acceptodds
Under review as a conference paper at ICLR 2027

Learning More by Training Less: Uncertainty-Guided Progressive Contraction for Efficient Reasoning Post-Training

Abstract

Large language models have achieved remarkable advances in complex reasoning. However, scaling reasoning post-training with ever-larger corpora can substantially increase training cost without proportional performance gains. In this study, we therefore ask: how far can reasoning supervision be compressed without sacrificing capability? To this end, we propose an uncertainty-guided progressive contraction framework that concentrates supervision and optimization on the reasoning trajectories most valuable for learning. Specifically, rather than treating a teacher-generated chain-of-thought (CoT) as a single realized path, we construct uncertainty-profiled reasoning supervision by building Monte Carlo reasoning trees around intermediate reasoning states to expose process-level uncertainty. This uncertainty first contracts the original corpus into a compact teacher-informed pool. A second, learner-aware contraction then performs online routing during SFT, concentrating optimization according to the student's evolving learning state. Together, the two stages progressively contract both supervision and optimization, substantially reducing the student training budget. Experiments across diverse reasoning benchmarks and LLM families show that our method uses as little as 0.22% of the optimization budget of full-data SFT, over 450 less training exposure, while maintaining or improving reasoning performance. Notably, our framework generalizes well across heterogeneous student families, black-box teachers, and weak-to-strong transfer. This indicates that it is not only a way to accelerate training but can also enhance reasoning capability.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Dynamic Data Selection, Dataset Distillation, Data-Effciient Learning

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_iJG9OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes two-stage data selection for reasoning SFT. Teacher trajectories are ranked by disagreement in final-answer correctness among continuations from intermediate steps. Correct trajectories with high uncertainty form a candidate pool. Student response losses then select a smaller training subset. Experiments cover multiple student families, weak-to-strong transfer, and API-only teachers. The paper reports comparable or better reasoning performance with student sample exposure as low as 0.22% of full-data SFT. This figure excludes teacher-side uncertainty profiling.

Critical strengths

The comparison within the same pool of correct teacher responses is useful. In Table 4, uncertainty-based selection improves the seven-benchmark average by 1.6–4.5 points over random selection. This is more informative than comparing only against full-data SFT. Reusing one teacher profile across different student families is also a useful design choice.

Critical weaknesses
  • The student-side savings are substantial, but the practical cost advantage remains unclear.
  • The learner-routing procedure is described inconsistently. Section 2.3 says scores and the active subset are updated during training. Appendix E instead fixes the subset and sampling weights after the first epoch.
Questions for authors
  • What rollout count and total teacher-profiling cost were used?
  • Which routing procedure produced the reported results: adaptive rescoring or the fixed subset in Appendix E?
  • Please clarify: Does Table 3 represent one checkpoint? Its entries match the per-benchmark maxima across Table 1. In particular, AIME25 = 93.3 and HMMT25 = 90.0 occur at different retention ratios.
3 comments

Dear Reviewer iJG9,

Thank you for the careful review and constructive suggestions. We appreciate your recognition of our work's strengths, e.g., strong reasoning performance, extensive experimental results, and useful method design.

We provide responses to address the comments below:

  • Q1/2: More details of the teacher profiling costs and end-to-end efficiency.
  • A1/2: Thanks for the suggestion. We would like to emphasize that our framework separates one-time, student-agnostic uncertainty profiling from subsequent learner-specific optimization. Thus, the teacher-side profiling can be constructed once and reused across downstream student training, highlighting its practical efficiency.

1. As suggested, we first provide a practical cost analysis of the complete pipeline on both Qwen3-8B/1.7B teachers (Table C-1), covering offline teacher profiling (2 rollouts) and online student training. Importantly, due to our decoupled design, the teacher-side profiling incurs only a one-time, student-agnostic cost, and can be reused across student training runs. Moreover, Table C-2 further compares our complete pipeline with full-data SFT. Even under this minimal reuse, training just one 4B/9B student using a shared profile, our method achieves 15.2%–15.8% end-to-end cost savings. Furthermore, Table C-3 demonstrates that the additional profiling cost is fully amortized after only two student training runs, with further reuse yielding increasing cumulative savings, highlighting our practical efficiency.

