Let LRMs Break Free from Overthinking via Self-Braking Tuning

Dear Reviewer FmRF,

We thank you for your thorough review and constructive feedback. We have tried to address each of your concerns point by point in the rebuttal.

Q1: About the Design of the Overthink Score

A1：

1. About the relation to prior ξ$_O$ metric [1]: We acknowledge that [1] is a valuable work, which we also cite. However, our formulation introduces two key enhancements better suited for our Self-Braking Tuning framework:

• Purpose: Unlike [1], we leverage RER not just for post-hoc analysis but as a core quantitative signal for data construction(Section 2.3), enabling step-wise adaptive truncation based on per-example overthink scores.

• Granularity: Their ξ$_O$ operates at the token level, whereas ours works at the reasoning step level, respecting logical boundaries. Our ablations confirm this structural granularity better supports overthinking mitigation (Section 4.3).

2. About the Reliability of OMR

a. OMR as an intentionally sensitive signal: We do not consider OMR unreliable, but rather intentionally sensitive to capture subtle linguistic cues of overthinking such as "Wait" or "Alternatively." Prior works [2,3] support using such markers to identify reasoning drift.

b. Balancing OMR's contribution: To address potential noise, we assign OMR a low weight in the Overthink Score calculation, ensuring it informs without dominating the objective.

c. Ablation validation: Removing OMR entirely shows its clear contribution:

OMR enhances both accuracy (+1.58%) and efficiency (-398 tokens) by capturing linguistic signals that RER alone misses, forming a complementary structural-linguistic detection mechanism.

3. On reasoning step segmentation: Reasoning steps are segmented by double newlines in model outputs—a widely validated approach in the literature [4,5]. Each segment separated by \n\n represents one reasoning step.

[1] "Do not think that much for 2+ 3=? on the overthinking of o1-like llms."ICML 2025

[2] "Demystifying long chain-of-thought reasoning in LLMs." ICLR 2025

[3] "Retro-search: Exploring untaken paths for deeper and efficient reasoning." arXiv:2504.04383

[4] "Can Large Language Models Detect Errors in Long Chain-of-Thought Reasoning." ACL 2025

[5] "The lessons of developing process reward models in mathematical reasoning." ACL 2025

Q2:About the Overthink Score

A2:

(1) On the Design Motivation for Setting β = 0.1 in the Overthink Score

We set β = 0.1 to prioritize structural reliability (ηs) while preserving linguistic sensitivity (κt). The 90:10 weighting ensures stability across different prompting styles while capturing subtle overthinking signals. Our ablation study validates this choice, showing consistent performance across threshold values.

(2) On the Empirical Justification for β = 0.1 through Ablation Study

To support this design choice, we conducted an ablation study by varying β ∈ {0.05, 0.1, 0.15, 0.2} for both SBT-E and SBT-D truncation strategies. The results are shown below:

These results show β = 0.1 yields optimal accuracy-efficiency trade-offs across both methods. Higher β values cause accuracy degradation, confirming our parameter choice.

Q3: About the SBT data construction procedure

A3： We would like to clarify that SBT-Exact and SBT-Dynamic are not final inference strategies, but data construction methods designed to guide the model toward learning when to stop reasoning.

(1) About the design of SBT-Exact:

We address your concern through both existing results from Section 4.2 and additional experiments. The systematic comparison across different solution counts demonstrates:

This analysis reveals that retaining two complete solutions provides the optimal balance: sufficient reasoning exposure for robust learning while establishing clear termination signals through answer convergence.

(2) About the design of SBT-Dynamic: SBT-D uses τ₁=0.2 based on systematic evaluation across multiple values, identifying about 60% of samples as overthinking. The step-wise evaluation enables problem-specific adaptive termination—simple problems stop after 3-4 steps while complex ones continue for 15-20 steps, determined by the model's own reasoning patterns rather than fixed cutoffs.

(3) About spontaneous overthinking control:

These augmentations develop model self-awareness of overthinking. After SBT, generation becomes fully autonomous—models learn to shorten outputs and signal stopping without external control (Fig. 1&2). Analysis of Qwen2.5-Math-7B-SBT-E on AIME shows substantial outputs naturally terminate early via epiphany sentences, achieving both higher efficiency and accuracy.

