acceptodds
Under review as a conference paper at ICLR 2027

DeCoR: Decoupled Collaborative Reasoning via Preference-Guided Pruning and Dual-Indicator Gating

Abstract

Large Language Models (LLMs) with long Chain-of-Thought (CoT) training demonstrate exceptional reasoning capabilities, yet reliance on excessive reasoning tokens degrades both performance and efficiency. Existing methods address this through specialized retraining or external constraints. While effective, these methods still retain a single-stream reasoning paradigm, leaving exploration and termination entangled within the same trajectory. To overcome this limitation, we propose DeCoR (coupled llaborative easoning), a novel framework for efficient reasoning that decouples exploration and termination through twin-model interaction. We first devise a preference-guided method to achieve structured layer pruning, which derives a functional auxiliary model that serves as an explorer under high predictive uncertainty. This design disentangles distinct reasoning roles while preserving a shared state and trajectory coherence. To coordinate these roles according to the evolving reasoning state, we further introduce a dual-indicator gating mechanism that monitors uncertainty and consistency of intermediate answers. It dynamically alternates between the base model for correction or termination and the auxiliary model for continuous exploration. Extensive experiments on five reasoning benchmarks demonstrate that our method significantly reduces CoT length by 32.5% on average and improves final accuracy by 2.8 to 6.5 percentage points. Our method also increases the density of effective reasoning information within concise outputs. The code will be made publicly available.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLMs, reasoning, overthinking, structured pruning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_6BDrOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DeCoR, an efficient reasoning framework that separates exploration and termination using a base model and a pruned auxiliary model. It combines preference-guided layer pruning with uncertainty- and consistency-based gating to determine when to continue reasoning, correct the trajectory, or terminate. Experiments on three 8B reasoning models and five benchmarks show reduced output length and improved accuracy.

Critical strengths
  • The shared-parameter and KV-cache implementation introduces almost no additional memory overhead. The analysis of cache reuse, auxiliary-model performance, and inference latency is also useful from a systems perspective.
Critical weaknesses
  • Insufficient justification/validation of the core assumptions. DeCoR assumes that high answer entropy indicates insufficient reasoning and that low-entropy, consistent answers indicate reliable convergence. However, the paper does not analyze correct-but-high-entropy predictions on challenging problems or incorrect-but-low-entropy predictions arising from model overconfidence. It also does not demonstrate that the pruned auxiliary model performs more semantically diverse or effective exploration rather than simply continuing the same reasoning trajectory.

  • Method complexity relative to the reported gains. DeCoR requires per-model offline calibration, a model-specific pruning mask, multiple gating mechanisms, and several hyperparameters. Given this added implementation complexity, the reported accuracy gains are relatively modest, and their reliability is unclear because no standard deviations or confidence intervals are provided. Moreover, DeCoR increases total FLOPs despite reducing output tokens and latency.

Questions for authors
  • DeCoR uses a single fixed pruning mask per base model across all tasks. Is the selected pruning mask stable across different calibration subsets or random seeds, and do task-specific masks result in materially different performance?
3 comments

We sincerely thank the reviewer for the professional comments, as well as the positive recognition of our work. We hope you might reconsider the rating in light of our clarifications. Detailed responses are provided point by point as follows:


Q1: Insufficient justification and empirical validation of core assumptions. DeCoR assumes that high answer entropy indicates insufficient reasoning and that low-entropy, consistent answers indicate reliable convergence. However, the paper analyzes neither correct yet high-entropy predictions on challenging problems nor incorrect yet low-entropy predictions arising from model overconfidence. It also does not demonstrate that the pruned auxiliary model performs more semantically diverse or effective exploration, rather than simply continuing the same reasoning trajectory.

A1: Many thanks for this insightful comment. To further investigate potential factors that may compromise final reasoning quality and accuracy, we quantify the proportion of challenging gating cases encountered by DeCoR across all datasets. We analyze five distinct cases: (1) Low-U Wrong: the proportion of incorrect answers among low-entropy intermediate answers. (2) Wrong Acceptance: the proportion of completed storage windows with an incorrect latest answer that still satisfy the consistency criterion and trigger termination. (3) Wrong-to-Correct: the proportion of trajectories containing low-entropy incorrect answers that ultimately achieve correct final answers. (4) High-U Correct: the proportion of correct answers among high-entropy intermediate answers. (5) Final Correct: the proportion of trajectories containing high-entropy correct answers that ultimately terminate with correct answers.

Base ModelLow-U Wrong (%)Wrong Acceptance (%)Wrong-to-Correct (%)High-U Correct (%)Final Correct (%)
R1-Distill-Llama-8B18.425.345.649.373.4
Qwen3-8B16.722.048.241.677.2
Ministral-3-8B16.220.447.352.869.5

The results show that predictive uncertainty is informative but imperfect, as both low-uncertainty errors and high-uncertainty correct answers occur. Importantly, the consistency gate filters out most incorrect candidates within completed low-uncertainty windows, while continued reasoning allows many trajectories with incorrect intermediate answers to reach the correct final answer. These findings support the value of combining uncertainty and consistency for more reliable reasoning control.