2. To fully address your comment, beyond GPU-hour accounting, Table C-4 further evaluates actual response-token consumption, total non-padding tokens, and training GPU-hours. At ρ=1%\rho=1\%, our method processes only 0.016% of response tokens and 0.019% of total non-padding tokens compared with full SFT, which validates the practical value of amortizing teacher-side profiling across learners.

We have included these additional results in Appendix K on pages 30-31 of the revised manuscript and will publicly release the uncertainty-profiled reasoning data to facilitate broader reuse.

Table C-1: The wall-clock GPU hours using Qwen3-8B/1.7B as the teacher on a 4-H200-GPU server. We report CPU hours on supervision contraction.

TeacherReasoning ElicitationMCRT ProfilingSupervision ContractionTotal Teacher-sideOnline Student (Probing + Routing + SFT)
Qwen3-8B3.114.40.000517.5≤0.07 (4B/9B)
Qwen3-1.7B1.86.10.00057.9≤0.10 (4B/14B)
Cost ScopeOne-timeOne-timeOne-timeOne-time; resuablePer student

Table C-2: End-to-end computational cost comparison between full SFT and Ours for one training run per student.

Cost (GPU-h)Full SFTOurs (30%)Ours (10%)Ours (5%)Ours (1%)
Teacher Reasoning Elicitation3.13.13.13.13.1
MCRT Profiling + Construction—14.414.414.414.4
4B student training8.470.070.030.020.01
9B student training9.230.070.030.020.01
Total GPU-hours20.8017.6417.5617.5417.52
End-to-end reduction—15.2%15.6%15.7%15.8%

Table C-3: Amortized computational efficiency and break-even analysis.

ρ\rho4B Online9B OnlineSavings per Run (4B / 9B)Break-even (4B / 9B)
30%0.070.078.40 / 9.162 runs
10%0.030.038.44 / 9.202 runs
5%0.020.028.45 / 9.212 runs
1%0.010.018.46 / 9.222 runs

Table C-4: Token-level computational efficiency for Qwen3.5-4B across contraction ratios.

SettingBudgetResponse TokensToken RatioTotal Non-padding TokensNon-padding RatioGPU hours
Full SFT100%149.15M100%153.54M100%8.47
Ours (30%)1.87%492.2K0.33%572.9K0.37%0.07
Ours (10%)0.72%167.5K0.11%199.3K0.13%0.03
Ours (5%)0.44%91.7K0.06%110.1K0.07%0.02
Ours (1%)0.17%23.3K0.016%29.9K0.019%0.01
  • Q3/4: Clarification on the learner-routing procedure.

  • A3/4: Thanks for pointing this out. We would like to clarify that we perform learner-aware routing after the probing epoch and use this selected active subset for the remaining two epochs to avoid additional rescoring overhead. We have clarified this in Sec. 2.3 and Appendix E of the revised manuscript.

  • Q5: Clarification on Table 3.

  • A5: Thanks for pointing this out. We clarify that all results in Table 3 are from a single Qwen3.5-9B checkpoint (ρ=30%\rho=30\%). We have corrected the HMMT25 accuracy from 90.0% to 86.7% in Table 3 and clarified the checkpoint configuration in the caption of the revised manuscript.

Reviewer_iJG9OpenReview ↗

Reply

Your answer has answered some of my questions. I will increase my score.Thank you

Dear Reviewer iJG9,

We would like to express our sincere gratitude for acknowledging our work and providing constructive feedback. Thank you again for your time and effort in reviewing our manuscript.