These findings further support that the self-braking behavior is not enforced by hard constraints, but emerges naturally during inference—yielding better performance with less computation.

Q4:About Performance Under SBT

A4: We agree that model performance should remain the priority. Our results reflect an inherent accuracy-efficiency trade-off, consistent with prior efficiency-focused work [1,2].

1. Performance preservation strategies: We actively mitigate performance degradation through several strategies:

(1) dual supervision using SBT-Exact and SBT-Dynamic for both fixed and adaptive stopping patterns,

(2) explicit braking cues providing natural language signals for sufficient reasoning depth.

(3) balanced training mixing full and truncated trajectories to preserve reasoning depth.

Under these measures, we achieve well-controlled performance impact with an average accuracy drop of 1.43% across all models and maximum drop of 2.86%. Notably, for general models (LLaMA), SBT-D actually improves both accuracy and efficiency, demonstrating the method's potential in broader settings.

2. Key insight: Baseline models may achieve correctness through redundancy rather than focused reasoning. Our approach teaches models to reach correct answers more efficiently, with the slight accuracy trade-off reflecting elimination of this redundant computation rather than loss of reasoning capability.

[1] Ma, Xinyin, et al. "CoT-Valve: Length-Compressible Chain-of-Thought Tuning." ACL 2025

[2] Xia, Heming, et al. "TokenSkip: Token Pruning for Vision-Language Models." ACL 2025.

Q5: On the Purpose and Effectiveness of Natural Language Guidance (NLG)

A5:

1. Essential Role of Explicit Stopping Signals: Without NLG, our training data would consist merely of variable-length reasoning trajectories with abrupt truncations. This creates a fundamental learning gap: models would have no explicit supervision signal for recognizing when reasoning becomes sufficient, instead passively waiting for external </think> markers. Our ablation study confirms this necessity:

Removing guidance reduces accuracy (-0.27%) and increases token consumption (+7.1%), demonstrating that explicit metacognitive cues are essential for self-regulation learning.

2. Superior Efficiency Over Special Tokens: While Table 5 shows comparable overall performance, detailed results in Table 9 reveal NLG's advantages on challenging tasks:

• AIME: 3381 vs. 3647 tokens (-7.3%), 14.17% vs. 12.34% accuracy (+14.8%)
• Overall: 1682 vs. 1797 tokens (-6.4% average)

Performance differences are most pronounced on complex reasoning, where natural language's semantic richness provides clearer termination cues.

3. Key Advantages of Natural Language Signals:

• Semantic alignment with pretraining makes them easier to learn than special tokens
• Self-awareness enables internal reasoning state reflection
• Fluent integration allows natural generation during autonomous inference

Q6: On the necessity and contribution of Masked Redundant Thinking (MRT) A6: We address your concern in two parts:

1. Foundational Motivation and Supporting Work: MRT builds on established practices: [1] uses masked incorrect reasoning for step-level self-correction, adopted in [2, 3]. These works demonstrate that masking redundant reasoning improves abstraction, robustness, and self-regulation—core objectives of Self-Braking Tuning.

2. Empirical Evidence and Ablation Study: Section 4.2 shows shorter masked spans improve reasoning efficiency. An ablation removing MRT yields:

While removing MRT marginally increases accuracy (+0.19%), it significantly raises token consumption (+37.84%), confirming MRT's essential role in efficiency gains without compromising effectiveness.

[1] "S³cmath: Spontaneous step-level self-correction makes large language models better mathematical reasoners." AAAI 2025

[2] "Masked Thought: Simply Masking Partial Reasoning Steps Can Improve Mathematical Reasoning Learning of Language Models." ACL 2024

[3] "Token Cleaning: Fine-Grained Data Selection for LLM Supervised Fine-Tuning." ICML 2025

Q7: On Paper Formatting

A7: We thank the reviewer for pointing out the formatting issues. We will address both concerns in the revised manuscript to ensure full compliance with the conference's guidelines.

Dear Reviewer FmRF,

Thank you once again for your time and thoughtful feedback on our submission.

In our recent responses, We have addressed your concerns regarding the adaptability of SBT-Exact, the empirical evidence supporting SBT-Dynamic, and the interpretability of our ablation results.