Our existing results already support the auxiliary model's semantic diversity and efficacy during reasoning. Figure 4 shows its higher-uncertainty exploration patterns, while Figure 7 demonstrates the gains of preference-guided pruning over ShortGPT and the unpruned variant. Moreover, as discussed for reviewer ieBf (Q2, Q4), our analysis of uncertainty and intermediate answer dynamics further supports DeCoR's collaborative reasoning. Nevertheless, we conduct an additional experiment to examine the auxiliary model's semantic dynamics. Specifically, from shared prefixes in DeCoR trajectories, the pruned auxiliary model and full base model independently generate segments under identical decoding settings and token budgets without gating. The same base model then generates intermediate answers before and after each continuation, isolating trajectory differences from evaluation. We define three metrics: (1) exploration gain (EG), the average increase in correct intermediate answer proportion before and after continuation, (2) answer discovery rate (ADR, %), calculated as the proportion of prefixes with no correct answers before exploration that yield at least one correct answer afterward, and (3) semantic exploration novelty (SEN, %), the manually audited proportion of continuations introducing novel and useful reasoning directions relative to their prefixes. We sample 120 problems across MATH500, AIME26, GPQA-D, and BBEH (30 each), extracting the first 15 segments of each DeCoR trajectory as shared prefixes. For each prefix, both models independently generate four 512-token continuations, followed by four intermediate answer generations with the base model. Results are shown below.

Base ModelExplorerEG ↑ADR (%) ↑SEN (%) ↑
R1-Distill-Llama-8BBase3.519.624.2
Auxiliary3.727.230.0
Qwen3-8BBase4.118.730.4
Auxiliary5.226.438.1
Ministral-3-8BBase3.925.027.9
Auxiliary3.630.435.2

It is clear that the auxiliary model achieves EG values comparable to those of the base model, while consistently improving ADR and SEN. This indicates that the auxiliary model not only maintains effective reasoning progress but also more frequently discovers correct answers and introduces novel reasoning directions, rather than merely extending existing reasoning trajectories.

We have incorporated these results into Appendix H and Appendix J.

Q2: Method complexity relative to reported gains. DeCoR requires per-model offline calibration, a model-specific pruning mask, multiple gating mechanisms, and several hyperparameters. Given this added implementation complexity, the reported accuracy gains are relatively modest, and their reliability is unclear because no standard deviations or confidence intervals are provided. Moreover, DeCoR increases total FLOPs despite reducing output tokens and latency.

A2: Many thanks for this comment. First, DeCoR requires only a one-time calibration for each base model on 400 samples, with an average cost of approximately 5 A100 GPU-hours. The resulting pruning mask is reused across all five benchmarks without additional training or task-specific recalibration. Moreover, the auxiliary pathway is implemented through in-place layer masking with shared parameters and a shared KV cache, introducing at most 0.02 GB of additional peak memory (Appendix L). Regarding gating complexity, DeCoR employs two complementary indicators based on simple uncertainty and consistency checks, with a unified hyperparameter configuration across all three base models and five benchmarks. Appendix D further validates parameter robustness. Thus, while DeCoR introduces additional calibration and inference control, it avoids repeated task-specific optimization and the overhead of maintaining two independently loaded models. Please refer to the response to reviewer 15U3's Q2 for further analysis of pruning generalizability.

To address the concern about statistical reliability, we additionally report 95% confidence intervals for accuracy across eight independent runs. As shown below, DeCoR achieves higher mean accuracy than Vanilla across all three base models and five benchmarks, with non-overlapping confidence intervals in 10 of the 15 comparisons. The remaining overlaps occur on AIME26 across all three base models, as well as on MATH500 and GPQA-D for Ministral-3-8B, indicating greater statistical uncertainty in these comparisons. These results provide further evidence for the reliability of accuracy improvements achieved by DeCoR across different models and benchmarks. DeCoR should be evaluated in terms of the overall trade-off between accuracy and efficiency, as it consistently improves accuracy while substantially reducing reasoning tokens and inference latency across diverse models and benchmarks (see Table 2).

Base ModelMethodGSM8KMATH500AIME26GPQA-DBBEH
R1-Distill-Llama-8BVanilla81.5 [81.2, 81.8]75.8 [74.4, 77.2]42.1 [38.7, 45.5]44.5 [43.4, 45.6]15.2 [15.0, 15.4]
DeCoR85.0 [83.9, 86.1]82.3 [79.6, 85.0]46.1 [44.3, 47.9]48.3 [46.4, 50.2]19.1 [18.3, 19.9]
Qwen3-8BVanilla84.1 [82.9, 85.3]79.5 [79.0, 80.0]45.2 [43.3, 47.1]46.1 [43.7, 48.5]16.6 [15.7, 17.5]
DeCoR88.2 [85.9, 90.5]82.3 [81.0, 83.6]49.1 [45.9, 52.3]49.7 [48.6, 50.8]20.8 [19.7, 21.9]
Ministral-3-8BVanilla91.2 [90.5, 91.9]84.1 [82.2, 86.0]41.2 [38.4, 44.0]50.2 [48.6, 51.8]18.9 [18.4, 19.4]
DeCoR94.0 [92.4, 95.6]87.2 [85.2, 89.2]45.6 [43.3, 47.9]53.0 [50.9, 55.1]24.3 [22.9, 25.7]