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

This paper proposes an uncertainty-guided progressive contraction framework for efficient reasoning post-training. The method first estimates process-level uncertainty by sampling multiple continuations from intermediate teacher reasoning states and retains a small pool of verified trajectories with high uncertainty. It then uses the student’s response loss to prioritize a smaller subset during supervised fine-tuning. Experiments across mathematical and scientific reasoning benchmarks, multiple student architectures, and different teacher settings report substantial reductions in student-side training exposure while maintaining or improving reasoning accuracy.

Critical strengths
  1. Promising empirical results under highly constrained student training budgets.

Table 1 reports competitive or improved performance using only 0.22–1.91% of the full-SFT training exposure. The experiments also cover multiple student families and teacher settings, including black-box and weak-to-strong distillation. These results suggest that carefully selecting reasoning supervision can substantially reduce the student optimization required for effective post-training.

  1. Useful ablations supporting the proposed selection strategy.

The paper goes beyond comparisons with full-data SFT. Table 4 compares uncertainty-based selection with random selection under matched budgets, Table 5 evaluates matched-budget distillation, and Table 6 examines the two components separately. Appendices F and I further compare alternative selection criteria and control for reasoning length. These experiments provide meaningful evidence that the proposed selection strategy can be more effective than several straightforward alternatives.

Critical weaknesses
  1. The efficiency claim does not account for the full computational cost of the method.

The central contribution is framed around extreme training efficiency, including a reported reduction to 0.22% of the full-SFT optimization budget. However, this metric measures student-side sample exposure and excludes the cost of constructing Monte Carlo reasoning trees. Section 2.2 requires multiple teacher continuations at intermediate reasoning states, potentially introducing substantial inference overhead. Although Appendix J acknowledges this limitation, the experiments do not quantify the complete teacher-profiling cost or establish the end-to-end computational break-even point. Consequently, the reported efficiency gains do not yet demonstrate that the overall pipeline is more computationally efficient than simpler distillation approaches.

  1. The necessity of the proposed uncertainty-guided mechanism is not fully established.

The paper shows that uncertainty-based selection outperforms random selection and several alternative statistics. However, these results do not fully establish whether the expensive Monte Carlo uncertainty estimation is necessary to achieve the reported gains. In particular, the main experiments do not provide a comprehensive cost-matched comparison against simpler student-loss-based selection or difficulty-based filtering that avoids reasoning-tree construction. Moreover, Table 9 shows only a modest improvement over random selection in the GSM8K setting (82.11% versus 81.58%). This leaves open whether the proposed uncertainty signal provides sufficient additional learning value to justify its computational complexity.

Questions for authors
  1. End-to-end efficiency: What are the total GPU hours and generated tokens required for teacher-side uncertainty profiling, including all Monte Carlo continuations? How does the complete pipeline compare with standard distillation when teacher inference and student training costs are both included?
  2. Necessity of uncertainty profiling: Under a matched total computational budget, how does the proposed method compare with simpler selection strategies based on student loss, teacher difficulty, or other inexpensive trajectory-level signals?
  3. Implementation consistency: Section 2.3 states that routing scores are updated during training, whereas Appendix E describes selecting a fixed high-loss subset after the first epoch. Which procedure is used in the reported experiments? Could the authors also reconcile the training-budget calculations in Table 1 and Appendix D?
10 comments

Dear Reviewer Zhyb,

We sincerely thank you for the careful review and insightful suggestions. We appreciate your recognition of our work's strengths, e.g., promising empirical results, effective method design, useful ablation studies, and broad generalization across diverse student and teacher settings.

We provide responses to address the comments below:

  • Q1/2: More results on the end-to-end computational efficiency.
  • A1/2: Thanks for the suggestion. We would like to highlight a key design advantage of our framework: it decouples one-time, student-agnostic uncertainty profiling from subsequent learner-specific optimization. This allows the resulting uncertainty profiles to be reused across downstream students without repeated teacher inference, highlighting its practical efficiency. Consequently, the profiling cost remains fixed, while optimization savings accumulate with each additional student training run.

