acceptodds
Under review as a conference paper at ICLR 2027

EduMCTS: Improving Large Language Model Reasoning through Adaptive Pedagogical Interaction

Abstract

High-quality reasoning trajectories serve as critical post-training resources for enhancing the reasoning abilities of large language models (LLMs). However, existing reasoning data augmentation methods mainly focus on exploring diverse candidate solutions and selecting among reasoning trajectories, while overlooking how to actively improve intermediate reasoning during trajectory generation. To mitigate this limitation, we propose EduMCTS, a reasoning data augmentation framework that constructs reasoning trajectories by selectively introducing pedagogical interventions during problem solving. Specifically, EduMCTS formalizes pedagogical strategies as executable search actions for intermediate reasoning states, evaluates candidate reasoning paths using local reasoning progress and final answer correctness, and leverages tree search to construct verified interactive reasoning trajectories. Extensive experiments across three LLMs and eleven reasoning benchmarks, spanning seven mathematical and four scientific reasoning tasks, demonstrate that EduMCTS consistently outperforms existing data augmentation methods and generalizes well to out-of-domain scientific reasoning. Further analyses support the role of pedagogical strategy search and show that models trained with EduMCTS exhibit adaptive pedagogical reasoning behaviors during inference. Our code and dataset are anonymously available at https://anonymous.4open.science/r/EduMCTS-B568/.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Reasoning Data Augmentation, Large Language Models, Instruction Fine-Tuning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_S1tFOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EduMCTS, a framework for synthesizing reasoning supervision through adaptive teacher–student interaction. It uses MCTS to search over four actions—autonomous continuation, hints, critiques, and Socratic questions—and evaluates their resulting trajectories using local progress and terminal correctness. Verified interaction trajectories are serialized for supervised fine-tuning. Experiments on three backbone models and eleven mathematical and scientific reasoning benchmarks show improvements over several data augmentation baselines.

Critical strengths
  1. The framework is clearly presented, with detailed prompts and implementation settings that make the data construction procedure understandable.

  2. EduMCTS achieves the highest overall average across all three backbones, including comparisons with relevant pedagogical and search-based baselines, MASTER and CRPS. Component ablations provide additional evidence for the usefulness of the proposed design.

Critical weaknesses
  1. The methodological novelty appears limited relative to existing work. MASTER and SoDa already construct reasoning supervision through pedagogical interactions, rStar searches over diverse reasoning actions, and ReST-MCTS* combines process-guided search with intermediate self-critique and training data generation. EduMCTS’s main distinction is to make typed pedagogical interventions searchable within this existing framework. While reasonable, this appears to be an incremental combination, and the paper does not sufficiently establish what distinctive capability it provides beyond these approaches.
Questions for authors
  1. Why was evaluation restricted to greedy decoding? Do the relative improvements remain under top-p sampling?

  2. Why do the components in Table 3 have different effects across benchmarks? For example, removing search substantially hurts AIME but slightly improves MMLU-Pro Math. Is there an intuitive explanation, and are these differences stable across runs?

  3. Why does SFT on the original solutions sometimes perform substantially worse than the backbone without SFT? Have the authors investigated whether this results from the supervision itself or sensitivity to learning rate and training duration?

5 comments
AuthorsOpenReview ↗

Responses to Reviewer S1tF (1/2)

Thank you for your thoughtful and helpful feedback. We will address your questions point by point below.

W1.The methodological novelty appears limited relative to existing work.

Thank you for your insightful comment. EduMCTS is inspired by prior work, and its core innovation lies in adaptively selecting effective pedagogical interventions based on intermediate reasoning states during data synthesis. This design addresses two limitations of existing approaches:

First, existing pedagogical interaction methods lack adaptive intervention selection across different reasoning states. MASTER uses predefined teaching scenarios, while SoDa primarily guides reasoning through Socratic questioning. Yet different reasoning states may require different interventions. EduMCTS formulates pedagogical strategies as search actions, adaptively deciding whether and how to intervene based on the student's actual reasoning progress.

