acceptodds
Under review as a conference paper at ICLR 2027

From Traces to Recipes: Constructing and Distilling Reusable Procedural Knowledge into Language Models

Abstract

Reasoning distillation trains small students on teacher solution traces, but each trace entangles a reusable procedure with the details of one problem, leaving the student to infer what is shared. Existing supervision thus trades reusability against verification: traces solve a concrete problem but are tied to it, whereas abstractions generalize across problems but are never tested on the model that must use them. We introduce DistilRecipe, which represents procedural knowledge as recipes: instance-independent procedures that state when a method applies, which operations to perform in dependency order, and how to check the result. A recipe is admitted only when a small frozen student following it answers correctly and a monitor confirms it applied and was followed. In context, a receipe bank built with one 2B anchor improves 0.8B–4B Qwen3.5 models in all 30 settings, by an average of 9.0% on MATH-500 and 5.6% on GSM8K. Used as answer-free privileged context, recipes also distill into a Qwen3.5-2B student that needs no recipe at test time and outperforms matched chain-of-thought distillation in all four comparisons, by 2.08–3.20%.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Knowledge Distillation, Reusable Reasoning Procedures, On-Policy Distillation

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_WwUvOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

DistilRecipe constructs reusable recipes containing applicability conditions, procedures, and checks, validated through a frozen student’s execution. A larger-model router selects a recipe for each problem, and recipe revisions must preserve success on previously supported problems. Recipes improve inference-time performance and are distilled into students that require no recipes at deployment. Experiments on GSM8K and MATH-500 show gains over matched recipe-free teacher–student distillation under SFT and OPD, although self-distillation results are mixed.

Critical strengths
  1. Execution-based recipe construction. Student execution, adherence checks, and replay of previously supported problems provide a concrete mechanism for maintaining reusable guidance.
  2. Useful distillation results. Matched teacher–student comparisons show consistent gains under both SFT and OPD without requiring recipes at deployment.
Critical weaknesses
  1. Recipe selection is insufficiently evaluated. Lexical retrieval may miss suitable recipes when the required procedure is not apparent from the problem’s wording. The paper does not distinguish coverage, retrieval, and routing failures or quantify their performance impact.
  2. The incremental contribution remains unclear. Comparisons with reference-conditioned OPSD and experiential-context OPCD are missing. Ablations of recipe fields, adherence verification, and historical replay are also absent, leaving unclear whether the proposed complexity is necessary.
Questions for authors
  1. How do oracle or best-of-bank recipe selection results compare with the current router, for both prompting and distillation?
  2. How does recipe conditioning compare with reference-solution-conditioned OPSD and experiential-context-conditioned OPCD under comparable conditions?
  3. What are the contributions of applicability, procedures, and checks? Do adherence verification and historical replay improve over answer-correctness filtering alone?
3 comments

Response to Reviewer WwUv (1/2)

We thank the reviewer for recognizing execution-based recipe construction and the matched distillation gains under both SFT and OPD, without recipes at deployment. We address the concerns below.

W1 / Q1. Recipe selection: separating coverage, retrieval and routing errors

We ran every shortlisted recipe, then every recipe in the bank, on every test problem (Qwen-2B, 2,048 tokens). A recipe counts as applicable only if the student, given it, follows it to the correct answer.

Recipe-freeRouterShortlist oracleFull-bank oracle
GSM8K71.5777.7982.2186.18
MATH-50059.6069.0075.4079.80

The gaps between columns attribute the router's errors to three causes (points; share of errors):

Cause of a router errorMeasured asGSM8KMATH-500
Routing: an applicable recipe was shortlisted but not chosenShortlist oracle − router4.4 (20%)6.4 (21%)
Retrieval: an applicable recipe exists but was not shortlistedFull-bank oracle − shortlist oracle4.0 (18%)4.4 (14%)
Coverage: no recipe in the bank is applicable100 − full-bank oracle13.8 (62%)20.2 (65%)

Lexical retrieval therefore accounts for only 14–18% of errors, while coverage accounts for about two thirds, which is why gains grow with the bank (Figures 6–7). Routing and retrieval together leave 8.4 points on GSM8K and 10.8 on MATH-500 that are already in the bank and recoverable by a better selector without authoring new recipes.