1. As suggested, in Table B-1, we first provide a complete cost breakdown of teacher profiling and student training using both Qwen3-8B/1.7B teachers. Notably, only student-side optimization incurs recurring costs (including complete probing, scoring, and SFT), which our framework reduces to ≤0.07 GPU hours for 4B/9B students and 0.10 GPU hours for 4B/14B students. Table B-2 compares the end-to-end costs with full SFT. Even with minimal reuse of training just one 4B and 9B students with a shared profile, the shared-profile setting achieves 15.2%–15.8% end-to-end cost savings across all ratios. Furthermore, break-even analyses in Table B-3 show that the additional profiling investment is fully amortized after only two student training runs, with further reuse yielding increasing cumulative savings, highlighting its efficiency.

2. Beyond sample exposure and GPU hours, in Table B-4, we further provide a complementary token-level efficiency. At ρ=1%\rho=1\%, student response-token and non-padding-token consumption drop to just 0.016% and 0.019% of full SFT, respectively, corresponding to over 6000× fewer response tokens, with only 0.01 GPU hours of online training, which validates the practical value of amortizing teacher-side profiling across learners. Together, these results show that our framework converts one-time teacher-side profiling into reusable reasoning supervision, enabling extreme student-side optimization efficiency and increasing computational savings across downstream learners.

3. We have included these analyses in Appendix K on pages 30-31 of the revised manuscript. We will also publicly release the uncertainty-profiled reasoning data to facilitate broader reuse.

Table B-1: The wall-clock GPU hours using Qwen3-8B/1.7B as the teacher on a 4-H200-GPU server with DAPO-17k. We report CPU hours on supervision contraction.

TeacherReasoning ElicitationMCRT ProfilingSupervision ContractionTotal Teacher-sideOnline Student (Probing + Routing + SFT)
Qwen3-1.7B1.86.10.00057.9≤0.10 (4B/14B)
Qwen3-8B3.114.40.000517.5≤0.07 (4B/9B)
Cost ScopeOne-timeOne-timeOne-timeOne-time; resuablePer student

Table B-2: End-to-end computational cost comparison between full SFT and Ours for one training run per student.

Cost (GPU-h)Full SFTOurs (30%)Ours (10%)Ours (5%)Ours (1%)
Teacher Reasoning Elicitation3.13.13.13.13.1
MCRT Profiling + Construction—14.414.414.414.4
4B student training8.470.070.030.020.01
9B student training9.230.070.030.020.01
Total GPU-hours20.8017.6417.5617.5417.52
End-to-end reduction—15.2%15.6%15.7%15.8%

Table B-3: Amortized computational efficiency and break-even analysis.

ρ\rho4B Online9B OnlineSavings per Run (4B / 9B)Break-even (4B / 9B)
30%0.070.078.40 / 9.162 runs
10%0.030.038.44 / 9.202 runs
5%0.020.028.45 / 9.212 runs
1%0.010.018.46 / 9.222 runs

Table B-4: Token-level computational efficiency for Qwen3.5-4B across contraction ratios.

SettingBudgetResponse TokensToken RatioTotal Non-padding TokensNon-padding RatioGPU hours
Full SFT100%149.15M100%153.54M100%8.47
Ours (30%)1.87%492.2K0.33%572.9K0.37%0.07
Ours (10%)0.72%167.5K0.11%199.3K0.13%0.03
Ours (5%)0.44%91.7K0.06%110.1K0.07%0.02
Ours (1%)0.17%23.3K0.016%29.9K0.019%0.01
  • Q3/4: Further analysis on process-level uncertainty.
  • A3/4: Thanks for your suggestion. To further examine the effectiveness of process-level uncertainty, we test correctness-, student-loss-, and teacher-difficulty-based criteria within our framework under a matched total computational budget (Table B-5). The results show that our method consistently achieves the best results across benchmarks, with particularly substantial gains on challenging competition-level mathematics (e.g., +13.4% on HMMT25 and +6.1% on HMMT26 at the 10% ratio). Thus, these results demonstrate that process-level uncertainty provides more effective learning value beyond conventional selection signals, while retaining high practical efficiency through reusable teacher-side profiling.