We acknowledge that DeCoR incurs additional FLOPs, while total FLOPs and wall-clock latency measure different aspects of computational efficiency. Conceptually, DeCoR resembles test-time scaling: rather than extending the reasoning chain, it allocates additional computation to intermediate answer assessment and selective correction, enabling more effective reasoning within a reduced token budget. By leveraging the pruned auxiliary model for exploration and reducing redundant reasoning through dual indicator gating, DeCoR consistently achieves lower output token counts and inference latency despite the additional arithmetic operations. Our latency measurements include intermediate answer generation, corrective interventions, and context prefilling overhead. Furthermore, Appendix K (Table 18) shows that shared KV caching achieves an 88.5%--91.6% reuse rate and reduces FLOPs by 76.2%--78.3% over the unshared variant. Thus, DeCoR improves practical reasoning efficiency and accuracy while accepting a moderate increase in total FLOPs.

We have incorporated confidence interval statistics into Appendix C.4.


Q3: DeCoR uses a single fixed pruning mask per base model across all tasks. Is the selected pruning mask stable across different calibration subsets or random seeds, and do task-specific masks result in materially different performance?

A3: Thank you for this valuable question. We conduct an additional calibration robustness analysis, detailed in our response to Reviewer 15U3's Q2. The results demonstrate that pruning masks are relatively stable across calibration subsets, while task-specific and out-of-domain calibration yield comparable performance, supporting the robustness and cross-task generalizability of DeCoR.

Dear Reviewer 6BDr,

We hope this message finds you well. As the discussion period is nearing its end, we wanted to ensure that we have addressed all of your concerns satisfactorily. If there is any additional feedback or points you would like us to consider, please let us know. Your insights are invaluable to us, and we are eager to resolve any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper.

Reviewer_15U3OpenReview ↗
Rating
3
Confidence
4(+1)
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes DeCoR, a framework for efficient reasoning in large language models that decouples exploration and termination in LLM reasoning via twin-model interaction: an auxiliary model is derived from the base model through preference-guided structured pruning (layer importance scored by KL divergence of next-token distributions, some of layers masked out), and during inference the auxiliary model generates reasoning segments while the base model produces intermediate answers, with a dual-indicator gating mechanism (entropy-based uncertainty and answer-consistency over a storage window) deciding when to continue, correct, or terminate. The main experiments evaluate three 8B reasoning models on five benchmarks, reporting accuracy, token count, and latency against multiple baselines, and the paper further includes analyses of reasoning dynamics and ablation studies, with additional supplementary experiments.

Critical strengths
  1. The efficient reasoning problem that the paper focuses on is an important topic in large language model reasoning.
  2. The design that separates "exploration" and "termination" is very interesting, and the twin-model framework is highly innovative.
  3. The paper reports a large and comprehensive set of experimental results and analyses, which provide strong support for the authors' claims.
  4. In addition to the effectiveness of the scheme on the efficient reasoning problem itself, the authors also provide explanations and analyses on engineering aspects such as KV cache and peak memory, which serve as a strong practical reference.
Critical weaknesses
  1. There are many hyperparameters, and they are coupled with one another. Although the appendix provides a sensitivity analysis, the search space is large, which raises the barrier to the practical application of the method.
  2. The data types in the calibration set are fixed. If the calibration distribution differs substantially from the target task distribution, is the scheme still reliable?
  3. The consistency gate uses "the answer remains stable within the window" as evidence of convergence. Could this assumption fail on open-ended generation tasks or tasks where the answer format is not unique?
Questions for authors

Please see Weaknesses. I have only minor concerns about this paper and look forward to the authors' response.

5 comments

We sincerely thank the reviewer for the positive and insightful feedback, and we are encouraged by the recognition of the importance of the efficient reasoning problem, the architectural innovation of decoupling exploration from termination within our twin-model framework, the comprehensiveness of our experimental results, and the practical value of our engineering analyses regarding KV cache and peak memory overhead. Your overall evaluation provides clear direction for our work, motivating us to continue refining and advancing this line of research. We carefully address your concerns point by point as follows:


Q1: There are many hyperparameters, and they are coupled with one another. Although the appendix provides a sensitivity analysis, the search space is large, which raises the barrier to the practical application of the method.

A1: Many thanks for this comment. The hyperparameters serve two distinct stages: preference-guided pruning (α,ρ\alpha,\rho) is performed during one-time offline calibration, while uncertainty and consistency parameters regulate online model collaboration. Despite a vast theoretical search space, a unified hyperparameter configuration performs consistently across all three base models and five benchmarks without task-specific tuning. Appendix D already provides extensive sensitivity analyses, including joint evaluation of the coupled window size ω\omega and consistency threshold ξ\xi.

Nevertheless, we conduct an additional joint sensitivity analysis on Qwen3-8B across MATH500 and GPQA-D, varying the uncertainty threshold τ∈{0.03,0.04,0.05}\tau\in\{0.03,0.04,0.05\} and the consistency configurations (ω,ξ)∈(3,0.5),(5,0.7),(6,0.8)(\omega,\xi)\in{(3,0.5),(5,0.7),(6,0.8)}, while keeping all other parameters fixed. This controlled experiment evaluates whether DeCoR maintains a favorable tradeoff between accuracy and efficiency under coupled parameter variations, without requiring an exhaustive hyperparameter search.