Second, existing reasoning search methods do not explicitly compare the effects of alternative feedback. rStar explores problem-solving actions such as subproblem decomposition and re-answering, while ReST-MCTS* uses shared Self-Critic feedback for candidate steps within each node expansion. Given the risks of unreliable self-correction [1, 2], EduMCTS jointly searches over pedagogical interventions and their induced student reasoning steps, evaluating their effects through local reasoning progress and terminal correctness to construct training trajectories.

Thus, EduMCTS explores not only how to continue reasoning, but also when and how to provide effective guidance based on the current reasoning state. This capability is also reflected in the fine-tuned model's reasoning behavior. Table 4 and Appendix C.2 present examples of the model improving intermediate reasoning through critiques, hints, and Socratic questioning.

Q1. Why was evaluation restricted to greedy decoding? Do the relative improvements remain under top-p sampling?

Thank you for your careful review and suggestion. Following common evaluation practices in LLM reasoning research [3, 4], we used greedy decoding to reduce randomness and ensure fair comparisons across methods.

To examine performance under different decoding strategies, we additionally evaluated EduMCTS and two strong baselines, MASTER and CRPS, using top-p = 0.9 with temperatures of 0.5 and 0.7. We conducted three independent evaluation runs on MATH-500, GSM8K, and AMC and report the mean accuracy below.

Table 1: Reasoning Accuracy (%) under Top-p Sampling (Qwen3-8B-Base)

TemperatureMethodMATH-500GSM8KAMCAvg.
0.5MASTER73.0791.4735.0066.51
0.5CRPS77.4093.1741.5370.70
0.5EduMCTS80.8794.0344.7773.22
0.7MASTER70.3391.1733.0764.86
0.7CRPS77.0792.8341.3070.40
0.7EduMCTS80.9392.9346.5373.46

EduMCTS achieves the highest accuracy on all three benchmarks and the highest overall average under both top-p sampling settings, indicating that its performance advantage does not depend on greedy decoding.

AuthorsOpenReview ↗

Responses to Reviewer S1tF (2/2)

Q2. Why do the components in Table 3 have different effects across benchmarks? For example, removing search substantially hurts AIME but slightly improves MMLU-Pro Math. Is there an intuitive explanation, and are these differences stable across runs?

Q2.1. “Why do the components in Table 3 have different effects across benchmarks?”

We appreciate your insightful observation. Prior work suggests that search benefits depend on task difficulty and model capability, with limited gains on tasks where accuracy is already high [5, 6]. In EduMCTS, MCTS guides pedagogical intervention selection during data synthesis. Search-guided synthesis provides larger downstream gains on AIME, which requires complex open-ended reasoning, while offering smaller gains on MMLU-Pro Math, a multiple-choice benchmark.

Q2.2. “are these differences stable across runs?”

To examine stability, we conducted three independent evaluation runs of EduMCTS and w/o Search on AIME and MMLU-Pro Math using Qwen3-8B-Base under the original evaluation settings.

Table 2: Cross-Run Stability of Search Ablation (Accuracy %, Mean ± Std.)

BenchmarkMethodRun 1Run 2Run 3Mean ± Std.
AIMEw/o Search10.0013.3310.0011.11 ± 1.92
AIMEEduMCTS17.7820.0022.2220.00 ± 2.22
MMLU-Pro Mathw/o Search79.4279.2078.5379.05 ± 0.46
MMLU-Pro MathEduMCTS79.7178.6878.0978.83 ± 0.82

EduMCTS consistently outperforms w/o Search on AIME across all three runs, confirming a consistent performance advantage. On MMLU-Pro Math, the mean accuracy differs by only 0.22 percentage points, with no consistent ranking across runs. This indicates that the slight advantage of w/o Search on MMLU-Pro Math is not stable across evaluations.

Q3. Why does SFT on the original solutions sometimes perform substantially worse than the backbone without SFT? Have the authors investigated whether this results from the supervision itself or sensitivity to learning rate and training duration?

This is a very sharp and insightful question. We conducted additional experiments on the same Ori data using a reduced learning rate, fewer training epochs, full-parameter SFT, DoRA, and QLoRA to investigate the performance degradation.