W2 / Q2. Comparison with OPSD and OPCD

All three methods use the same on-policy objective and differ only in the teacher's context (§3.5): the reference solution (OPSD [1]), one experiential context shared by all problems (OPCD [2]), or the recipe routed from the question (ours). We use the same 9B teacher, 2B student, problems and schedule, with three seeds:

Teacher's privileged contextMATH-500GSM8K
CoT63.53 ± 0.6480.07 ± 0.78
Reference solution (OPSD [1])64.46 ± 0.4280.44 ± 0.59
One shared experiential context (OPCD [2])64.53 ± 0.9079.26 ± 0.52
Distillrecipe (ours)66.33 ± 0.5883.47 ± 1.01

Distillrecipe gives a consistent gain: +1.8 to +4.2 points over OPSD and OPCD.

Response to Reviewer WwUv (2/2)

W2 / Q3. Ablation of recipe fields: applicability, procedure and checks

The three fields act at different points of the pipeline (§3.2, App. A). Applicability states which problems a procedure fits; it is read by the router, whose profiles contain applicability and procedure but not checks (App. A.2), and by the student, who is told to use a recipe only when its applicability matches (App. B.3). The procedure is the ordered list of operations the student executes and the monitor audits (App. B.4). The checks are post-conditions applied at the end of an execution, such as reconstructing a total (App. E.3) or confirming that the answer covers the whole domain (App. E.4).

We removed each field from the routed recipe at inference and, separately, removed applicability from routing only while still showing the full recipe (Qwen 2B recipient, 2,048 tokens; change from the full recipe in parentheses):

recipe shown to the studentMATH-500GSM8K
Full recipe69.0077.79
Without procedure60.40 (−8.6)70.62 (−7.2)
Without applicability63.60 (−5.4)72.56 (−5.2)
Applicability removed from routing only62.60 (−6.4)71.92 (−5.9)
Without checks66.80 (−2.2)75.87 (−1.9)
No recipe59.6071.57

Each field matters in the way its role predicts:

  • Procedure. Without it, the recipe loses its entire value: accuracy falls to within one point of using no recipe.
  • Applicability. Without it, the student still has the procedure but applies it where it does not fit, and loses more than half of the gain. Removing it from routing alone costs at least as much, so it is also the signal the router relies on.
  • Checks. They add a smaller but consistent gain, slightly larger on MATH-500, where the final form of the answer is part of its correctness.

W2 / Q3. Ablation of adherence verification and historical replay

We rebuilt the MATH bank with the same author, authoring problems and router, changing only the acceptance rule ( MATH-500, 2,048 tokens):

Acceptance ruleRecipes0.8B2BMean gain
No recipe–38.0059.60–
No verification (author's recipes as written)17842.2064.40+4.5
Monitor judgment only (no student execution)7342.4066.00+5.4
Final-answer correctness only (no monitor)8043.4065.40+5.6
Full e·α·h (ours)6845.6069.00+8.5

Adherence verification clearly improves over answer-correctness filtering: it adds 2.9 points on the mean (2.2 for the 0.8B recipient, 3.6 for the 2B), with a smaller bank. Answer-only filtering is barely better than no verification at all, and the construction logs explain why: 26.7% of the executions that reached the correct answer did not follow the recipe, so an answer-only rule accepts them as evidence for a procedure the student never used. Judging applicability without execution is no better, because a recipe can read as applicable and still be one the student cannot carry out. Only the combination identifies procedures the student can execute.

Historical replay protects the evidence the bank is built on. A revision is committed only if it still solves every problem its predecessor solved, so the support recorded for a recipe stays true after every edit (Prop. 1). Without replay, a revision that fixes its triggering problem can silently break earlier ones. The bank would keep recipes whose recorded support no longer holds. This happens often in practice: during MATH construction, 496 replays on inherited supports failed, each one a previously solved problem that a revision passing its own triggering problem would have broken.

We thank the reviewer again for the constructive feedback, which has strengthened the paper. We hope our response addresses the concerns, and we welcome any further questions.