τ\tau(ω\omega, ξ\xi)MATH500 Acc ↑MATH500 TC ↓MATH500 Lat ↓GPQA-D Acc ↑GPQA-D TC ↓GPQA-D Lat ↓
0.03(3, 0.5)80.91323.5324.3846.32176.2041.37
0.03(5, 0.7)81.71481.0726.9149.22564.9253.71
0.03(6, 0.8)79.81710.2631.8648.02813.7157.23
0.04(3, 0.5)81.51184.6221.4048.81925.4336.42
0.04(5, 0.7)82.31342.7524.3249.72213.6541.77
0.04(6, 0.8)80.41572.3429.1848.52562.7252.27
0.05(3, 0.5)80.11105.5719.9448.11816.6334.22
0.05(5, 0.7)81.31364.0924.8347.72304.5542.65
0.05(6, 0.8)79.51493.4327.7747.32453.8950.14

It is clear that the default configuration achieves the highest accuracy on both benchmarks, while neighboring settings maintain competitive performance. In particular, varying τ\tau from 0.03 to 0.05 with (ω,ξ)=(5,0.7)(\omega,\xi)=(5,0.7) changes accuracy by at most 1.0 and 2.0 percentage points on MATH500 and GPQA-D, respectively. More permissive consistency criteria generally reduce reasoning costs at the expense of accuracy, whereas stricter criteria increase token consumption and latency without consistently improving accuracy. These findings demonstrate a clear trade-off between accuracy and efficiency under coupled parameter variations. Together with the unified configuration adopted across three base models and five benchmarks, these results show that DeCoR does not require extensive hyperparameter tuning for each specific task to achieve competitive performance.

Q2: The data types in the calibration set are fixed. If the calibration distribution differs substantially from the target task distribution, is the scheme still reliable?

A2: Thank you for this valuable question. In our original experiments, each base model employs a single pruning mask constructed from 400 calibration samples drawn equally from AQuA-RAT [1] and CommonsenseQA [2]. The same mask is reused across all five benchmarks without task-specific recalibration, demonstrating that DeCoR remains effective across diverse mathematical and knowledge-based reasoning tasks despite differences between calibration and evaluation distributions. Here, we conduct an additional experiment to further investigate the pruning sensitivity to calibration data. We sample two disjoint calibration subsets of 400 samples each from AQuA-RAT and CommonsenseQA, both distinct from the subset utilized in the main text. In addition, we collect two domain-aligned calibration subsets of 400 samples each from the training splits of GSM8K [3] and SciQ [4], along with 400 samples from the CaseHOLD [5] training split, a legal reasoning benchmark, to evaluate pruning robustness under substantial calibration distribution shifts. Using these five subsets, we recompute the pruning masks for Qwen3-8B and evaluate performance on MATH500 and GPQA-D, with the results summarized in the table below.

Calibration DatasetMask Overlap ↑MATH500 Acc ↑MATH500 TC ↓MATH500 Lat ↓GPQA-D Acc ↑GPQA-D TC ↓GPQA-D Lat ↓
General-1 (Original)100%82.31342.7524.3249.72213.6541.77
General-287.5%81.91357.2024.5448.82243.4942.16
General-3100%82.31342.7524.3249.72213.6541.77
Math-Specific87.5%82.61419.7823.9449.22292.8342.85
Science-Specific75%81.51390.2125.1950.02280.7541.35
Legal75%80.91395.9926.0748.52275.6741.84

The results demonstrate that preference-guided pruning exhibits strong robustness to calibration data variations. First, masks constructed from different general purpose calibration subsets achieve an 87.5%–100% overlap with the original mask while maintaining comparable performance. Second, domain-aligned calibration produces moderately different pruning masks but offers only marginal accuracy improvements on the corresponding benchmarks. Third, out-of-domain calibration using the legal reasoning dataset CaseHOLD yields 75% mask overlap, with accuracy decreases of only 1.4 and 1.2 percentage points on MATH500 and GPQA-D, respectively. These findings suggest that DeCoR does not critically depend on a particular calibration subset or task-specific pruning mask to achieve competitive performance, supporting the practicality of using a single mask across different reasoning tasks.

We have incorporated these analyses into Appendix I.


Q3: The consistency gate uses the stability of the answer within the window as evidence of convergence. Could this assumption fail on open-ended generation tasks or tasks where the answer format is not unique?

A3: This is indeed an important issue. We agree that exact answer consistency may be insufficient for open-ended tasks where semantically equivalent answers can have different forms. In the current implementation, DeCoR primarily targets reasoning tasks with objectively verifiable final answers, using a constrained \boxed{} format to facilitate intermediate answer comparison. Furthermore, consistency is not used in isolation for termination but is combined with the predictive uncertainty indicator, although their combination does not guarantee correctness. For open-ended generation, a natural extension is to replace exact matching with semantic equivalence checks or task-specific consistency verification. These adaptations require further validation and remain a direction for future work.