Table 3: Sensitivity Analysis of Fine-Tuning on Original Trajectories (Qwen3-8B-Base)

MethodMATH-500GSM8KAMCAvg.
w/o SFT77.6092.3142.9170.94
Ori LoRA (Original)59.3085.8621.4055.52
Ori LoRA (Low LR)73.6789.3341.0368.01
Ori LoRA (2 Epochs)51.6084.0015.9350.51
Ori Full SFT39.7375.8710.4742.02
Ori DoRA47.1384.2714.4348.61
Ori QLoRA59.5383.2321.9054.89

All tested configurations, including full-parameter SFT, remain below the backbone without SFT, indicating that the performance degradation is not specific to LoRA or a particular training configuration.

We further used GPT-5 to assess all 1,080 Ori trajectories and found that 20.65% contain substantive reasoning defects, providing additional evidence of data quality issues that may contribute to the performance degradation. The complete hyperparameter experiments, GPT-5 quality analysis, and representative examples have been added to the revised Appendix D.1 (Table 14), D.2 (Table 15), and D.3.

References (for Parts 1/2 and 2/2)

[1] LLMs cannot find reasoning errors, but can correct them given the error location

[2] Large Language Models Cannot Self-Correct Reasoning Yet

[3] DART-Math: Difficulty-Aware Rejection Tuning for Mathematical Problem-Solving

[4] An Empirical Study of LLM Reasoning Ability Under Strict Output Length Constraint

[5] Tree of Thoughts: Deliberate Problem Solving with Large Language Models

[6] Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters

AuthorsOpenReview ↗

We sincerely request your valuable feedback

Dear Reviewer S1tF,

Thank you very much for your insightful and constructive comments! In our previous detailed responses, we have provided further clarifications and additional experimental evidence:

W1: We clarified the methodological distinction between EduMCTS and existing pedagogical and reasoning search methods. Specifically, EduMCTS jointly searches over pedagogical interventions and their induced student reasoning steps, enabling adaptive guidance based on intermediate reasoning states.

Q1: We conducted additional evaluations using top-p sampling at temperatures of 0.5 and 0.7, confirming that EduMCTS maintains its performance advantage under different decoding settings.

Q2: We explained the different effects of search across AIME and MMLU-Pro Math and conducted three independent evaluation runs, showing consistent gains on AIME and no stable ranking difference on MMLU-Pro Math.

Q3: We investigated the degradation caused by fine-tuning on original trajectories through additional training configurations and a GPT-5-based data quality analysis, with the findings included in the revised Appendix D.

We sincerely hope that these clarifications and additional results help address your concerns, particularly regarding the methodological contribution of EduMCTS. We look forward to your valuable feedback.

Best regards,

The Authors

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

This paper introduces EduMCTS, a training data synthesis method that organizes intermediate reasoning states into a search tree, formalizes pedagogical strategies, including HINT, CRITIQUE, and SOCRATIC, as executable search actions, and then uses MCTS to explore pedagogical actions and construct verified interactive reasoning trajectories for SFT. Experiments on different models and multiple benchmarks show that EduMCTS achieves better SFT performance than the compared baselines.

Critical strengths
  1. The proposed method is reasonable and effective. Compared with other data synthesis methods for SFT, EduMCTS achieves better performance on multiple benchmarks across different models.

  2. EduMCTS instantiates teaching as a concrete action space with different types of pedagogical strategies and integrates teacher feedback into the final trajectory for SFT training, which provides insights for future data synthesis methods.

Critical weaknesses
  1. The computational cost is not discussed. It is unclear whether EduMCTS still outperforms other data synthesis methods under a comparable computational budget.
Questions for authors
  1. Is it possible that the performance drop from the original trajectories is caused by LoRA training? Have the authors tried full-parameter SFT?

  2. Could the authors provide a computational cost comparison across different training data synthesis methods? For example, what are the number of total generated tokens and total GPU hours?

  3. Following my last question, if under a comparable computational budget, some baselines (e.g., Ori) can have more trajectories for SFT, will EduMCTS still show improvement?

5 comments
AuthorsOpenReview ↗

Responses to Reviewer WuMp (1/2)

Thank you for your thoughtful and helpful feedback. We will address your questions point by point below.

W1. The computational cost is not discussed.

Thank you for your careful review. In our original experiments, all baselines followed their respective default configurations and synthesis procedures. We measured the actual API calls and token consumption of EduMCTS, CRPS, and MASTER during data synthesis, as shown below.