As the discussion phase is drawing to a close, we would greatly appreciate it if you could take a moment to review our responses and let us know if you have any further questions or suggestions. Your insights have been extremely valuable in helping us improve the clarity and rigor of our work.

Thank you again for your constructive engagement throughout the review process.

Best regards,

The authors of Submission8232

Dear Reviewer FmRF,

Thank you very much for your continued engagement and for raising a fundamental and insightful question:

"But the more fundamental question for achieving adaptive thinking is whether more difficult problems also require more reasoning solutions."

We fully understand your concern, and would like to address it by analyzing the relationship between the number of reasoning solutions and reasoning capability from two complementary perspectives — during training data construction and during inference behavior.

---

1. On the Number of Solutions in Data Construction

In our initial rebuttal, we compared the effects of retaining 1, 2, and 3 solutions during SBT-E construction. Here, we further provide their respective performance on the most challenging dataset — AIME:

These results clearly demonstrate that increasing the number of solutions does not improve performance; in fact, retaining 3 solutions degrades both accuracy and efficiency compared to retaining 2. This suggests that increasing the number of reasoning solutions beyond two offers diminishing returns in terms of meaningful reasoning content, and often introduces redundancy. These results clearly indicate that a greater number of reasoning solutions does not translate into enhanced problem-solving ability.

2. On the Number of Solutions During Inference

As shown in Figure 3(b), we analyzed the average number of solutions generated per question across datasets of varying difficulty using DeepSeek-R1-Distill-Qwen-7B:

• GSM8K (easiest): 4.54 solutions, over 1,000 tokens
• MATH500 (intermediate): 4.82 solutions, over 2,000 tokens
• AIME (most difficult): 5.64 solutions, approximately 5,000 tokens

Despite the significant difference in problem difficulty, the difference in the number of generated solutions between the easiest and hardest datasets is, on average, only about one additional solution. However, this gap is reflected far more substantially in the content — with nearly a fivefold increase in tokens, largely concentrated within the Foundation Solution.

This suggests that more difficult problems are not solved by proportionally increasing the number of reasoning solutions, but rather by increasing the complexity and depth of each individual solution.

This directly supports our central design choice in SBT-E:

When we fix the number of reasoning solutions at two, what varies is the level of detail and complexity within each individual solution, depending on the difficulty of the problem. In other words, the model adapts by making each reasoning path richer and deeper, rather than by generating more separate reasoning paths. This behavior aligns with the reasoning patterns that we have observed in LRMs. This approach helps maintain stable training while providing the model with enough flexibility to determine when to stop reasoning, without producing excessive redundant or unnecessary solutions.

We hope this analysis helps clarify that the adaptivity in SBT-E and SBT-D does not lie in generating more reasoning solutions for harder problems, but rather in allowing each solution to naturally scale in complexity with problem difficulty — a form of adaptive reasoning that we believe directly addresses your core concern.

We sincerely hope this clarifies the core of your concern, and we are grateful for your thoughtful push to refine the framing of our work.

Best regards,

The authors of Submission8232

Dear Reviewer S6DA,

We thank you for your thoughtful review and valuable feedback. We appreciate your recognition of our framework’s motivation, the effectiveness of our results, and the contribution of a useful dataset. We have carefully addressed your comments in the rebuttal.

Q1: About the new terms

A1： We appreciate your feedback on the complexity of our terminology. Our intention was to precisely distinguish between different reasoning concepts and analysis dimensions, but we understand this may have impacted readability. In future versions, we will strive to streamline naming, ensure stylistic consistency, and provide clearer, more intuitive explanations upon first use to improve overall accessibility.

Q2: About the evaluation scope

A2: Thank you for the suggestion. We understand the importance of evaluating beyond a single domain and offer the following clarification and supporting results:

1. Focus on the math domain: Our main evaluation centers on math-related tasks because the SBT framework is trained entirely on math data. Overthinking behaviors are especially prominent and easier to analyze in this setting, given the well-defined reasoning structure and objective correctness of math problems. This makes math particularly suitable for developing and initially validating our approach.