References

[1] Ling et al., "Program Induction by Rationale Generation: Learning to Solve and Explain Algebraic Word Problems", ACL (2017).

[2] Talmor et al., "CommonsenseQA: A Question Answering Challenge Targeting Commonsense Knowledge", NAACL (2019).

[3] Cobbe et al., "Training Verifiers to Solve Math Word Problems", arXiv:2110.14168 (2021).

[4] Welbl et al., "Crowdsourcing Multiple Choice Science Questions", Proceedings of the 3rd Workshop on Noisy User-generated Text (2017).

[5] Zheng et al., "When Does Pretraining Help? Assessing Self-Supervised Learning for Law and the CaseHOLD Dataset", Proceedings of the 18th International Conference on Artificial Intelligence and Law (2021).

Dear Reviewer 15U3,

We hope this message finds you well. As the discussion period is nearing its end, we wanted to ensure that we have addressed all of your concerns satisfactorily. If there is any additional feedback or points you would like us to consider, please let us know. Your insights are invaluable to us, and we are eager to resolve any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper.

Reviewer_ieBfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DeCoR, a framework for efficient reasoning that alternates between a pruned auxiliary explorer and the full base model. The auxiliary model is constructed using layer-wise changes in predictive distributions. The base model periodically generates intermediate answers, whose uncertainty and consistency determine whether to continue, correct, or terminate reasoning. Experiments on three 8B models and five benchmarks report improved accuracy with fewer generated tokens and lower latency.

Critical strengths
  1. The framework offers a practical way to combine inexpensive exploration with selective full-model intervention without additional parameter training. Sharing model weights makes this design more attractive than maintaining two independently loaded models.

  2. The evaluation covers three model families and five benchmarks, with ablations for pruning, gating, and cache sharing. The explicit inclusion of intermediate answers and corrective tokens in token accounting, and of intervention overhead in latency measurement, strengthens the empirical evaluation.

Critical weaknesses
  1. Appendix B specifies 30 AIME26 problems, and Appendix C.1 defines accuracy using eight responses per problem. Under this protocol, accuracy must equal (100m/240) for an integer number of correct responses (m). However, the DeCoR accuracies of 46.1%, 49.1%, and 45.6% in Table 2 cannot be obtained even after rounding to one decimal place. This may reflect an undocumented aggregation procedure.

  2. Appendix D measures baseline termination uncertainty using the last reasoning token, but measures DeCoR termination uncertainty using the average entropy of an intermediate answer already selected by the low-uncertainty gate. Consequently, the observed entropy separation is partly built into the measurement and controller. Although the pruning and gating ablations support the usefulness of the components, these analyses do not establish the proposed mechanistic explanation.

Questions for authors

1.Does the exploration–termination entropy separation remain when all methods are measured on the same type of output, with the same entropy definition and sampling temperature, without conditioning the comparison on passing the low-uncertainty gate?

  1. How are incorrect trajectories and temporarily correct answers handled in the reasoning deliberation ratio? What evidence distinguishes a higher ratio caused by useful verification from one caused by redundant reasoning after the answer is already available?
4 comments

We sincerely thank the reviewer for the constructive feedback and for recognizing the value of our work. We kindly invite you to reconsider the rating in light of our clarifications. Detailed responses are provided point by point as follows:


Q1: Appendix B specifies 30 AIME26 problems, and Appendix C.1 evaluates 8 responses per problem. Thus, accuracy must equal 100m/240100m/240 for an integer mm. However, Table 2 reports DeCoR accuracies of 46.1%, 49.1%, and 45.6%, which cannot be obtained even with rounding. This may reflect an undocumented aggregation procedure.

A1: Many thanks for identifying this critical point. As explicitly stated in Section 4.1, "We report accuracy (Acc, %), output token count (TC) and reasoning latency (Lat, s), averaged across 8 independent runs." Under this protocol, 8 reasoning trajectories are evaluated per problem across 8 independent runs. Therefore, each AIME26 problem is evaluated using 64 trajectories, yielding 30×8×8=1,92030 \times 8 \times 8 = 1{,}920 trajectories in total. The reported accuracy is averaged across all 8 runs, corresponding to a step resolution of 100/1,920≈0.0521%100 / 1{,}920 \approx 0.0521\%. Consequently, the reported accuracies of 46.1%, 49.1%, and 45.6% are mathematically consistent with our setup (e.g., 885/1,920≈46.09%→46.1%885 / 1{,}920 \approx 46.09\% \to 46.1\%). We apologize for the lack of clarity on this point in the original manuscript. We have clarified this evaluation protocol in Appendix C.1 to eliminate any ambiguity.


Q2: Appendix D measures baseline termination uncertainty via the last reasoning token, but measures DeCoR using the average entropy of an intermediate answer already selected by the low-uncertainty gate. Thus, the observed entropy separation is partly built into the measurement and controller. While pruning and gating ablations demonstrate component utility, these analyses do not establish the proposed mechanistic explanation.