Table 1: Synthesis Computational Cost Comparison across 1,080 Problems

MethodAPI CallsInput Tokens / QOutput Tokens / QTotal Tokens / QFull-run Tokens
MASTER10,8008,8566,09514,95116.15M
CRPS32,91627,99610,26138,25741.32M
EduMCTS89,71879,70833,224112,932121.97M

EduMCTS consumes approximately 2.95× and 7.55× as many tokens as CRPS and MASTER, respectively. Following your suggestion, we independently ran CRPS three times (3×3\times) and MASTER eight times (8×8\times) per seed problem to approximately match the synthesis token budgets. Following [1], we used GPT-4o to score candidate trajectories and selected the best trajectory through rejection sampling to construct the training sets. We then fine-tuned Qwen3-8B-Base using the same training configuration and evaluated it on three widely used reasoning benchmarks.

Table 2: Downstream Performance under Comparable Synthesis Budgets (Qwen3-8B-Base)

MethodMATH-500GSM8KAMCAvg.
MASTER (Original)72.7091.9336.9467.19
MASTER (Equal Budget)76.4793.0038.5669.34
CRPS (Original)78.0093.1442.1671.10
CRPS (Equal Budget)78.8793.3542.2971.50
EduMCTS82.2094.1348.1374.82

EduMCTS consistently outperforms the baselines under comparable synthesis budgets, further supporting its effectiveness when accounting for synthesis costs.

Q1. Is it possible that the performance drop from the original trajectories is caused by LoRA training? Have the authors tried full-parameter SFT?

Following your suggestion, we conducted five experiments on the same Ori data: full-parameter SFT, a reduced LoRA learning rate (1×10−51\times10^{-5}), fewer training epochs (2 epochs), DoRA, and QLoRA. The experimental configurations and results have been added to the revised Appendix D.1.

Table 3: Sensitivity Analysis of Fine-Tuning on Original Trajectories (Qwen3-8B-Base)

MethodMATH-500GSM8KAMCAvg.
w/o SFT77.6092.3142.9170.94
Ori LoRA (Original)59.3085.8621.4055.52
Ori LoRA (Low LR)73.6789.3341.0368.01
Ori LoRA (2 Epochs)51.6084.0015.9350.51
Ori Full SFT39.7375.8710.4742.02
Ori DoRA47.1384.2714.4348.61
Ori QLoRA59.5383.2321.9054.89

All tested fine-tuning configurations, including full-parameter SFT, perform below the backbone without SFT, indicating that the performance degradation is not specific to LoRA.

To further investigate the cause, we used GPT-5 to assess all 1,080 Ori training trajectories and found that 20.65% contain substantive reasoning errors or incomplete supervision. Detailed statistics and representative cases have been added to the revised Appendix D.2 and D.3. Prior work has shown that errors and noise in reasoning trajectories can introduce misleading supervision and degrade reasoning performance [2]. These findings suggest that reasoning quality issues in the original trajectories may be an important factor contributing to the observed degradation.

AuthorsOpenReview ↗

Responses to Reviewer WuMp (2/2)

Q2. Could the authors provide a computational cost comparison across different training data synthesis methods?

Thank you for your suggestion. All methods in our experiments synthesize data through a remotely hosted Qwen3-235B API. As the actual GPU usage of the API server is unavailable, we compare computational costs using measured API calls and input/output token consumption, as reported in W1 (Table 1). The total token consumption is 121.97M for EduMCTS, 16.15M for MASTER, and 41.32M for CRPS.

Q3. If under a comparable computational budget, some baselines can have more trajectories for SFT, will EduMCTS still show improvement?

Thank you for this insightful question. In W1 (Table 2), we compare EduMCTS with strong baselines under approximately matched synthesis budgets. EduMCTS achieves the best performance under this setting.

To further examine whether baselines benefit from more training trajectories, we conducted additional experiments with expanded training sets. Since Ori trajectories are directly sampled from the original dataset without additional API costs, we constructed 4× and 8× Ori training sets from NuminaMath-1.5. For MASTER (8×) and CRPS (3×), we used all trajectories generated under the expanded synthesis budgets for SFT, without rejection sampling. All experiments used Qwen3-8B-Base.