References

[1] Zhao, S., Xie, Z., Liu, M., Huang, J., Pang, G., Chen, F., and Grover, A. Self-distilled reasoner: On-policy self-distillation for large language models. In ICML, 2026.

[2] Ye, T., Dong, L., Wu, X., Huang, S., and Wei, F. On-policy context distillation for language models. arXiv:2602.12275, 2026.

AuthorsOpenReview ↗

Follow-up before the discussion deadline

Dear Reviewer WwUv,

We truly appreciate the time and effort you've already dedicated to reviewing our submission. We prepared our rebuttal with great care in hopes of directly addressing the valuable points raised in your review, including an oracle analysis of recipe selection, comparisons with OPSD and OPCD, ablations of the recipe fields, and adherence verification.

If these results address your concerns, we would be grateful if you would consider raising your score. We would also welcome any further thoughts or clarifications you might wish to share.

Thank you again for your contribution to the review process.

Best regards,

The Authors

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

This paper introduces DistilRecipe, a framework for transferring recipes (reusable procedural knowledge) from large to smaller language models. A recipe is admitted only when a small frozen student following it answers correctly and a monitor confirms it applied and was followed. In context, a recipe bank built with one 2B anchor improves 0.8B--4B Qwen3.5 models in all 30 settings. Used as answer-free privileged context, recipes also distill into a Qwen3.5-2B student that needs no recipe at test time and outperforms matched chain-of-thought distillation in all four comparisons, by 2.08--3.20%.

Critical strengths
  1. The idea of recipe abstraction before distillation is intuitive and well-motivated.
  2. Student-grounded verification is a meaningful contribution. Rather than accepting an abstraction merely because a stronger model produces or judges it, the method requires the frozen target student to successfully use it to solve problems. This addresses potential issues in existing distillation methods: the produced abstractions may contain errors or are too far away from the student's solution space.
Critical weaknesses
  1. My main concern is lack of strong baselines. The paper acknowledges several reusable abstraction and reasoning trace distillation approaches, as long as training-based distillation algorithms, but the experiments only compare DistilRecipe against recipe-free inference, random recipes, and CoT distillation. Without stronger baseline comparisons, it's difficult to estimate the effectiveness of the proposed approach.
  2. The empirical scope is fairly narrow relative to what's claimed in the introduction. All experiments are done on mathematical reasoning on only GSM8K and MATH-500, using models exclusively from the Qwen3.5 family. So it is unclear whether recipes capture broadly reusable procedural knowledge or whether the approach is particularly effective for math.

Post rebuttal comment: The concern about lack of strong baselines has been partially addressed after the rebuttal, although I still believe that RLAD would be a strong baseline and should have been included as well. I gave an overall score of 3 but I'm lean toward 2.5.

Questions for authors

See the weaknesses.

5 comments

Response to Reviewer Tcvp (1/2)

We thank the reviewer for recognizing recipe abstraction as well motivated and student-grounded verification as a meaningful contribution. We address both concerns in detail below.

W1. Stronger baselines

Our original comparisons were controlled ablations, each isolating one factor (recipe vs. none, routed vs. random, recipe vs. CoT distillation). To also position DistilRecipe against prior methods, we added baselines in both settings, under our protocol (same models, problems, decoding and scorer).

In-context. Running prior reusable-abstraction methods end to end would confound how procedures are admitted with how they are produced and used. For example, [1] curates behaviors with a 70B model for 32B–70B solvers and retrieves them by MATH topic labels, which our router never sees, and RLAD [2] trains its abstraction generator with reinforcement learning. We therefore implement each method's admission rule inside our pipeline, with the same 27B author, 1,000 authoring problems, router and recipients, so differences reflect the admission rule alone, which is where DistilRecipe departs from these methods (MATH-500, 2,048 tokens):

Acceptance ruleClosest prior regimeRecipes0.8B2BMean gain
No recipe––38.0059.60–
No verification (author's recipe as written)Behavior handbooks, thought templates [1, 3]17842.2064.40+4.5
Monitor judgment only (no student execution)Judge-rated guides [4]7342.4066.00+5.4
Final-answer correctness only (no monitor)RLAD [2]8043.4065.40+5.6
Full e·α·h (ours)DistilRecipe6845.6069.00+8.5

DistilRecipe gives the largest gain with the fewest recipes, nearly double the unverified bank. Answer-only or judge-only filtering adds about one point; requiring the student both to reach the correct answer and to follow the procedure adds four. This supports the reviewer's point that abstractions accepted without the student can be wrong or lie outside its solution space.