2. Generalization beyond math: We agree that overthinking is not confined to mathematics. To test the broader applicability of our method, we additionally evaluate SBT on two out-of-domain benchmarks, MMLU-Redux and GPQA-Diamond, which cover a range of factual and commonsense reasoning challenges. The results below show that our methods (SBT-E and SBT-D) achieve consistent token savings while maintaining competitive accuracy across diverse model backbones—without any out-of-domain fine-tuning:

These results confirm that our proposed overthinking-aware strategies transfer well to non-mathematical reasoning tasks and generalize across model scales and architectures. Specifically, we observe consistent token reductions ranging from 26% to 65% across all model-task combinations, with accuracy drops typically limited to 1-3%. Notably, smaller models (LLaMA3.2-1B) show the largest efficiency gains (65% token reduction) and sometimes even improved accuracy, while larger models achieve more moderate but still substantial savings (26-40%). The fact that math-trained models maintain competitive performance on diverse reasoning tasks without domain-specific retraining demonstrates the fundamental nature of our self-regulation mechanism—suggesting that learning to recognize reasoning sufficiency is a transferable meta-cognitive skill applicable across reasoning domains.

Q3: About Framework Generalizability and Improvements

A3:

1. On Multiple Hyperparameters and Complexity

Our Self-Braking Tuning framework actually requires fewer hyperparameters than existing overthinking mitigation approaches. While we use two primary thresholds (τ₁ and τ₂), the secondary threshold is simply set as τ₁ + 5%, reducing this to essentially one core hyperparameter. In contrast:

• RL-based approaches require reward function design, learning rates, exploration parameters, and length penalty coefficients [1-3]
• External constraint methods need token budgets, stopping criteria, and confidence thresholds [4,5]
• Dynamic inference methods require multiple confidence thresholds, early-stopping parameters, and uncertainty estimation hyperparameters [6-8]

Our approach represents a significant simplification in the hyperparameter space while maintaining effectiveness.

2. On Clear Hyperparameter Selection Guidance

We agree that guiding parameter selection is essential; therefore, we dedicate ourselves to providing systematic empirical guidance for hyper-parameter choice:

(1) For threshold τ₁: Our analysis in Section 4.1 demonstrates that τ₁ = 0.2 consistently achieves optimal performance across all model architectures and task difficulties

(2) For weighting parameter β: We supplemented our experiments with additional validations that confirmed the choice of β = 0.1.

This systematic analysis shows β = 0.1 consistently achieves the best accuracy-efficiency trade-off for both SBT variants, providing clear selection guidance rather than arbitrary parameter choices.

3. On Robustness Demonstrated by Consistent Improvements in Table 2-5

(1) Purpose of Comprehensive Ablation Studies: These ablation studies serve a critical purpose: identifying near-optimal configurations that maintain our primary objective of dramatic efficiency gains (30-60% token reduction) while preserving accuracy. Our extensive experiments systematically explore the parameter space across:

• Multiple threshold values (0.2, 0.3, 0.4)
• Alternative configurations (step-level vs. token-level, natural language vs. special tokens)
• Various model architectures (Qwen2.5-Math, Llama-3.1, Llama-3.2)

This systematic exploration establishes optimal operating points for Self-Braking Tuning across diverse experimental conditions.

(2) Beyond Peak Performance: Prioritizing Universal Applicability:

However, our focus extends beyond simply chasing peak numerical performance. Instead, we aim to develop a generalizable and stable methodology that gracefully balances performance and efficiency across diverse conditions. The consistent patterns observed across all experiments—rather than dramatic variations—indicate that SBT achieves reliable performance without requiring extensive retuning for different models or tasks. This stability is precisely what enables practical deployment and broader applicability.

(3) Evidence of Framework Robustness: The reviewer's observation about "marginal improvements" in Tables 2-5 actually demonstrates our framework's key strength: robustness. The consistency of results across:

• Multiple model architectures: Qwen2.5-Math, Llama-3.1, Llama-3.2
• Varying task difficulties: GSM8K to AIME
• Different experimental configurations: All ablation combinations

This consistency confirms that our core mechanism—teaching models autonomous reasoning regulation—is fundamentally sound and transferable. This robustness and reliability across diverse conditions validates that SBT represents a mature, deployable solution rather than a method overfitted to specific experimental settings.

[1] Aggarwal, Pranjal, and Sean Welleck. "L1: Controlling how long a reasoning model thinks with reinforcement learning." arXiv preprint arXiv:2025.

[2] Luo, Haotian, et al. "O1-pruner: Length-harmonizing fine-tuning for o1-like reasoning pruning." arXiv preprint arXiv:2501.12570, 2025.