Table B-5: Comparison of uncertainty-guided selection with other criteria across contraction ratios on Qwen3.5-4B.

MethodRatioAIME24AIME25AIME26AMCHMMT25HMMT26MATH-500GSM8KGPQA-D
Correctness1%80.060.070.087.356.760.694.695.874.2
Student Loss1%83.373.373.390.366.766.796.895.274.8
Teacher Difficulty1%83.376.780.090.370.063.696.895.075.3
Ours1%90.083.386.794.076.769.798.496.377.8
Correctness5%83.366.773.391.860.060.694.695.275.3
Student Loss5%80.070.076.790.366.766.797.495.275.3
Teacher Difficulty5%80.070.070.090.363.360.697.295.276.8
Ours5%86.773.380.093.370.069.797.896.377.3
Correctness10%80.063.380.087.356.766.797.495.274.2
Student Loss10%80.073.373.393.363.372.796.695.275.8
Teacher Difficulty10%83.376.773.392.563.366.797.495.075.3
Ours10%86.783.383.394.876.778.898.096.277.3
Correctness30%80.066.780.092.563.366.797.495.875.3
Student Loss30%80.073.380.093.366.760.697.496.275.3
Teacher Difficulty30%80.076.776.790.366.766.797.895.075.3
Ours30%86.783.386.796.376.769.798.496.477.8
  • Q5: Clarification of the implementation.

  • A5: Thanks for pointing this out. We would like to clarify that, in our three-epoch training schedule, learner-aware routing selects an active subset after the initial probing epoch, which is then fixed for the remaining two epochs to avoid additional rescoring overhead.

    We have clarified this in Sec. 2.3. We have also reconciled the budget calculations in Table 1 and Appendix D of the revised manuscript.

AuthorsOpenReview ↗

Looking forward to the reply

Dear Reviewer Zhyb:

Thank you again for your time and constructive feedback. In response to your comments, we have conducted additional experiments and analyses, along with clarifications of the learner-routing implementation. We have also incorporated these revisions into the updated manuscript.

As the discussion period is approaching its end, we would greatly appreciate your feedback on whether our responses have adequately addressed your concerns. If there are further concerns or questions, we are willing to address them.

Thanks again for taking time to review our work and provide insightful comments.

Authors of Submission 38506

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

This paper proposes uncertainty-guided progressive contraction, a two-stage framework for reducing supervised reasoning post-training while preserving the benefits of teacher-generated chains of thought. First, Monte Carlo reasoning trees estimate uncertainty across teacher continuations to select informative, verified trajectories. Second, student-side probing identifies high-loss examples for targeted supervised fine-tuning. Experiments across multiple student models, teacher settings, and reasoning benchmarks demonstrate competitive accuracy while reducing student-side training exposures to just 0.22–1.91% of conventional full-data SFT. The main contribution is an efficient two-stage supervision allocation strategy, although its overall computational efficiency remains uncertain without fully accounting for teacher-side profiling and student-side selection costs.

Critical strengths
  1. Coherent two-stage problem formulation. The method distinguishes a teacher-side question—what reasoning paths contain informative uncertainty—from a learner-side question—what the current student still needs to learn. This makes the design more purposeful than indiscriminate filtering, and Eqs. 5–7 specify both selection stages clearly. The separation could also make the teacher profile reusable across different students.

  2. A measurable process-level selection signal. The continuation-correctness estimator in Eqs. 1–4 gives a concrete, behavior-based interpretation of reasoning uncertainty rather than relying only on the confidence or correctness of one completed CoT. The analysis behind Figure 1 helps explain why apparently correct or confident trajectories may remain informative. Importantly, the score does not require teacher logits or hidden states.

  3. Substantial empirical compression of student supervision. Table 1 reports results for Qwen3.5-4B and 9B over nine mathematical/science metrics, including several configurations using under 1% of the full-SFT trajectory exposure. The size of the reported exposure reduction makes the research problem consequential even if the final cost savings turn out to be smaller than the headline ratios.