Table 4: Downstream Performance with Expanded Baseline Training Data (Qwen3-8B-Base)

MethodMATH-500GSM8KAMCAvg.
Ori (Original)59.3085.8621.4055.52
Ori (4× Data)37.7382.8013.2044.58
Ori (8× Data)39.9380.1011.9744.00
MASTER (8× Data)74.0090.7537.8167.52
CRPS (3× Data)75.6093.0042.0770.22
EduMCTS82.2094.1348.1374.82

EduMCTS maintains the best performance across all three benchmarks, even when the baselines use substantially more training trajectories.

References (for Parts 1/2 and 2/2)

[1] Crowd Comparative Reasoning: Unlocking Comprehensive Evaluations for LLM-as-a-Judge

[2] Understanding the Effect of Noise in LLM Training Data with Algorithmic Chains of Thought

AuthorsOpenReview ↗

We sincerely request your valuable feedback

Dear Reviewer WuMp,

Thank you very much for your thoughtful and constructive comments! Following your suggestions, we have provided detailed responses and conducted several additional experiments:

W1 & Q2: We reported the API calls and token consumption of EduMCTS and the main baselines, and conducted additional experiments under approximately matched synthesis budgets, showing that EduMCTS maintains its performance advantage.

Q1: We investigated the performance degradation on original trajectories through full-parameter SFT, different learning rates, training durations, DoRA, and QLoRA. We further examined all 1,080 original trajectories using GPT-5, identifying reasoning quality issues that may contribute to the degradation.

Q3: We expanded the baseline training sets, including 4×/8× Ori data, 8× MASTER data, and 3× CRPS data. The results show that EduMCTS continues to outperform these baselines despite their larger training sets.

We sincerely hope that these additional experiments and analyses help address your questions. We look forward to your valuable feedback.

Best regards,

The Authors

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

This paper proposes EduMCTS, a reasoning data synthesis framework that integrates pedagogical interventions into Monte Carlo Tree Search. It introduces four actions (Self, Hint, Critique, and Socratic) and combines process-level progress evaluation with outcome rewards to construct verified reasoning trajectories for SFT. Experiments on three backbone models across eleven reasoning benchmarks demonstrate consistent improvements over existing data augmentation methods.

Critical strengths
  1. Interesting formulation. Integrating pedagogical interventions into MCTS provides an intuitive approach to constructing reasoning supervision.
  2. Consistent empirical improvements. The method outperforms several representative data augmentation baselines across three models and eleven benchmarks, supported by component ablations.
Critical weaknesses

W1. Unclear contribution of pedagogical supervision.

The w/o Teach ablation changes both trajectory generation and supervision, leaving the contribution of pedagogical feedback unclear. It remains uncertain whether the gains stem from pedagogical interaction or simply from the resulting high-quality reasoning trajectories. A comparison using the same verified trajectories with teacher reflections removed would help clarify this distinction.

W2. Insufficient analysis of synthesis efficiency.

EduMCTS relies on a 235B model and multiple MCTS iterations. Although baselines share the synthesis backbone, their computational budgets are not explicitly matched. Comparisons of model calls, token costs, and performance under equal budgets would help justify the additional overhead relative to CRPS.

W3. Limited validation of LLM-based evaluation.

Both process rewards and terminal verification rely on LLM judgments. Despite access to reference answers, their reliability is not independently assessed. Since these judgments guide search and training data selection, further validation of their accuracy would strengthen the results.

Questions for authors

Q1. How does pedagogical supervision differ from existing reasoning paradigms, such as CoT, self-reflection, self-correction, and self-questioning? Could the authors further isolate its contribution by removing teacher feedback from the same verified trajectories while retaining the student reasoning steps?

Q2. Could the authors report the synthesis costs of EduMCTS and the main baselines, particularly MASTER and CRPS, and discuss whether the improvements persist under comparable computational budgets?

4 comments
AuthorsOpenReview ↗

Responses to Reviewer oCxS (1/2)

Thank you for your detailed reviews. We will address your questions point by point below.

W1. Unclear contribution of pedagogical supervision.