Training-based distillation. OPSD [5] and OPCD [6] share our on-policy objective, so we compare against them directly: same 9B teacher, 2B student, problems and schedule, three seeds, and only the teacher's privileged context differs:

Teacher's privileged contextMATH-500GSM8K
CoT63.53 ± 0.6480.07 ± 0.78
Reference solution (OPSD [5])64.46 ± 0.4280.44 ± 0.59
One shared experiential context (OPCD [6])64.53 ± 0.9079.26 ± 0.52
Distillrecipe (ours)66.33 ± 0.5883.47 ± 1.01

Only the recipe gives a consistent gain: +1.8 to +4.2 points over OPSD and OPCD.

Response to Reviewer Tcvp (2/2)

W2. Empirical scope: model families and new distributions

GSM8K and MATH are the standard benchmarks for mathematical reasoning in recent distillation work (e.g., [7–10]), and our MATH protocol (a bank built from 1,000 training problems, evaluated on MATH-500) follows the closest prior work on reusable reasoning behaviors [1]. Within this scope, the evidence is broader than two benchmarks suggest:

Different procedure types. GSM8K recipes are arithmetic schemas over story quantities, while MATH spans competition problems in seven subjects that need symbolic methods such as domain restriction and modular search (App. E). One bank built with a 2B anchor improves three model sizes in all 30 settings.

Domain-agnostic design. Recipes, routing, monitoring and replay operate over abstract roles (§3.2, App. A.1); only the answer evaluator and prompt wording are domain-specific. A new domain needs only a new bank, and gains grow with the authoring budget (Figures 6–7).

Beyond Qwen. The banks were authored and verified with Qwen models, yet applied unchanged they improve three other model families in all eight settings, by up to +12.0 points:

RecipientMATH-500: no recipe → recipeGSM8K: no recipe → recipe
Gemma-3-1B-it39.00 → 51.00 (+12.0)48.37 → 56.94 (+8.6)
Gemma-3-4B-it70.20 → 78.40 (+8.2)87.49 → 89.61 (+2.1)
Phi-4-mini-instruct58.80 → 65.00 (+6.2)88.17 → 92.65 (+4.5)
Llama-3.2-3B-instruct44.40 → 49.00 (+4.6)78.47 → 80.89 (+2.4)

Beyond the authoring distribution. Without new authoring, the GSM8K bank also improves accuracy on SVAMP [11], GSM-Symbolic [12] and GSM-Hard [13] in all nine settings, by +0.5 to +11.3 points:

Dataset (n)RecipientNo recipeRecipeGain
SVAMP (300)0.8B64.3369.33+5.0
2B77.3388.67+11.3
4B93.6795.67+2.0
GSM-Symbolic (1,000)0.8B46.3051.80+5.5
2B71.5072.00+0.5
4B90.0094.40+4.4
GSM-Hard (1,319)0.8B27.4531.01+3.6
2B39.6546.63+7.0
4B60.2767.55+7.3

We thank the reviewer again for the constructive feedback, which has strengthened the paper. We hope our response addresses the concerns, and we welcome any further questions.

References

[1] Didolkar, A., Ballas, N., Arora, S., and Goyal, A. Metacognitive reuse: Turning recurring LLM reasoning into concise behaviors. arXiv:2509.13237, 2025.

[2] Qu, Y., Singh, A., Lee, Y., Setlur, A., Salakhutdinov, R., Finn, C., and Kumar, A. RLAD: Training LLMs to discover abstractions for solving reasoning problems. In ICLR, 2026.

[3] Yang, L., Yu, Z., Zhang, T., Cao, S., Xu, M., Zhang, W., Gonzalez, J. E., and Cui, B. Buffer of thoughts: Thought-augmented reasoning with large language models. In NeurIPS, 2024.

[4] Alkiek, K., Jurgens, D., and Vydiswaran, V. G. V. Big reasoning with small models: Instruction retrieval at inference time. arXiv:2510.13935, 2025.