[3] Shen, Yi, et al. "Dast: Difficulty-adaptive slow-thinking for large reasoning models." arXiv preprint arXiv:2503.04472, 2025.

[4] Han, Tingxu, et al. "Token-budget-aware llm reasoning." arXiv preprint arXiv:2412.18547, 2024.

[5] Xu, Silei, et al. "Chain of draft: Thinking faster by writing less." arXiv preprint arXiv:2502.18600, 2025.

[6] Yang, Chenxu, et al. "Dynamic early exit in reasoning models." arXiv preprint arXiv:2504.15895, 2025.

[7] Zhang, Jintian, et al. "Lightthinker: Thinking step-by-step compression." arXiv preprint arXiv:2502.15589, 2025.

[8] Manvi, Rohin, Anikait Singh, and Stefano Ermon. "Adaptive inference-time compute: Llms can predict if they can do better, even mid-generation." arXiv preprint arXiv:2410.02725, 2024.

Question: Is this a typo? It seems that the model with the lowest token cost shown in Table 1 — Llama-3.1-8B-Instruct — is marked incorrectly.

Thank you for pointing this out. We must acknowledge that the lowest token cos for LLaMA-3.1-8B-Instruct was incorrectly marked due to an oversight on our part. We appreciate your careful reading and will correct this in the revised version.

Dear Reviewer Bv4y,

We thank you for your thorough review and constructive feedback. We sincerely appreciate your positive recognition of the novelty of our framework, the clarity of our analysis, and the overall quality of the writing. We have addressed each of your concerns point by point in the rebuttal.

---

W1: About the performance under SBT

A1: 1. On the Trade-off Between Accuracy and Efficiency: We would like to clarify that the trade-off between reasoning efficiency and accuracy is not a limitation of our method per se, but rather an inherent challenge in complex multi-step reasoning. Reducing redundant steps can, in some cases, limit the model’s ability to recover answers through repetition or self-correction. This phenomenon is particularly visible in math-specialized models (e.g., Qwen2.5-Math), which rely heavily on deep token-level exploration. Similar trade-offs between accuracy and efficiency have been observed in other studies as well, such as in [1] and [2], further validating this inherent challenge in multi-step reasoning tasks.

2. About the Accuracy: We recognize that maintaining strong performance is critical, and we adopt multiple strategies to mitigate potential drops. These include diverse supervision (SBT-Exact + SBT-Dynamic), explicit braking prompts, and balanced training between full and early-stopped traces. As a result, the overall performance drop is strictly controlled (≤ 2.86%), with an average gap of only 1.43% across models—while reducing token usage by up to 60%.

3. On General Models: Importantly, SBT is not merely efficient—it also enhances accuracy in some settings. On general-purpose models such as LLaMA,, the SBT-D variant yields consistent gains in both performance and efficiency. This highlights the method’s broader applicability beyond math-heavy tasks, and suggests that SBT can serve as a general-purpose optimization strategy for step-wise reasoning.

[1] "CoT-Valve: Length-Compressible Chain-of-Thought Tuning." ACL 2025.

[2] "TokenSkip: Token Pruning for Vision-Language Models." ACL 2025.

---

W2: Addressing Module-Level Ablation Studies

A2:

We'd like to explain our original experimental design and provide the requested module-level evaluation.

1. Original Experimental Motivation:

Our initial focus was on comparing different guidance mechanisms (natural language vs. special tokens) rather than evaluating the necessity of guidance signals themselves. This decision was based on preliminary observations that models without explicit termination cues showed poor self-regulation behavior, making the comparison between guidance types more immediately relevant than the presence vs. absence comparison.

2. Complete Module Ablation Results:

We have now conducted the comprehensive ablation study removing natural language guidance entirely:

Key Findings from Complete Ablation: The removal of natural language guidance results in both reduced accuracy (-0.27%) and increased token consumption (+7.1%), demonstrating that explicit metacognitive cues provide measurable benefits for both reasoning quality and efficiency.

3. Individual Module Contributions:

This complete ablation clarifies each component's role:

• Natural Language Guidance: +0.27% accuracy, -6.6% tokens (vs. no guidance)
• Natural Language vs. Special Tokens: +0.05% accuracy, -6.4% tokens (from Table 9)

The results confirm that explicit termination signals are essential for effective self-regulation, with natural language providing the most efficient implementation of this mechanism.