  4. Transfer across model and teacher regimes. Figure 2 and Tables 2–4 examine heterogeneous student architectures, a smaller teacher supervising larger students, and API-only teachers. This breadth is useful because a selector tied to one tokenizer or model-internal uncertainty signal would be much less deployable.

Critical weaknesses
  1. The central efficiency claim omits a major part of the pipeline. The Abstract and Introduction emphasize “0.22%” and “over 450× less training exposure,” but Eq. 8 explicitly defines only student-side trajectory exposures. Constructing the Monte Carlo reasoning trees entails multiple teacher continuations per intermediate state, outcome verification, and possible API expenditure, none of which is included in that ratio. Table 1 labels its GPU-hour column “Overhead,” but does not establish a complete teacher-plus-student cost ledger. Consequently, the paper supports a striking reduction in student SFT exposure, not yet an equally striking reduction in end-to-end compute, money, or elapsed time. This distinction is central to its main contribution, not a minor reporting issue.

  2. Trajectory counts are not a hardware-neutral optimization budget. The numerator and denominator of Eq. 8 count instances rather than response tokens or FLOPs. Reasoning sequences vary substantially in length; the probing pass, repeated loss scoring, data loading, and rerouting may further change cost. The appendix does provide length-based selection controls (Table 10), which is useful, but does not replace a token- and GPU-time-normalized accounting of the actual training procedure. The claimed speedup should therefore be described as an exposure ratio unless validated against measured resources.

  3. The accuracy–budget comparison needs a stronger Pareto frontier. Table 1 contrasts the proposed tiny active set with a three-epoch full-SFT reference; this does not by itself show that the method dominates well-tuned early-stopped SFT or random/correct-only/high-loss selection at the same token, wall-clock, and update budgets. Table 5 and the appendix selection comparisons partially address this issue, so the concern is not an absence of any controls. Rather, these controls should be extended to the strongest operating points and reported as a multi-budget frontier with identical optimizer settings and accounting boundaries. Otherwise some apparent gains could reflect the full-data baseline being overtrained or evaluated at an unfavorable budget.

Questions for authors
  1. Please provide a complete per-student and amortized cost breakdown: teacher-generated tokens/rollouts, number of profiled intermediate states, verification cost, GPU-hours/API charges, student probing/rerouting, and SFT training. For how many student runs would the initial profiling investment break even?

  2. How variable are the results in Table 1 across student-training seeds, teacher sampling seeds, and evaluation generations? Could you report paired intervals, particularly for AIME24/25/26 and HMMT?

5 comments

Dear Reviewer nNZw,

We sincerely thank you for the insightful comments and constructive suggestions. We appreciate your recognition of our work's strengths, e.g., principled methodological design, meaningful research contributions, and strong empirical results.

We provide responses to address the comments as follows:

  • Q1/2/3: Additional results on the end-to-end efficiency.
  • A1/2/3: Thanks for the suggestion. We would like to emphasize that our framework separates one-time, student-agnostic uncertainty profiling from subsequent learner-specific optimization. Thus, the teacher-side profiling can be constructed once and reused across downstream student training, highlighting its practical efficiency.

1. As suggested, in Table A-1, we first provide a complete cost breakdown of offline teacher-side contraction and online learner-aware optimization using both Qwen3-8B/1.7B teachers. Notably, only student-side optimization incurs recurring costs (including complete probing, scoring, and SFT), which our framework reduces to ≤0.07 GPU hours for 4B/9B students and 0.10 GPU hours for 4B/14B students. In Table A-2, we further compare the end-to-end costs with full SFT. Even with this minimal reuse, i.e., training just one 4B and 9B student using a shared teacher profile, our method already achieves 15.2%–15.8% end-to-end cost savings across all ratios ρ\rho. The break-even analyses in Table A-3 show that the entire profiling investment is amortized after only two student training runs, with further reuse yielding increasingly substantial cumulative savings, highlighting our practical efficiency.