[5] Zhao, S., Xie, Z., Liu, M., Huang, J., Pang, G., Chen, F., and Grover, A. Self-distilled reasoner: On-policy self-distillation for large language models. In ICML, 2026.

[6] Ye, T., Dong, L., Wu, X., Huang, S., and Wei, F. On-policy context distillation for language models. arXiv:2602.12275, 2026.

[7] Agarwal, R., Vieillard, N., Zhou, Y., Stanczyk, P., Ramos Garea, S., Geist, M., and Bachem, O. On-policy distillation of language models: Learning from self-generated mistakes. In ICLR, 2024.

[8] Xu, W., Han, R., Wang, Z., Le, L. T., Madeka, D., Li, L., Wang, W. Y., Agarwal, R., Lee, C.-Y., and Pfister, T. Speculative knowledge distillation: Bridging the teacher-student gap through interleaved sampling. In ICLR, 2025.

[9] Li, Y., Yue, X., Xu, Z., Jiang, F., Niu, L., Lin, B. Y., Ramasubramanian, B., and Poovendran, R. Small models struggle to learn from strong reasoners. arXiv:2502.12143, 2025.

[10] DeepSeek-AI. DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning. arXiv:2501.12948, 2025.

[11] Patel, A., Bhattamishra, S., and Goyal, N. Are NLP models really able to solve simple math word problems? In NAACL, 2021.

[12] Mirzadeh, I., Alizadeh, K., Shahrokhi, H., Tuzel, O., Bengio, S., and Farajtabar, M. GSM-Symbolic: Understanding the limitations of mathematical reasoning in large language models. In ICLR, 2025.

[13] Gao, L., Madaan, A., Zhou, S., Alon, U., Liu, P., Yang, Y., Callan, J., and Neubig, G. PAL: Program-aided language models. In ICML, 2023.

AuthorsOpenReview ↗

Follow-up before the discussion deadline

Dear Reviewer Tcvp,

We truly appreciate the time and effort you've already dedicated to reviewing our submission. We prepared our rebuttal carefully to directly address the valuable points raised in your review, including comparisons with stronger baselines, comparisons with OPSD and OPCD, and transfer to other model families and new distributions.

If these results address your concerns, we would be grateful if you would consider raising your score. We would also welcome any further thoughts or clarifications you might wish to share.

Thank you again for your contribution to the review process.

Best regards,

The Authors

Reviewer_x9nTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces DistilRecipe, a framework for extracting reusable procedural knowledge from reasoning problems and distilling it into smaller language models. Unlike conventional chain-of-thought (CoT) distillation, which transfers instance-specific reasoning trajectories, DistilRecipe represents generalizable reasoning strategies as structured recipes comprising applicability conditions, ordered procedures, and verification checks. The framework constructs a recipe bank through an iterative process involving a larger recipe author, a frozen student model, and a procedural monitor. Recipes are accepted based on answer correctness, applicability, and procedural adherence, while revisions must preserve previously verified executions. The resulting recipes can either guide frozen models during inference or serve as privileged context for supervised fine-tuning (SFT) and on-policy distillation (OPD), enabling students to internalize procedural knowledge without requiring recipes at deployment. Experiments on GSM8K and MATH demonstrate that recipe-conditioned inference consistently improves accuracy across Qwen3.5 models of different sizes and generation budgets, achieving average absolute gains of 5.62 and 9.00 percentage points, respectively. Recipe-based teacher–student distillation also outperforms matched CoT distillation by 2.08–3.20 percentage points across four comparisons.

Critical strengths
  1. A well-motivated framework for extracting and transferring reusable procedural knowledge. The paper addresses an important limitation of conventional reasoning distillation: individual CoT trajectories entangle generalizable procedures with instance-specific details. Representing procedural knowledge explicitly through applicability conditions, ordered operations, and verification checks is a meaningful approach to improving knowledge transfer. The student-grounded verification mechanism is particularly interesting. By requiring successful execution and procedural adherence before accepting a recipe, the framework attempts to ensure that the extracted knowledge is executable by the target student rather than merely plausible to a stronger teacher. The support-preserving revision mechanism further provides a principled way to maintain consistency during recipe-bank construction.