A2: Many thanks for this insightful comment. As detailed in Appendix E, measurement protocols differ across reasoning behaviors: baseline exploration and termination are measured at the last reasoning token immediately preceding each newline delimiter or the end of reasoning, whereas correction and termination in DeCoR are evaluated by averaging token level entropy over intermediate answers generated by the base model. We acknowledge that termination entropy in DeCoR is conditioned on the low uncertainty and consistency criteria. However, a strictly matched comparison of intrinsic exploration and termination states would require additional operational assumptions, given their different outputs, contexts, and decision mechanisms.

Importantly, the decoupling in DeCoR is functional rather than conditioned on an explicit separation of intrinsic exploration and termination states: preference-guided pruning establishes complementary reasoning pathways, while dual indicator gating coordinates exploration, correction, and termination behaviors through collaborative twin-model interaction. Accordingly, Figure 4 is intended to characterize the uncertainty patterns associated with different reasoning behaviors under the proposed controller, rather than establish an independent comparison of intrinsic uncertainty separation on matched outputs.

Response to Q2 (Continued)

To validate whether intermediate answer uncertainty is associated with answer correctness independently of the gating mechanism, we conduct an additional controlled experiment across MATH500, AIME26, and GPQA-D. Specifically, we generate intermediate answers along reasoning trajectories using the original base model at four predefined token positions: 512, 1024, 1536, and 2048. All intermediate generations follow the same prompt template, decoding configurations, and entropy formulations. This entire process is performed without modifying the original reasoning trajectories or filtering answers based on the gating mechanism in DeCoR. Subsequently, we visualize the uncertainty distributions of correct and incorrect answers. In addition, we employ two metrics: the mean uncertainty difference between incorrect and correct answers (ΔHap\Delta H_{\text{ap}}) and uncertainty discrimination score (UDS, %), which quantifies the ability of predictive uncertainty to distinguish correct from incorrect answers.

ΔHap=1N−∑j=1N−Hj−−1N+∑i=1N+Hi+\Delta H_{\text{ap}} = \frac{1}{N_-}\sum_{j=1}^{N_-}H_j^- - \frac{1}{N_+}\sum_{i=1}^{N_+}H_i^+

UDS=100N+N−∑i=1N+∑j=1N−[I(Hj−>Hi+)+12I(Hj−=Hi+)]\mathrm{UDS} = \frac{100}{N_+N_-}\sum_{i=1}^{N_+}\sum_{j=1}^{N_-}\left[\mathbb{I}(H_j^->H_i^+) + \frac{1}{2}\mathbb{I}(H_j^-=H_i^+)\right]

Here, Hi+H_i^+ and Hj−H_j^- denote the predictive entropy of the ii-th correct and jj-th incorrect intermediate answers; N+N_+ and N−N_- denote their counts; and I(⋅)\mathbb{I}(\cdot) is the indicator function.

The respective metric values for the three base models are reported below.

MethodΔHap\Delta H_{\text{ap}}↑UDS (%) ↑
Vanilla (R1-Distill-Llama-8B)0.03954.3
ARLCP0.04458.7
DTSR0.02951.1
DeCoR0.25670.4
MethodΔHap\Delta H_{\text{ap}}↑UDS (%) ↑
Vanilla (Qwen3-8B)0.04859.7
IBPO0.05161.2
DEER0.03453.5
DeCoR0.27273.0
MethodΔHap\Delta H_{\text{ap}}↑UDS (%) ↑
Vanilla (Ministral-3-8B)0.06258.8
ConPress0.07160.5
SoT0.06056.2
DeCoR0.29573.3

The results show that DeCoR achieves the highest ΔHap\Delta H_{\text{ap}} and UDS across the three base models. This indicates that under a unified protocol, DeCoR exhibits a much clearer correspondence between answer correctness and predictive uncertainty.

To further assess measurement robustness, we evaluate the entropy of the last token of the same reasoning segments and intermediate answers used in Figure 4(b) for Qwen3-8B across all five benchmarks, while preserving the original gating mechanism, as it is an indispensable component enabling DeCoR to trigger correction and termination behaviors. This provides a consistent entropy definition across different methods and reasoning behaviors. The results are shown as follows:

MethodExploreContinueHigh-U CorrectInconsistency CorrectTerminate
Vanilla0.562———0.523
IBPO0.525———0.491
DEER0.538———0.502
DeCoR0.7750.4930.5690.4070.244

We also measure the entropy of the last reasoning token preceding each behavior in Figure 4(b), without involving any intermediate answer tokens. This aligns the measurement target across methods by consistently evaluating reasoning tokens rather than mixing reasoning and intermediate answer tokens. The results are shown as follows:

MethodExploreContinueHigh-U CorrectInconsistency CorrectTerminate
Vanilla0.562———0.523
IBPO0.525———0.491
DEER0.538———0.502
DeCoR0.7750.5230.6870.5710.439

Together, these analyses indicate that the uncertainty patterns observed in Figure 4 are not solely attributable to differences in entropy aggregation or output types. In particular, the pre-termination reasoning states exhibit lower uncertainty than general exploration states under matched measurements, providing additional evidence for the functional coordination of exploration and termination in DeCoR.

We have incorporated the corresponding figures and tables following Figure 4 and in Appendix F and Appendix G.

