0.0 )
5255 ( -29% )
COT
95.8
2152
95.6
5166
66.7
12393
60.0
12835
90.0
7920
Table 1: Performance on mathematical reasoning benchmarks. Metrics include Acc ( ↑ ) and Tokens ( ↓ ). Changes relative to the COT are highlighted in orange for Acc and blue for Tokens. Best results within each group are bolded.
Figure 4: Accuracy and token usage on GPQA and HumanEval across different model scales.
Figure 5: Ablation study results. Left: training and test loss curves for GRU-based Pilot under different inputs. Right: test-set correlation (Pearson/Spearman) between predicted step scores and supervision targets.
Figure 6: Strategy ablation results on MATH500 (Qwen3-14B). Accuracy and token usage of COT , w/o FAST , w/o SLOW , and SAT.
Figure 7: Attribution analysis of token savings on MATH500 (Qwen3-14B). Left: decomposition of SAT’s total token savings into reflection-related and branching-related components. Right: Pearson correlation between per-sample total token savings and each component.
Figure 8: Length-limit failures on aggregated math benchmarks. Number of incorrect predictions caused by exceeding maximum length, aggregated over GSM8K, MATH500, AIME 2024, AIME 2025, and AMC.
Figure 9: Performance analysis on MATH500 (Qwen3-14B). Left: Absolute accuracy comparison. Right: Normalized computational costs (Tokens/s, Time, VRAM) relative to the COT baseline ( 1.0× ).
Figure 10: Hyperparameter robustness on MATH500 (Qwen3-14B). Sensitivity of SAT to τfast and τslow .
Figure 11: State allocation across dataset difficulty and model scales. We report the step-level thinking mode ratios on GSM8K (easy), MATH500 (medium), and AIME 2025 (hard), for Qwen3-14B and Qwen3-4B.
Figure 12: Outcome-conditioned token savings on AIME2025 (Qwen3-8B). SAT saves substantially more tokens on correct instances than on failed ones, where many runs hit the max-token limit.
Category
Feature Name
Description & Definition
Semantic
Step Embedding
The 384-dimensional dense vector encoded by gte-small representing the semantic content of the current reasoning step text.
Uncertainty (Static)
canonical_logprobs
The log-probability of the sampled token: logP(xt∣x<t) .
canonical_selected_rank
The rank of the sampled token in the vocabulary distribution (1-based).
canonical_entropy
The Shannon entropy of the local probability distribution (truncated to Top- K ).
canonical_logit_gap
The difference between the largest logit (Top-1) and the second largest logit (Top-2).
canonical_margin
The probability margin between the Top-1 and Top-2 tokens: P(xtop1)−P(xtop2) .
Table 2: List of Input Features for the Lightweight Pilot. The features are categorized into Semantic inputs and Uncertainty inputs (subdivided into Static and Dynamic metrics).
Control Tag
Instruction / Semantics
[Fast_Step]
Indicates the current step appears easy. The model is instructed to keep reasoning brief and avoid unnecessary details.
[Slow_Step]
Indicates the current step appears difficult. The model is instructed to perform detailed, expansive reasoning.
[Normal_Step]
Indicates moderate difficulty. The model is instructed to resume standard step-by-step reasoning. (Corresponds to the Normal state).
[Skip_Step]
Indicates the step is excessively difficult and further elaboration is unlikely to be helpful. The model is instructed to summarize existing reasoning, make a reasonable guess, and quickly output the final answer.
Table 3: Definitions of Control Tags. These tags are inserted into the context to modulate the reasoning style of the subsequent step.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 13: Baselines vs. SAT. Early-Stopping halts after a high-confidence first-pass answer and fails; Suppressing Reflective eliminates all verification steps, preventing necessary self-correction and leading to the same error. In contrast, SAT skips redundancy on easy steps but preserves verification on hard steps, achieving the correct answer efficiently.
Model
Method
GSM8K
MATH500
AIME 2024
AIME 2025
AMC
Acc ( ↑ , %)
Tokens ( ↓ )
Acc ( ↑ , %)
Tokens ( ↓ )
Acc ( ↑ , %)
Tokens ( ↓ )
Acc ( ↑ , %)
Tokens ( ↓ )
Acc ( ↑ , %)
Tokens ( ↓ )
COT
95.1
2136
95.0
4823
60.0
12662
50.0
12720
92.5
7408
TALE
–
–
89.1 ( -5.9 )
2657 ( -45% )
48.9 ( -11.1 )
9727 ( -23% )
–
–
86.7 ( -5.8 )
5107 ( -31% )
NoThinking
–
–
84.9 ( -10.1 )
988 ( -80% )
24.4 ( -35.6 )
4504 ( -64% )
–
–
70.0 ( -22.5 )
1710 ( -77% )
Dynasor
–
–
90.1 ( -4.9 )
3877 ( -20% )
54.3 ( -5.7 )
9912 ( -22% )
–
–
86.7 ( -5.8 )
6233 ( -16% )
DEER
94.5 ( -0.6 )
1250 ( -41% )
92.6 ( -2.4 )
3214 ( -33% )
63.3 ( +3.3 )
9327 ( -26% )
55.0 ( +5.0 )
12039 ( -5% )
87.5 ( -5.0 )
4906 ( -34% )
Appendix
Table 4: Comparisons with additional baselines. Metrics include Acc ( ↑ ) and Tokens ( ↓ ). Changes relative to the COT are highlighted in orange for Acc and blue for Tokens. Best results within each group are bolded.