  2. Consistent empirical improvements and a reasonably controlled distillation evaluation. The experimental results provide encouraging evidence for the utility of reusable recipes. Recipe-conditioned inference improves accuracy across all 30 evaluated combinations of model size, benchmark, and generation budget. The improvements persist even with larger generation budgets, suggesting that recipes provide useful procedural guidance beyond simply compensating for short reasoning outputs. The distillation experiments are also valuable because they evaluate students without access to recipes at test time. In particular, the teacher–student comparisons use matched training problems, teacher models, and optimization objectives, providing evidence that recipe-conditioned supervision offers benefits beyond conventional CoT distillation.

Critical weaknesses
  1. Insufficient ablation studies to establish the contribution of the proposed student-grounded verification mechanism.

Although student-grounded recipe verification is presented as a central methodological contribution, the experiments do not adequately isolate its effectiveness.

The framework combines several components, including recipe abstraction by a strong 27B model, question-based routing, student execution verification, procedural adherence monitoring, and support-preserving revision. However, the paper does not compare the complete framework against variants that construct recipes without student verification, use answer correctness alone, or remove the adherence monitor and support-preservation mechanism.

Consequently, it remains unclear whether the improvements primarily arise from the proposed verification mechanism or simply from providing high-quality procedural instructions generated by a substantially stronger model.

The random-recipe experiment demonstrates the importance of matching recipes to problems, but it does not establish the necessity of student-grounded verification. Comparisons with unverified teacher-generated recipes or closely related reusable-strategy methods would substantially strengthen the central contribution.

This limitation is consequential because the distinction between teacher-generated reasoning abstractions and student-verified executable procedures is fundamental to the paper's novelty.

  1. Limited experimental scope and insufficient evidence for the robustness of distillation improvements.

The evaluation is restricted to two mathematical reasoning benchmarks, GSM8K and MATH, using models from a single architecture family. While the recipe-conditioned inference improvements are substantial, the advantages over matched CoT distillation are relatively modest, ranging from 2.08 to 3.20 percentage points.

The paper does not report results across multiple training seeds, confidence intervals, or statistical significance tests. This makes it difficult to determine the reliability of these relatively small improvements, particularly on MATH-500. Additionally, the negative result of self-distillation with OPD on MATH-500 suggests that the effectiveness of recipe-based knowledge transfer may depend considerably on the teacher's capabilities and training configuration.

More broadly, both benchmarks involve relatively structured mathematical procedures. It remains uncertain whether the same recipe representation and verification process can generalize to less structured reasoning tasks that require flexible planning, diverse intermediate decisions, or adaptive problem-solving strategies.

These limitations weaken the evidence supporting the broader claim that reusable procedural knowledge provides a generally superior unit for reasoning distillation.

Questions for authors

Q1. How much of the improvement can be attributed specifically to student-grounded verification?Could the authors compare DistilRecipe against recipe banks generated by the same 27B model without student execution filtering, as well as variants using only final-answer correctness or omitting the procedural adherence monitor? Ideally, these comparisons should control for the number of recipes, authoring examples, and computational budget. Such results would help determine whether the proposed verification mechanism provides benefits beyond strong teacher-generated procedural instructions.

Q2. How robust are the reported improvements over CoT distillation? Could the authors report results across multiple training seeds or provide paired statistical significance analyses for the four teacher–student distillation comparisons? Given the relatively small improvements, understanding their variance would materially affect confidence in the claimed superiority of recipe-based distillation.

Q3. How sensitive is the framework to recipe-routing quality, and what proportion of held-out problems receive genuinely applicable recipes? The experiments show that incorrect recipe assignment can substantially degrade performance. Could the authors quantify the accuracy of recipe routing or provide a breakdown of performance for correctly versus incorrectly matched recipes? This would clarify how much of the observed improvement depends on the quality of the recipe bank versus the effectiveness of the routing mechanism, and how reliably the framework might transfer to new problem distributions.

3 comments

Response to Reviewer x9nT (1/2)

We thank the reviewer for the thoughtful and constructive review, and for recognizing the motivation, and the consistent empirical improvements. We address the concerns below.