---

W3: On Hyperparameter Sensitivity and Practical Usability

A3:

We appreciate the reviewer's focus on practical deployment considerations. Our analysis reveals that the perceived hyperparameter complexity is actually a strength of interpretable design:

1. Interpretable and modular parameter design: Unlike black-box optimization methods, our SBT framework introduces hyperparameters that are explicitly interpretable and modular. Each parameter controls observable behaviors directly linked to model outputs—for example, the threshold corresponds to measurable overthinking patterns in reasoning trajectories. This transparency enables practitioners to adjust parameters intuitively and purposefully rather than relying on blind hyperparameter search, making the framework more accessible than alternative approaches requiring complex reward engineering.

2. Robust performance across parameter ranges: Our threshold ablation experiments (Table 2) demonstrate that while performance fluctuates between different values, variations are smooth and bounded. Even the worst-performing threshold achieves substantial efficiency gains—specifically, 41.5% token reduction (1917 vs. 3278 tokens) while maintaining competitive accuracy. This indicates our framework is not fragile: moderate deviations from the “optimal” threshold do not lead to catastrophic drops in performance, as long as the deviation is not too large.

3. Adaptive trade-offs by design, not deficiency: The accuracy-token usage trade-off represents a deliberate design feature rather than a limitation. SBT explicitly enables flexible early termination to avoid unnecessary overthinking. This adaptability makes our approach flexible rather than rigid, allowing practitioners to optimize for their specific requirements without framework modifications.

---

Q1: Evaluation on Long CoT-Capable Models

A1:

We appreciate the reviewer's suggestion to evaluate SBT on models with inherent long CoT capabilities. This was in fact an aspect we considered early in our experimentation process. We did conduct preliminary experiments on DeepSeek-R1-Distill-Qwen-7B to address this question, and the results provide important insights into our method's applicability:

1. Experimental Results on DeepSeek-R1-Distill-Qwen-7B:

While SBT achieves a 13.4% token reduction, the accuracy drop (3.21%) is more significant than observed with base models, suggesting limited effectiveness when applied to already long CoT-optimized models.

2. Data Scale Analysis: The modest performance gains can be attributed to a fundamental data scale imbalance. DeepSeek-R1-Distill models have been trained on approximately 800K high-quality long CoT reasoning trajectories, while our SBT-enhanced dataset contains only 92K examples. This significant disparity (8.7× difference) means that the pre-existing long CoT patterns learned from the larger dataset naturally dominate during fine-tuning.

The limited scale of our SBT dataset is insufficient to substantially alter the reasoning behaviors already established through extensive training on much larger long CoT corpora. This explains why the efficiency improvements are modest—the model's existing reasoning patterns, optimized through large-scale CoT training, are more resistant to modification by our smaller intervention dataset.

3. Method Positioning: Our results indicate that SBT is most effective for base models that have not undergone extensive long CoT training, where our 92K examples can meaningfully shape reasoning behaviors. For models already trained on large-scale CoT data, SBT would require proportionally larger datasets to achieve competitive performance.

This finding suggests an important design consideration: SBT is optimally positioned as a primary training enhancement rather than a secondary refinement technique. Future work could explore scaling our data construction methodology to generate larger SBT datasets comparable to existing long CoT training corpora.

---

Q2: On Illustrative Examples and Summary Statistics

A2: We appreciate the reviewer's constructive suggestion for enhancing the clarity of our data construction framework. We address both recommendations below:

1. Illustrative Examples We have already provided illustrative examples of our framework's impact in Figures 1 and 2. The consistency between our training datasets and the trained models' inference behavior generally reflects the effectiveness of our data construction strategy. We will include more detailed illustrative examples in the appendix of the camera-ready version.

2. Summary Statistics of Dataset Construction Impact While we briefly mentioned the proportion of overthinking cases in Section 4.1, we provide more comprehensive statistics here to better illustrate our framework's data construction impact:

Total dataset size: 92,064 examples

This comprehensive data transformation directly translates to the substantial efficiency gains (30-60% token reduction) demonstrated in our experimental results, confirming the effectiveness of our systematic approach to overthinking identification and mitigation.