2. To fully address your comment, beyond GPU-hour accounting, we further measured the actual response-token consumption, total non-padding tokens, and training GPU-hours (Table A-4). At ρ=1%\rho=1\%, student response-token and non-padding-token consumption drop to just 0.016% and 0.019% of full SFT, respectively, corresponding to over 6000× fewer response tokens, with only 0.01 GPU hours of online training, which validates the practical value of amortizing teacher-side profiling across learners.

3. Following your suggestion, we have clarified that Eq. 8 measures student-side sample exposure, and included these additional results in Appendix K on pages 30-31 in the revised manuscript. We will also publicly release our uncertainty-profiled CoT datasets, enabling community reuse without additional teacher-side profiling costs.

Table A-1: The wall-clock GPU hours using Qwen3-8B/1.7B as the teacher on a 4-H200-GPU server. We report CPU hours on supervision contraction.

TeacherReasoning ElicitationMCRT ProfilingSupervision ContractionTotal Teacher-sideOnline Student (Probing + Routing + SFT)
Qwen3-1.7B1.86.10.00057.9≤0.10 (4B/14B)
Qwen3-8B3.114.40.000517.5≤0.07 (4B/9B)
Cost ScopeOne-timeOne-timeOne-timeOne-time; resuablePer student

Table A-2: End-to-end computational cost comparison between full SFT and Ours for only one training run per student.

Cost (GPU-h)Full SFTOurs (30%)Ours (10%)Ours (5%)Ours (1%)
Teacher Reasoning Elicitation3.13.13.13.13.1
MCRT Profiling + Construction—14.414.414.414.4
4B student training8.470.070.030.020.01
9B student training9.230.070.030.020.01
Total GPU-hours20.8017.6417.5617.5417.52
End-to-end reduction—15.2%15.6%15.7%15.8%

Table A-3: Amortized computational efficiency and break-even analysis.

ρ\rho4B Online9B OnlineSavings per Run (4B / 9B)Break-even
30%0.070.078.40 / 9.162 runs
10%0.030.038.44 / 9.202 runs
5%0.020.028.45 / 9.212 runs
1%0.010.018.46 / 9.222 runs

Table A-4: Comparison of cumulative sample exposures, response-text tokens, non-padding tokens, and wall-clock GPU hours using Qwen3.5-4B/9B students.

StudentSettingBudgetResponse TokensToken RatioTotal Non-padding TokensNon-padding RatioGPU hours
Qwen3.5-4BFull SFT100%149.15M100%153.54M100%8.47
Ours (30%)1.87%492.2K0.33%572.9K0.37%0.07
Ours (10%)0.72%167.5K0.11%199.3K0.13%0.03
Ours (5%)0.44%91.7K0.06%110.1K0.07%0.02
Ours (1%)0.17%23.3K0.016%29.9K0.019%0.01
Qwen3.5-9BFull SFT100%149.15M100%153.54M100%9.23
Ours (30%)1.87%491.9K0.33%572.4K0.37%0.07
Ours (10%)0.72%166.6K0.11%198.7K0.13%0.03
Ours (5%)0.44%91.3K0.06%109.7K0.07%0.02
Ours (1%)0.17%24.2K0.016%31.2K0.020%0.01
  • Q4: Additional analysis on the Pareto frontier.

  • A4: Thanks for the suggestion. To better illustrate the Pareto frontier, we have added a dedicated accuracy–budget Pareto frontier illustration in Appendix J (Fig. 6) on page 29 of the revised manuscript. Specifically, we compare with early-stopped SFT, correct-only selection, and high-loss selection across four budget settings. Our method consistently outperforms others, demonstrating that its gains are not an artifact of an unfavorable full-SFT reference, but stem from more effective utilization of limited supervision and optimization budgets.

  • Q5: Additional variability results.