W1 / Q1. Ablation of student-grounded verification

We rebuilt the MATH bank with the same author (Qwen3.5-27B), 1,000 authoring problems, router and attempt budget, changing only the acceptance rule. The rows are the variants the reviewer requested (MATH-500, 2,048 tokens):

Acceptance ruleRecipes0.8B2BMean gain
No recipe–38.0059.60–
No verification (author's recipe as written)17842.2064.40+4.5
Monitor judgment only (no student execution)7342.4066.00+5.4
Final-answer correctness only (no monitor)8043.4065.40+5.6
Full e·α·h (ours)6845.6069.00+8.5

Verification nearly doubles the gain of the same author's recipe (+4.5 → +8.5). The unverified bank is exactly the reviewer's alternative hypothesis: procedures from the same 27B model on the same authoring problems. It is the weakest of the four variants despite holding the most recipes, so neither the author's strength nor the bank size explains the gain.

Both checks are needed. Either check alone, the answer without the monitor or the monitor without execution, adds only about one point over no verification; the full rule adds four, as the factorization in §3.3 predicts. The construction logs show why: 26.7% of answer-correct executions did not follow the recipe, and an answer-only rule, would have accepted them as evidence.

Cost. Verification is a one-time cost per domain (3,431 anchor executions and 2,935 monitor calls for MATH), amortized over every later use: the same frozen bank serves all recipients, including models from other families, supplies the privileged context for distillation, and adds nothing at inference. Replay guarantees that a revision never loses a problem its predecessor solved (Prop. 1); during MATH construction, 496 failed replays blocked revisions that had passed their triggering problem.

W2 / Q2. Robustness of the gains over CoT distillation

We ran three seeds per arm with the same problems and schedule (mean ± sd; Qwen3.5-2B student, 2,048 tokens):

ObjectiveBenchmarkBaseCoT distillationRecipe distillation
SFTMATH-50059.6060.60 ± 0.5363.80 ± 1.11
SFTGSM8K71.5775.29 ± 1.2178.19 ± 1.12
OPDMATH-50059.6063.53 ± 0.6466.33 ± 0.58
OPDGSM8K71.5780.07 ± 0.7883.47 ± 1.01

The gains reproduce the single runs of Table 2, are 2.5–4.6× the pooled seed sd, and favor the recipe in all four comparisons (by 2.8–3.4 points).

W2. Scope: other model families and the choice of benchmarks

The method is not math-specific by design: recipes, routing, monitoring and replay operate over abstract roles, and the domain enters only through the answer evaluator and prompt wording (§3.2, App. A.1, B). GSM8K and MATH are the standard benchmarks for mathematical reasoning in recent distillation work (e.g., [2–5]), and our MATH protocol (a bank built from 1,000 training problems, evaluated on MATH-500) follows the closest prior work on reusable reasoning behaviors [6]. The two benchmarks also differ in kind: GSM8K recipes are arithmetic schemas over story quantities, while MATH spans competition problems in seven subjects that need symbolic methods (App. E).

Recipes also transfer across model families. Although the banks were built and verified with Qwen models, they improve Gemma-3, Phi-4-mini and Llama-3.2 recipients in all eight settings, by up to +12.0 points, without any re-authoring

RecipientMATH-500: no recipe → recipeGSM8K: no recipe → recipe
Gemma-3-1B-it39.00 → 51.00 (+12.0)48.37 → 56.94 (+8.6)
Gemma-3-4B-it70.20 → 78.40 (+8.2)87.49 → 89.61 (+2.1)
Phi-4-mini-instruct58.80 → 65.00 (+6.2)88.17 → 92.65 (+4.5)
Llama-3.2-3B-instruct44.40 → 49.00 (+4.6)78.47 → 80.89 (+2.4)

Response to Reviewer x9nT (2/2)

Q3. Sensitivity to routing

We ran every shortlisted recipe, then every recipe in the bank, on every test problem (Qwen-2B, 2,048 tokens). We compared it to the routed recipe (as reported in the paper).