Dear Reviewer 6Pbn,

We sincerely appreciate the reviewer's insightful and constructive feedback, which highlights the strengths of our work and provides valuable guidance for future improvements. We have tried to address each of your concerns point by point in the rebuttal.

Q1: About the reliance on specific reasoning delimiters in SBT

A1: We respectfully address the concern about generalizability limitations. The reviewer's concern appears to center on the <think>... </think> delimiter dependency, but this represents a misunderstanding of both our approach's flexibility and the current landscape of reasoning models.

1. Current reasoning model landscape: The vast majority of state-of-the-art reasoning models—including OpenAI o1, DeepSeek-R1, QwQ, Gemini 2.0 Flash Thinking, and Kimi-1.5—already employ explicit reasoning delimiters during inference. This is not a limitation of our approach but rather an alignment with the dominant architectural paradigm in modern LRMs.

2. Framework adaptability: Our SBT framework is highly flexible regarding delimiter formats. Adapting to different thinking markers is straightforward and easily implementable—we can simply modify the delimiter patterns in our parsing logic. The core methodology—identifying overthinking patterns through efficiency metrics and linguistic markers—remains unchanged regardless of the specific delimiter format used. For models without explicit delimiters, our framework can be adapted to identify reasoning segments through other structural cues or linguistic patterns.

Rather than representing a limitation, our focus on reasoning-capable models addresses the most practically relevant segment of current LRM deployment, where overthinking poses the greatest computational challenge. The widespread adoption of structured reasoning in leading models suggests that delimiter-based approaches like ours are becoming increasingly relevant, not obsolete.

Q2: On Threshold Adaptability and Generalizability

A2: We agree that the choice of predefined thresholds can influence the effectiveness of the SBT framework. However, we would like to clarify that our method demonstrates robustness across a reasonably wide range of threshold values, and address the underlying concern about manual tuning requirements.

1. Empirical Evidence of Threshold Robustness

As shown in Section 4.1, although we selected the best-performing threshold (0.2) to maximize performance, other threshold settings still achieved strong results. Specifically:

• Threshold 0.3: 44.0% average token reduction with 2.3% accuracy drop
• Threshold 0.4: 43.0% average token reduction with 2.0% accuracy drop
• Even under suboptimal thresholds, we maintain an average token reduction of 43.5%

This indicates that SBT does not require precise tuning to be effective and retains strong performance under relaxed configurations. The performance degradation is gradual rather than cliff-like, suggesting inherent stability in our approach.

2. Cross-Model Generalization Without Retuning

More importantly, our experiments demonstrate that the same threshold (0.2) works effectively across diverse model architectures without requiring model-specific adjustments:

• Mathematical specialists (Qwen2.5-Math-1.5B/7B): 48.9% and 30.7% token reduction respectively
• General-purpose models (Llama-3.2-1B and Llama-3.1-8B): 54.2% and 62.8% token reduction respectively

This cross-model consistency suggests that overthinking patterns exhibit structural similarities that transcend specific architectures, making our thresholds naturally transferable.

3. Task-Agnostic Performance

Our evaluation across four distinct mathematical reasoning benchmarks (GSM8K, MATH500, AIME, AMC23) with varying difficulty levels shows consistent improvements using identical thresholds. The framework adapts organically to task complexity through the Overthink Score mechanism itself, rather than requiring manual threshold adjustments for each benchmark.

4. Theoretical Foundation for Threshold Selection

The threshold range (0.2-0.4) corresponds to identifying 40-60% of reasoning instances as containing overthinking, which aligns with our structural analysis in Section 2.1 showing that Evolution Solutions (where overthinking primarily occurs) constitute a significant portion of reasoning trajectories. This provides a principled basis for threshold selection rather than purely empirical tuning.

5. Comparison with Existing Approaches

We note that most existing overthinking mitigation methods require significantly more complex parameter tuning:

• RL-based approaches require reward function design, learning rates, and exploration parameters, and so on.[1-3]

• External constraint methods need token budgets and stopping criteria calibration, and so on.[4,5]

• Dynamic inference methods require confidence thresholds and early-stopping parameters, and so on.[6-8]

In contrast, our approach requires tuning of essentially one primary hyperparameter (τ₁), with the secondary threshold (τ₂) set as a simple offset (τ₁ + 5%).