Pedagogical interaction in EduMCTS serves two complementary purposes: guiding the generation of high-quality student reasoning steps during data synthesis and providing explicit pedagogical supervision by retaining teacher feedback during SFT. Following your suggestion, we conducted an additional ablation (w/o Reflection) by removing teacher feedback from EduMCTS-synthesized data. We fine-tuned Qwen3-8B and Llama3.1-8B and evaluated them on three widely used reasoning benchmarks.

Table 1: Controlled Ablation on Pedagogical Supervision

MethodMATH-500GSM8KAMCAvg.
Full EduMCTS (Qwen3)82.2094.1348.1374.82
w/o Teach (Qwen3)79.7092.7647.7673.41
w/o Reflection (Qwen3)81.1392.8046.2773.40
Full EduMCTS (Llama3.1)56.2086.6618.6653.84
w/o Teach (Llama3.1)54.6085.2017.4152.40
w/o Reflection (Llama3.1)53.7384.2316.1751.38

The results show that removing teacher feedback reduces performance across all three benchmarks for both models. These findings indicate that EduMCTS benefits not only from the high-quality student reasoning steps generated through pedagogical interaction but also from the additional supervision provided by explicitly retaining teacher feedback during SFT.

W2. Insufficient analysis of synthesis efficiency.

Thank you for your careful review. To ensure objective comparisons in our original experiments, all baselines followed their default configurations and data synthesis procedures. We measured the actual LLM API calls and token consumption of EduMCTS and two strong baselines, CRPS and MASTER, as shown in Table 2.

Table 2: Synthesis Computational Cost Comparison across 1,080 Problems

MethodAPI CallsInput Tokens / QOutput Tokens / QTotal Tokens / QFull-run Tokens
MASTER10,8008,8566,09514,95116.15M
CRPS32,91627,99610,26138,25741.32M
EduMCTS89,71879,70833,224112,932121.97M

The token consumption of EduMCTS is approximately 2.95× that of CRPS and 7.55× that of MASTER. Following your suggestion, we further conducted comparative experiments under approximately matched synthesis token budgets. Specifically, we independently ran CRPS three times and MASTER eight times for each seed problem. Following the method in [1], we used GPT-4o for rejection sampling to select the best trajectory for each method. The results show that EduMCTS maintains a clear performance advantage after budget matching.

Table 3: Downstream Performance Comparison under Matched Computational Budgets (Qwen3-8B-Base)

MethodMATH-500GSM8KAMCAvg.
MASTER (Original)72.7091.9336.9467.19
MASTER (Equal Budget)76.4793.0038.5669.34
CRPS (Original)78.0093.1442.1671.10
CRPS (Equal Budget)78.8793.3542.2971.50
EduMCTS82.2094.1348.1374.82

W3. Limited validation of LLM-based evaluation.

Thank you for your careful review. Prior work has employed LLM-based evaluators to guide reasoning search and data synthesis [2, 3]. Similarly, EduMCTS uses LLM-as-a-Judge for local reasoning progress scoring and terminal correctness verification.

To validate LLM-based evaluators, we randomly sampled 200 complete trajectories and 200 intermediate reasoning transitions. Two mathematics PhD students independently annotated these samples with model scores concealed, and a third adjudicated unresolved disagreements to establish the human gold standard. We also compared against GPT-5 and Gemini-3.1-Flash-Lite. Annotation procedures, criteria, and inter-annotator agreement are detailed in the revised Appendix E.1–E.3.

Table 4: Validation of Evaluators against Human Expert Gold Standard (%)

ModelVϕV_\phi Acc.VϕV_\phi Prec.VϕV_\phi Rec.EϕE_\phi ExactEϕE_\phi Polarity
Qwen3-235B (Ours)94.5098.7794.6778.5094.50
Gemini-3.1-Flash-Lite93.5094.8397.6381.0093.00
GPT-586.0088.9595.2781.0093.00

For terminal verification (VϕV_\phi), Acc., Prec., and Rec. denote accuracy, precision, and recall. For process evaluation (EϕE_\phi), Exact measures agreement on the five-level scale, while Polarity measures agreement on positive, neutral, and negative progress.

Qwen3-235B achieves 94.50% terminal verification accuracy and high process polarity agreement. These results support the reliability of our evaluators, with further analysis provided in the revised Appendix E.4.