SelectionGSM8KMATH-500
Recipe-free71.5759.60
Router (12-recipe shortlist → one recipe selected)77.7969.00
Applicable recipe in the shortlist (oracle)82.2175.40
Applicable recipe anywhere in the bank (oracle)86.1879.80

Most test problems (80–86%) have a genuinely applicable recipe in the bank, and the router realizes about 59% of the headroom within its shortlist. Matching matters: a routed recipe adds 6.2–9.4 points, while a random one costs 16.2–42.5 (Table 1). Of the router's remaining errors, 62–65% have no applicable recipe in the bank at all (coverage), 14–18% have one outside the shortlist (retrieval), and 20–21% have one in the shortlist that was not chosen (routing). The binding constraint on a new distribution is therefore the authoring budget, not the router (Figure 6).

W2 / Q3. Transfer to new distributions without new authoring

We applied the frozen GSM8K bank, with no new authoring and the same router, prompts and 2,048-token budget, to SVAMP [7] (word-problem variations designed to defeat surface heuristics) and to two perturbed versions of GSM8K: GSM-Symbolic [8] (new names and numbers) and GSM-Hard [9] (larger numbers):

Dataset (n)RecipientNo recipeRecipe
SVAMP (300)0.8B64.3369.33
2B77.3388.67
4B93.6795.67
GSM-Symbolic (1,000)0.8B46.3051.80
2B71.5072.00
4B90.0094.40
GSM-Hard (1,319)0.8B27.4531.01
2B39.6546.63
4B60.2767.55

Recipes improve accuracy in all nine settings (+0.5 to +11.3), with an average gain of +5.2, close to the +5.6 they give on GSM8K itself. Nearby distributions are therefore served by the existing bank without new authoring, while a distant domain would need a new bank (Figures 6–7).

We thank the reviewer again for the constructive feedback, which has strengthened the paper. We hope our response addresses the concerns, and we welcome any further questions.

References

[1] Qu, Y., Singh, A., Lee, Y., Setlur, A., Salakhutdinov, R., Finn, C., and Kumar, A. RLAD: Training LLMs to discover abstractions for solving reasoning problems. In ICLR, 2026.

[2] Agarwal, R., Vieillard, N., Zhou, Y., Stanczyk, P., Ramos Garea, S., Geist, M., and Bachem, O. On-policy distillation of language models: Learning from self-generated mistakes. In ICLR, 2024.

[3] Xu, W., Han, R., Wang, Z., Le, L. T., Madeka, D., Li, L., Wang, W. Y., Agarwal, R., Lee, C.-Y., and Pfister, T. Speculative knowledge distillation: Bridging the teacher-student gap through interleaved sampling. In ICLR, 2025.

[4] Li, Y., Yue, X., Xu, Z., Jiang, F., Niu, L., Lin, B. Y., Ramasubramanian, B., and Poovendran, R. Small models struggle to learn from strong reasoners. arXiv:2502.12143, 2025.

[5] DeepSeek-AI. DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning. arXiv:2501.12948, 2025.

[6] Didolkar, A., Ballas, N., Arora, S., and Goyal, A. Metacognitive reuse: Turning recurring LLM reasoning into concise behaviors. arXiv:2509.13237, 2025.

[7] Patel, A., Bhattamishra, S., and Goyal, N. Are NLP models really able to solve simple math word problems? In NAACL, 2021.

[8] Mirzadeh, I., Alizadeh, K., Shahrokhi, H., Tuzel, O., Bengio, S., and Farajtabar, M. GSM-Symbolic: Understanding the limitations of mathematical reasoning in large language models. In ICLR, 2025.

[9] Gao, L., Madaan, A., Zhou, S., Alon, U., Liu, P., Yang, Y., Callan, J., and Neubig, G. PAL: Program-aided language models. In ICML, 2023.

AuthorsOpenReview ↗

Follow-up before the discussion deadline

Dear Reviewer x9nT,

We truly appreciate the time and effort you've already dedicated to reviewing our submission. We prepared our rebuttal with great care in hopes of directly addressing the valuable points raised in your review, including an ablation of student-grounded verification, three-seed robustness results, a routing sensitivity analysis, and transfer to other model families and new distributions.

If these results address your concerns, we would be grateful if you would consider raising your score. We would also welcome any further thoughts or clarifications you might wish to share.

Thank you again for your contribution to the review process.

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.