Figure 14: Process behavior analysis on MATH500. Average reflect_steps and branch_steps (lower is better) for COT , w/o FAST , w/o SLOW , and SAT, computed from the thought traces using the forward window scheme ( W=1 ).
Figure 15: Window-size sensitivity for token-savings attribution. We report (i) contribution ratios of reflection-/branch-related savings and (ii) Pearson correlations between per-sample total savings and each component, under window sizes W∈{0,1,2} .
Figure 16: Generation Throughput Comparison. The average number of tokens generated per second by COT and SAT using Qwen3-32B. The consistent throughput confirms that SAT introduces negligible per-step latency overhead.
Large Reasoning Models (LRMs) have shown remarkable reasoning capabilities, yet they still suffer from overthinking, generating redundant reasoning steps which incur substantial token consumption. Existing methods, such as suppressing reflective keywords or forcing shorter reasoning lengths, attempt to mitigate this issue but inevitably truncate necessary steps and induce underthinking, thereby compromising performance. To address this dilemma, we investigate the latent representations and observe that efficient reasoning steps naturally cluster into a concentrated region in latent space, while those deviating from this region tend to produce verbose sequences. To leverage this, we keep reasoning focused within this region via a quadratic program which projects deviating hidden states back into the region. Then we propose a novel training-free framework to achieve efficient reasoning that reduces token generation costs without sacrificing performance. Extensive experiments conducted on four models ranging from 1.5B to 14B, and across six benchmarks in math reasoning, coding, and scientific QA, validate the effectiveness of our method, up to a 12.1% improvement in accuracy while reducing generated tokens by 11.8% to 52.8%. Codes are available at https://github.com/hzn18/Opt4Reasoning.
Large Reasoning Models (LRMs) have achieved remarkable success on complex tasks by leveraging long chain-of-thought (CoT) trajectories, yet they frequently exhibit overthinking on simple queries, resulting in significant token overhead and reduced inference efficiency. However, existing compression methods predominantly apply uniform length reduction or rely on coarse-grained difficulty estimation, often leading to performance degradation on difficult problems. To address this limitation, we propose Confidence-Adaptive Thinking (CAT), a framework that incorporates the model's intrinsic self-certainty signals as confidence into the preference optimization process, which autonomously modulates reasoning lengths based on problem difficulty. Experimental results show that CAT consistently outperforms state-of-the-art baselines on reasoning accuracy across multiple benchmarks on different base models. Our work enables LRMs to effectively compress confident responses while deliberating on uncertain ones, offering a potentially robust solution for balancing accuracy and latency in practical industrial scenarios.
Qizhi Jiang, Shuo Wang, Pei Ke +2
Laboratory of Intelligent Collaborative Computing, University of Electronic Science and Technology of China, Chengdu, China · Ubiquitous Intelligence and Trusted Services Key Laboratory of Sichuan Province
Large Reasoning Models (LRMs) often suffer from overthinking due to redundant verification steps. Existing approaches for mitigating overthinking, such as fast-slow thinking switching and reasoning trajectory compression, fail to make a fine-grained distinction between beneficial and redundant steps within the LRM's reasoning process, and may thus impair reasoning capability in their pursuit of efficiency. To simultaneously improve reasoning efficiency and capability, we propose EvoThink, a framework that reduces redundant verification and encourages the exploration of new reasoning paths. EvoThink comprises two key components: Self-Pruning Training (SPT), an unsupervised method that iteratively prunes redundant reasoning steps and self-trains on the concise trajectories; and Aha-Moment Preference Optimization (AMPO), which, inspired by genetic algorithms, identifies valuable failed reasoning attempts, synthesizes from-wrong-to-right aha-moment data, and optimizes the model to internalize this reasoning pattern. Extensive evaluations across mathematical reasoning and code generation benchmarks demonstrate that EvoThink not only substantially reduces inference-time token usage but also improves the reasoning capability of LRMs.