  • A5: Thanks for the comment. We provide mean improvement ± standard deviation results over full SFT across both generation (Table A-5) and training seeds (Table A-6) using Qwen3.5-9B on AIME24/25/26 and HMMT25/26 across three independent runs. The results demonstrate consistent gains on these challenging benchmarks.

Table A-5: Variability on the evaluation generation seeds.

RatioAIME24AIME25AIME26HMMT25HMMT26
1%2.2±1.92.2±1.91.1±1.911.1±1.914.1±1.7
5%1.1±1.93.3±0.01.1±1.97.8±1.914.1±1.7
10%2.2±3.80.0±3.30.0±3.310.0±5.817.2±1.7
30%2.2±1.94.4±1.9-1.1±1.914.4±1.916.2±3.5

Table A-6: Variability on the training seeds.

RatioAIME24AIME25AIME26HMMT25HMMT26
1%1.1±1.91.1±3.9-2.2±1.912.3±2.014.2±4.6
5%0.0±3.32.2±1.93.3±3.410.0±3.416.2±1.7
10%2.2±1.91.1±1.91.1±1.915.6±1.914.2±4.6
30%2.2±1.93.3±3.31.1±1.912.3±2.014.2±4.6
AuthorsOpenReview ↗

Looking forward to the reply

Dear reviewer nNZw:

Thanks so much again for your time and constructive suggestions. In response to your comments, we have conducted additional experiments and analyses, including end-to-end efficiency analyses, stronger accuracy-budget comparisons, and robustness analyses. We have also revised our manuscript to further incorporate these new results.

As the discussion period is coming to a close, we would greatly appreciate your feedback on whether these additional results and revisions have addressed your concerns. If there are further questions, please feel free to let us know.

Thanks again for taking the time to review our work and provide insightful feedback.

Authors of Submission 38506

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

General Description:

Dear Area Chairs and Reviewers,

We sincerely thank you for the time and effort in reviewing our work. We greatly appreciate the constructive feedback and the recognition from all reviewers—nNZw (R1), Zhyb (R2), and iJG9 (R3)—highlighting key strengths of our work, including (1) principled methodological design (R1, R2), (2) promising and strong empirical results (R1, R2, R3), (3) informative experimental comparisons and ablations (R2, R3), and (4) meaningful contributions (R1).

Besides, the reviewers' concerns are mainly concentrated on (1) additional efficiency analyses and experiments (R1, R2, R3), and (2) specific refinements to improve clarity (R2, R3). In response, we have systematically addressed these concerns through extensive additional experiments, detailed analyses, and corresponding manuscript revisions.

Updates of Experiments and Analyses:

Specifically, we have provided the following additional evidence:

  1. End-to-end efficiency analysis (R1: Q1/2/3, R2: Q1/2, R3: Q1/2): We provided detailed teacher- and student-side cost accounting, token-level measurements, and amortization analyses, demonstrating over 15% end-to-end computational savings even after fully accounting for the one-time offline teacher profiling costs, along with a break-even point of only two student training runs. (Table A/B/C-1/2/3/4, Appendix K)
  2. Stronger Pareto frontier analysis (R1: Q4): Following the reviewer's suggestion, we added additional comparisons with early-stopped SFT and selection criteria, demonstrating consistently superior accuracy–budget trade-offs and forming a stronger Pareto frontier. (Fig. 6, Appendix J)
  3. Variability ayalysis (R1: Q5): We conducted additional analyses across seeds, showing consistent performance improvements. (Table A-5/6)
  4. Effectiveness of the uncertainty-guided mechanism (R2: Q3/4): We added controlled experiments and demonstrated that our method consistently achieves the best results, particularly on challenging competition-level mathematics. (Table B-5)

We have carefully addressed each individual comment provided by the reviewers, and these results and corresponding clarifications have been incorporated into the revised manuscript, which has been highlighted for ease of reference.

We sincerely appreciate the reviewers' constructive feedback and hope that our detailed responses and revised manuscript have satisfactorily addressed their concerns. We welcome any further discussion.

Sincerely,

Authors of Submission 38506

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.