Q3: Does the exploration–termination entropy separation still hold when all methods are evaluated on the same output type, using identical entropy definitions and sampling temperatures, without conditioning on passing the low-uncertainty gate?

A3: Thanks for this important question. We conducted additional experiments extending the entropy comparison to further validate DeCoR's role decoupling mechanism. Specifically, we adopt the same reasoning token output type, entropy formulation, and sampling temperature (1.0) across methods. We also conduct unified intermediate answer generation without conditioning on the gating decisions of DeCoR. Note that termination in DeCoR is inherently defined by the gating mechanism. Removing this condition would alter the decision process of the method rather than provide a directly comparable termination event. These complementary analyses support the functional coordination of exploration and termination in DeCoR without assuming an intrinsic separation of reasoning states. Please refer to the response to Q2 for detailed results.


Q4: How does the reasoning deliberation ratio account for incorrect trajectories and temporarily correct answers? What evidence distinguishes a higher ratio driven by productive verification from one caused by redundant reasoning after an answer is already available?

A4: Thank you for raising this valuable question. As described in Section 4.3, RDR is computed only for reasoning trajectories with correct final answers, using the token position where the correct answer first appears as the onset. Trajectories with incorrect final answers are excluded, even if a correct answer appears during reasoning. For trajectories that eventually terminate with a correct answer, the initial onset is retained even if subsequent intermediate answers temporarily become incorrect before returning to correctness. To further distinguish effective reasoning from potentially redundant continuation, we conduct an additional experiment of intermediate answer dynamics using three complementary metrics: (1) Wrong-to-Correct Rate (WCR, %), the proportion of trajectories with at least one incorrect intermediate answer but a correct final answer. (2) Correct-to-Wrong Rate (CWR, %), the proportion of trajectories with at least one correct intermediate answer but an incorrect final answer. (3) Post-Stabilization Token Ratio (PSTR, %), the proportion of reasoning tokens generated after the answer becomes consistently correct. For DeCoR, intermediate answers are obtained through its native generation mechanism, whereas those for Vanilla are generated at the corresponding token positions using the same base model, with additional generation every 512 tokens when its reasoning trajectory extends beyond the termination of DeCoR. We report the evaluation metrics across the three base models on MATH500, AIME26, GPQA-D, and BBEH as follows:

Base ModelMethodWCR (%) ↑CWR (%) ↓PSTR (%) ↓
R1-Distill-Llama-8BVanilla34.221.328.6
DeCoR38.915.520.4
Qwen3-8B-ThinkingVanilla36.718.026.5
DeCoR39.514.320.8
Ministral-3-8B-ReasoningVanilla37.618.324.5
DeCoR41.213.918.5

DeCoR consistently outperforms Vanilla across all three base models, improving WCR by 2.8--4.7 percentage points while reducing CWR by 3.7--5.8 percentage points. Moreover, DeCoR decreases PSTR by 5.7--8.2 percentage points, indicating less redundant reasoning after answer stabilization. These findings suggest that the higher RDR under DeCoR stems from more effective error correction and reduced post stabilization overhead, rather than merely reflecting prolonged reasoning after reaching the correct answer.

We have placed this table following Figure 6 in the main text, accompanied by an explanatory description.

Dear Reviewer ieBf,

We hope this message finds you well. As the discussion period is nearing its end, we wanted to ensure that we have addressed all of your concerns satisfactorily. If there is any additional feedback or points you would like us to consider, please let us know. Your insights are invaluable to us, and we are eager to resolve any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their thoughtful and constructive feedback. We appreciate their recognition of the motivation of DeCoR, the novelty of its twin model collaborative reasoning framework, and its consistent improvements in reasoning accuracy and efficiency. We have carefully addressed the concerns raised in the reviews through additional controlled experiments, more detailed methodological analyses, and clarifications of our evaluation protocols.

To facilitate assessment by the Area Chair and reviewers, we summarize below our core contributions, the key additions and clarifications in our rebuttal, and the updates in the revised manuscript.


Core Contributions of Our Work

  • Decoupled Collaborative Reasoning: We propose DeCoR, a novel framework that functionally decouples exploration and termination, enabling more effective reasoning through dynamic twin-model collaboration.
  • Preference-Guided Pruning: We develop a structured pruning strategy for reasoning LLMs that quantifies layer influence on the next token distribution and evaluates layerwise importance, ensuring the construction of a reliable auxiliary explorer from the base model.
  • Dual-Indicator Gating: We introduce a gating mechanism comprising an uncertainty indicator and a consistency indicator to dynamically coordinate exploration, correction, and termination between the base model and the auxiliary model.
  • Performance Improvements: Extensive experiments across three base models and five benchmarks demonstrate consistent accuracy gains alongside substantial reductions in reasoning tokens and inference latency, without additional training. Further analyses reveal distinct uncertainty patterns associated with exploration and termination, alongside stronger discrimination between correct and incorrect intermediate answers. DeCoR also achieves higher reasoning deliberation ratios, more effective error correction, fewer answer regressions, and less redundant reasoning after answer stabilization.