AuthorsOpenReview ↗

Responses to Reviewer oCxS (2/2)

Q1. How does pedagogical supervision differ from existing reasoning paradigms, and can its contribution be isolated by removing teacher feedback while retaining the same student reasoning steps?

Q1.1. How does pedagogical supervision differ from existing reasoning paradigms?

Thank you for your question. We further clarify the distinctions between EduMCTS and existing reasoning paradigms:

  1. Difference from CoT: CoT typically generates reasoning steps along a single sequential path, where early errors can propagate through subsequent steps [4]. EduMCTS explores alternative reasoning branches through tree search and uses local reasoning progress and terminal correctness feedback to select more effective reasoning paths.

  2. Difference from Self-reflection and Self-correction: These methods primarily improve reasoning through self-examination and error correction, yet unreliable feedback can hinder error identification and even degrade performance [5]. EduMCTS not only generates pedagogical feedback but also evaluates its actual guiding effect through local reasoning progress and terminal correctness, constructing training trajectories that combine effective pedagogical interventions with high-quality subsequent reasoning steps.

  3. Difference from Self-questioning: Self-questioning advances reasoning through self-generated questions, while automatically generated guiding questions may be irrelevant or poorly timed [6]. EduMCTS extends beyond questioning alone, incorporating Socratic questioning, hints, and critiques into a unified search space and adaptively selecting more effective pedagogical strategies through tree search to construct training trajectories.

Overall, EduMCTS's core innovation is to formulate pedagogical interventions as searchable and adaptively selectable decision actions. By incorporating effective teacher feedback alongside student reasoning into SFT, the model learns both pedagogical reasoning strategies and high-quality reasoning processes.

Q1.2. Can its contribution be isolated by removing teacher feedback while retaining the same student reasoning steps?:

As detailed in W1 (Table 1), we conducted an additional ablation that removes all teacher feedback while preserving the same student reasoning steps. Removing teacher feedback reduces performance across all three benchmarks for both Qwen3-8B and Llama3.1-8B, indicating that explicit pedagogical feedback provides additional supervisory value.

Q2. Could the authors report the synthesis costs of EduMCTS and the main baselines, and discuss whether the improvements persist under comparable computational budgets?

As detailed in W2 (Tables 2–3), we report the API calls and token consumption of EduMCTS, MASTER, and CRPS under a consistent measurement protocol, along with additional performance comparisons under matched synthesis budgets. EduMCTS maintains its advantage across multiple reasoning benchmarks under comparable computational budgets, further supporting the effectiveness of the proposed method.

References (for Parts 1/2 and 2/2)

[1] Crowd Comparative Reasoning: Unlocking Comprehensive Evaluations for LLM-as-a-Judge

[2] rStar-Math: Small LLMs Can Master Math Reasoning with Self-Evolved Deep Thinking

[3] ReST-MCTS*: LLM Self-Training via Process Reward Guided Tree Search

[4] The Art of SOCRATIC QUESTIONING: Recursive Thinking with Large Language Models

[5] When Can LLMs Actually Correct Their Own Mistakes? A Critical Survey of Self-Correction of LLMs

[6] Improving Socratic Question Generation using Data Augmentation and Preference Optimization

AuthorsOpenReview ↗

We sincerely request your valuable feedback

Dear Reviewer oCxS,

Thank you for taking the time to review our paper and for raising these important questions. In our previous responses, we have provided detailed clarifications and additional experimental evidence addressing your concerns:

W1 & Q1: We clarified the distinctions between EduMCTS and existing reasoning paradigms and introduced the w/o Reflection ablation, demonstrating that explicit pedagogical supervision provides additional benefits beyond the generated student reasoning steps.

W2 & Q2: We reported detailed API calls and token consumption and conducted additional comparisons under approximately matched synthesis budgets, showing that EduMCTS maintains its performance advantage.

W3: We conducted a human evaluation involving three mathematics PhD students and compared our evaluators with GPT-5 and Gemini-3.1-Flash-Lite, providing additional evidence for the reliability of our evaluation mechanisms.

We hope these additional experiments and clarifications help address your concerns. We look forward to your valuable feedback.