Our focus is not on chasing peak metrics through intensive hyperparameter search, but on developing a generalizable and stable framework that gracefully balances accuracy and efficiency across different models and tasks with minimal manual intervention.

[1] Aggarwal, Pranjal, and Sean Welleck. "L1: Controlling how long a reasoning model thinks with reinforcement learning." COLM 2025.

[2] Luo, Haotian, et al. "O1-pruner: Length-harmonizing fine-tuning for o1-like reasoning pruning." arXiv preprint arXiv:2501.12570, 2025.

[3] Shen, Yi, et al. "Dast: Difficulty-adaptive slow-thinking for large reasoning models." arXiv preprint arXiv:2503.04472, 2025.

[4] Han, Tingxu, et al. "Token-budget-aware llm reasoning." ACL 2025.

[5] Xu, Silei, et al. "Chain of draft: Thinking faster by writing less." arXiv preprint arXiv:2502.18600, 2025.

[6] Yang, Chenxu, et al. "Dynamic early exit in reasoning models." arXiv preprint arXiv:2504.15895, 2025.

[7] Zhang, Jintian, et al. "Lightthinker: Thinking step-by-step compression." arXiv preprint arXiv:2502.15589, 2025.

[8] Manvi, Rohin, Anikait Singh, and Stefano Ermon. "Adaptive inference-time compute: Llms can predict if they can do better, even mid-generation." arXiv preprint arXiv:2410.02725, 2024.

Q3: About the performance

A3: We take the issue of premature termination seriously and have carefully designed our framework to mitigate it through the following three strategies:

1. Adaptive Reasoning Preservation: Both SBT-E and SBT-D are designed to preserve the Foundation Solution and at least part of the Evolution Solution, ensuring the model retains essential reasoning and self-correction capabilities. In particular, SBT-D employs a dynamic, step-wise overthinking score to determine problem-specific stopping points. This allows more complex problems to retain longer reasoning chains when necessary, reducing the risk of premature termination. Additionally, for trajectories not identified as overthinking, we preserve the full reasoning process to enhance the model’s capacity when handling genuinely complex tasks.

2. Exposure via Masked Redundant Reasoning: Instead of removing overthinking completely, we retain a small portion of redundant reasoning as masked content. This allows the model to observe overthinking patterns without reinforcing them, helping it to learn the distinction between necessary elaboration and redundancy.

3. Empirical Safeguards: Extensive results across benchmarks (e.g., MATH500, AIME) show that Self-Braking Tuning maintains, and occasionally even improves, accuracy while significantly reducing token usage. Notably, SBT-D on Llama-3.1-8B improves accuracy by +2.62% on MATH500 while cutting tokens by 59%, suggesting that the model benefits from eliminating unnecessary or distracting reasoning without compromising essential steps.

Q4: About the Overthink Marker

A4: Thank you for your comment. We would like to address your concerns from the following aspects:

1. Ease of Identifying Overthink Markers While specific overthink markers may vary between models, identifying these markers is straightforward. By manually inspecting or using a language model to analyze just a few dozen reasoning examples, one can easily collect a set of frequently occurring indicator tokens. This simple process allows adapting the marker set to different models without much effort.

2. Common Linguistic Features in Reasoning Models Additionally, recent studies[1-5] have highlighted that tokens like "wait," "perhaps," and "alternatively" frequently appear in reasoning models and are often indicative of overthinking. This has already become a widespread linguistic feature of reasoning models.

In summary, while specific markers may vary, our methodology for detecting and mitigating overthinking is adaptable and can be generalized across different reasoning models.

[1]"Wait, We Don't Need to 'Wait'! Removing Thinking Tokens Improves Reasoning Efficiency.", arXiv:2506.08343

[2]"Missing Premise Exacerbates Overthinking: Are Reasoning Models Overthinking?", arXiv:2504.06514

[3]"Demystifying Long Chain-of-Thought Reasoning in LLMs.", ICML 2025.

[4]"Retro-Search: Exploring Untaken Paths for Deeper and Efficient Reasoning.", arXiv:2504.04383

[5]"LLMs Can Easily Learn to Reason from Demonstrations: Structure, Not Content, Is What Matters!", arXiv:2502.07374