- 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
-
The framework is clearly presented, with detailed prompts and implementation settings that make the data construction procedure understandable.
-
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
- 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
-
Why was evaluation restricted to greedy decoding? Do the relative improvements remain under top-p sampling?
-
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?
-
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
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.)
| Benchmark | Method | Run 1 | Run 2 | Run 3 | Mean ± Std. |
|---|---|---|---|---|---|
| AIME | w/o Search | 10.00 | 13.33 | 10.00 | 11.11 ± 1.92 |
| AIME | EduMCTS | 17.78 | 20.00 | 22.22 | 20.00 ± 2.22 |
| MMLU-Pro Math | w/o Search | 79.42 | 79.20 | 78.53 | 79.05 ± 0.46 |
| MMLU-Pro Math | EduMCTS | 79.71 | 78.68 | 78.09 | 78.83 ± 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)
| Method | MATH-500 | GSM8K | AMC | Avg. |
|---|---|---|---|---|
| w/o SFT | 77.60 | 92.31 | 42.91 | 70.94 |
| Ori LoRA (Original) | 59.30 | 85.86 | 21.40 | 55.52 |
| Ori LoRA (Low LR) | 73.67 | 89.33 | 41.03 | 68.01 |
| Ori LoRA (2 Epochs) | 51.60 | 84.00 | 15.93 | 50.51 |
| Ori Full SFT | 39.73 | 75.87 | 10.47 | 42.02 |
| Ori DoRA | 47.13 | 84.27 | 14.43 | 48.61 |
| Ori QLoRA | 59.53 | 83.23 | 21.90 | 54.89 |
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
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
Responses to Reviewer S1tF (1/2)
Thank you for your thoughtful and helpful feedback. We will address your questions point by point below.
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.
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)
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.