Best regards,

The Authors

AuthorsOpenReview ↗

Second Follow-up

Dear Reviewer oCxS,

We hope you are doing well. As the author–reviewer discussion deadline approaches, we would like to share three additional experiments and analyses addressing your concerns.

Your review focuses on pedagogical supervision (W1/Q1), synthesis efficiency (W2/Q2), and LLM-based evaluation (W3). Our previous rebuttal provided the w/o Reflection ablation, API call and token comparisons under matched synthesis budgets, and human evaluation by three mathematics PhD students. The following experiments provide further evidence.

(1) Contribution of Pedagogical Supervision (W1 & Q1)

As clarified in Q1, EduMCTS formulates pedagogical interventions as searchable and adaptively selectable actions, unlike CoT, self-reflection, self-correction, and self-questioning. We further conducted a Mask Teacher Loss ablation, retaining the same EduMCTS trajectories while masking the loss on teacher feedback tokens. Results on Qwen3-8B-Base and Llama3.1-8B-Instruct are averaged over three independent evaluation runs.

Table 1: Additional ablation of pedagogical supervision.

MethodMATH-500GSM8KAMCAvg.
Full EduMCTS (Qwen3)82.2094.1348.1374.82
Mask Loss (Qwen3)81.4793.2344.2872.99
Full EduMCTS (Llama3.1)56.2086.6618.6653.84
Mask Loss (Llama3.1)54.9384.4716.9052.10

Masking teacher feedback supervision reduces performance across all three benchmarks for both models. Together with w/o Reflection, this supports the contribution of explicit pedagogical supervision beyond high-quality student reasoning steps.

(2) Data Efficiency with Reduced Training Data (W2 & Q2)

Following our matched-budget comparisons, we further fine-tuned Qwen3-8B-Base using only 50% and 75% of the EduMCTS training data to examine data efficiency.

Table 2: Downstream performance with different training data sizes (Qwen3-8B-Base).

MethodSamplesMATH-500GSM8KAMCAvg.
MASTER (Equal Budget)1,08076.4793.0038.5669.34
CRPS (Equal Budget)1,08078.8793.3542.2971.50
EduMCTS (50%)54080.7393.1745.3073.07
EduMCTS (75%)81081.2093.7745.8073.59
EduMCTS (100%)1,08082.2094.1348.1374.82

Encouragingly, even with only 50% of the training data, EduMCTS outperforms MASTER and CRPS with expanded synthesis budgets and rejection sampling in average accuracy. These results further support the effectiveness and data efficiency of EduMCTS-generated trajectories.

(3) Predictive Validity of Process Evaluation (W3)

Our previous human evaluation validated the agreement between process scores and expert judgments. We further examined whether higher process scores are associated with subsequent rollout success during search.

We analyzed 3,284 non-terminal rollout records from 200 sampled problems. For each rollout, we computed the normalized average process score along the preceding path and recorded the subsequent terminal verification outcome. We fitted logistic regression models controlling for search depth and reported the odds ratio (OR) per 0.5 increase in process score, with 95% confidence intervals from 2,000 problem-level bootstrap resamples.

Table 3: Association between path-level process scores and subsequent rollout success.

Search DepthRolloutsOR (+0.5 Score)95% CI
Depth 27212.09[1.43, 3.25]
Depth 32,4113.50[2.30, 5.76]
Depth 41521.68[0.78, 5.50]
Depth-adjusted Overall3,2842.90[1.96, 4.57]

Higher process scores are associated with subsequent rollout success (depth-adjusted OR: 2.90, 95% CI: [1.96, 4.57]). This complements our human evaluation by providing further evidence of the predictive validity of process scores during search.


Your Strengths section recognizes "Integrating pedagogical interventions into MCTS provides an intuitive approach to constructing reasoning supervision" and "Consistent empirical improvements." Reviewer WuMp noted that "The proposed method is reasonable and effective" and highlighted the pedagogical action space. Reviewer S1tF noted that "Component ablations provide additional evidence for the usefulness of the proposed design."

These new analyses further support pedagogical supervision, data efficiency, and process evaluation, directly addressing your concerns.

We hope these additional results help address your concerns. We would greatly appreciate your feedback after considering our rebuttal and these new analyses. Even a brief reply would help us provide further clarification during the remaining discussion period.

Thank you again for your time and consideration.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.