Key Additions and Clarifications during Rebuttal

We carefully addressed the reviewers' concerns through additional experiments and methodological clarifications, with corresponding updates incorporated into the revised manuscript:

  • To address Reviewer ieBf's concerns (Q2–Q4) regarding mechanistic interpretation and reasoning effectiveness, we added unified intermediate answer probing, matched entropy measurements, and WCR / CWR / PSTR analyses to examine uncertainty patterns, error correction, and redundant reasoning.
  • To address Reviewer 6BDr's concerns (Q1) regarding gating reliability and auxiliary exploration, we conducted challenging uncertainty case analyses and matched-prefix experiments to validate answer discovery and semantic exploration novelty.
  • To address Reviewers 15U3 (Q1–Q2) and 6BDr (Q3) regarding parameter complexity and pruning generalizability, we conducted joint sensitivity and calibration robustness experiments, examining parameter interactions, mask stability, and out-of-domain performance.
  • To address Reviewer 6BDr's concerns (Q2) regarding statistical reliability and computational overhead, we added 95% accuracy confidence intervals and clarified the FLOPs and latency trade-off and benefits of shared KV caching.
  • To address Reviewers ieBf (Q1) and 15U3 (Q3) regarding evaluation ambiguity and applicability, we clarified the AIME26 accuracy aggregation protocol and the limitations of exact answer consistency for open ended generation.

Updates in the Revised Manuscript

  • Main Text, Appendix F, and Appendix G: Responses to Reviewer ieBf (Q2–Q3).
  • Main Text: Response to Reviewer ieBf (Q4).
  • Appendix C.1: Response to Reviewer ieBf (Q1).
  • Appendix H and Appendix J: Response to Reviewer 6BDr (Q1).
  • Appendix C.4: Response to Reviewer 6BDr (Q2).
  • Appendix I: Responses to Reviewer 6BDr (Q3) and Reviewer 15U3 (Q2).

We believe these additional experiments and clarifications substantially strengthen the empirical evidence, methodological understanding, and practical applicability of DeCoR. Detailed results and explanations are provided in our individual responses to the reviewers, with corresponding manuscript revisions where indicated.

We sincerely appreciate the reviewers' time and consideration, and we welcome any further questions or suggestions.

AuthorsOpenReview ↗

Summary Regarding Review Process for Paper 12277

Dear AC, SAC, and PCs,

As the discussion period approaches its conclusion, we would like to provide a concise summary of the review and rebuttal process for our paper to facilitate your final assessment.

We are encouraged that the reviewers recognized our core contributions. Reviewer 15U3 highlighted the innovation of the twin-model framework, the comprehensiveness of our experiments, and the practical value of our engineering analyses. Reviewer ieBf acknowledged the practical combination of lightweight exploration and selective base model intervention, together with the comprehensive evaluation. Reviewer 6BDr particularly recognized the memory-efficient implementation and systems-level analyses.


Responses to Reviewers' Concerns

We carefully responded to the reviewers' concerns through additional experiments and methodological clarifications, with corresponding updates incorporated into the revised manuscript:

  • To address Reviewer ieBf's concerns (Q2–Q4) regarding mechanistic interpretation and reasoning effectiveness, we added unified intermediate answer probing, matched entropy measurements, and WCR / CWR / PSTR analyses to examine uncertainty patterns, error correction, and redundant reasoning (Section 4.3, Appendices F–G).
  • To address Reviewer 6BDr's concerns (Q1) regarding gating reliability and auxiliary exploration, we conducted challenging uncertainty case analyses and matched-prefix experiments to validate answer discovery and semantic exploration novelty (Appendices H and J).
  • To address Reviewers 15U3 (Q1–Q2) and 6BDr (Q3) regarding parameter complexity and pruning generalizability, we conducted joint sensitivity and calibration robustness experiments, examining parameter interactions, mask stability, and out-of-domain performance (Rebuttal, Appendices D and I).
  • To address Reviewer 6BDr's concerns (Q2) regarding statistical reliability and computational overhead, we added 95% accuracy confidence intervals and clarified the FLOPs and latency trade-off and benefits of shared KV caching (Appendices C.4, K, and L).
  • To address Reviewers ieBf (Q1) and 15U3 (Q3) regarding evaluation ambiguity and applicability, we clarified the AIME26 accuracy aggregation protocol and the limitations of exact answer consistency for open-ended generation (Rebuttal, Appendix C.1).

Detailed results and explanations are available in the individual reviewer responses and the revised manuscript.


Reviewer Feedback and Score Updates

We sincerely appreciate Reviewer 15U3's follow-up on November 18, explicitly confirming that, after carefully reviewing our rebuttal and supplementary materials, they maintained their positive score and increased their confidence from 3 to 4. We are grateful for this thoughtful reassessment.

As of this summary, neither of the other two reviewers has provided a follow-up response or made any public score changes. We hope our extensive responses and supplementary evidence will facilitate their further assessment and the committee's final evaluation.


We believe the rebuttal and manuscript revisions have substantially strengthened DeCoR's empirical validation, methodological clarity, and practical relevance. We sincerely thank all reviewers, Area Chairs, Senior Area Chairs, and Program Chairs for their time, constructive feedback, and careful consideration.

Best regards,

The Authors of Paper 12277

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.