acceptodds
Under review as a conference paper at ICLR 2027

Does self-distillation improve reasoning?

Abstract

In self-distillation, a teacher is constructed by prompting a student model with privileged information or environmental feedback, and logit-level feedback is derived from the difference between the two models' output distributions. This paradigm has recently gained traction as an alternative to policy-gradient reinforcement learning, promising dense supervision where reward signals are sparse and resting on the assumption that the self-conditioned teacher approximates the optimal KL-regularized policy. This paper critically evaluates whether self-distillation methods indeed extract useful guidance from privileged information in reasoning tasks or whether their gains are driven by simpler mechanisms. We report four findings: (1) Self-Distillation Policy Optimization (SDPO) only improves performance on LiveCodeBench v6 problems seen during training whereas policy-gradient training on the same reward generalizes to held-out problems; (2) SDPO's reported gains in science reasoning and tool use are largely recoverable through supervised fine-tuning on final answers without chain-of-thought; (3) the gains of On-Policy Self-Distillation (OPSD) in math can be attributed to distilling thinking-mode behavior into the non-thinking student, a mechanism that does not require privileged information; and (4) SDPO and OPSD gradients are not aligned with the policy gradient, contradicting the optimal-teacher assumption. We evaluate across Qwen3 models from 1.7B to 14B and Olmo-3-7B models. Together, these results indicate a need for robust methods that leverage privileged information to improve reasoning performance. Our code is available at https://anonymous.4open.science/r/does-sd-improve-reasoning/.

Primary area:
foundation or frontier models, including LLMs
Keywords:
reinforcement learning, large language models, distillation, self-distillation, policy optimization, rich feedback

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_FvvkOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper critically re-examines recent claims that self-distillation improves reasoning, focusing on SDPO and OPSD. Through reproductions and controlled experiments across Qwen3 models from 1.7B to 14B and Olmo-3-7B, the paper argues that several reported gains have alternative explanations. For SDPO on code, gains largely disappear under problem-disjoint evaluation, suggesting memorization of training problems rather than transferable reasoning. On science and tool-use tasks, supervised fine-tuning on final answers matches or exceeds SDPO, questioning whether privileged reasoning traces are necessary. For OPSD, the paper finds that performance is largely explained by transferring the teacher's thinking-mode behavior rather than by the particular privileged solution supplied to the teacher. Finally, it compares self-distillation updates with reward-optimization gradients and finds poor alignment on held-out prompts.

Critical strengths
  • This is a valuable corrective study because it proposes concrete alternative explanations and changes the experimental setting to test them. The problem-disjoint coding split cleanly distinguishes generalization to new test cases of the same specification from generalization to unseen problems, showing that these are substantially different claims.
  • The analysis of OPSD is particularly convincing. Varying the teacher's thinking mode independently of whether the reference solution is correct, absent, or shuffled directly separates thinking-mode transfer from privileged-information transfer. The output-format artifacts and response-length analysis provide additional evidence for the same explanation, and the results are reproduced across model sizes and families.
Critical weaknesses
  • Two interpretations are stronger than the corresponding interventions support. Final-answer SFT matching SDPO shows that explicit rationales or privileged CoT supervision are unnecessary for these benchmark gains; it does not show that the task or the model's computation requires no reasoning. Similarly, the gradient experiment compares self-distillation updates on training prompts with a GRPO gradient on held-out prompts. Negative cosine similarity is evidence of poor alignment with generalizable reward optimization, but it does not directly contradict Proposition 3.1, which assumes the same prompt distribution, the ideal KL objective, and the true optimal teacher.
  • The broad conclusion is based on two particular methods, SDPO and OPSD, and several related failure modes have also been reported by concurrent work discussed in the paper. The experiments provide strong evidence about these implementations, but they do not cover self-distillation teachers conditioned on the student's actual attempt and verifier outcome, or other methods that construct the teacher differently. The title and general conclusion should be scoped accordingly.
Questions for authors
  1. For the gradient analysis, what happens when both gradients are estimated on exactly the same prompts and checkpoint, before separately evaluating whether either direction improves held-out reward? This would distinguish objective misalignment from train-test distribution shift.
  2. Which conclusions are intended to extend to methods whose teacher observes the student's actual attempt and verifier outcome, rather than a privileged reference solution or thinking-mode prompt? This boundary materially affects whether the paper is about SDPO/OPSD or self-distillation more generally.
3 comments

Thank you for your positive feedback on the value of the paper and its analysis! We revise our manuscript based on your feedback below. Please note that for the revision, all changes are in red and new material is Section 4.3, Section 4.5, and Appendices D.3.2, D.3.4, and D.6.1.

Final-answer SFT matching SDPO shows that explicit rationales or privileged CoT supervision are unnecessary for these benchmark gains; it does not show that the task or the model's computation requires no reasoning

  • We clarify that our claim is that these benchmark gains do not establish improved reasoning, rather than that SDPO performs no reasoning
  • We add new experiments showing that SDPO degrades aggregate performance on domains not trained on by up to 16.6 percentage points whereas GRPO increases average performance in 4 of 5 training domains (revision Section 4.3)
  • Together with the coding overfitting results (Section 4.1), the new experiments demonstrate limited transfer under the tested SDPO recipe

Similarly, the gradient experiment compares self-distillation updates on training prompts with a GRPO gradient on held-out prompts. Negative cosine similarity is evidence of poor alignment with generalizable reward optimization, but it does not directly contradict Proposition 3.1, which assumes the same prompt distribution, the ideal KL objective, and the true optimal teacher.

  • We withdraw the teacher-optimality conclusion as gradient misalignment is not a test of teacher optimality. Revised Section 4.5 instead examines the alignment of practical self-distillation gradients with an estimated held-out reward gradient

For the gradient analysis, what happens when both gradients are estimated on exactly the same prompts and checkpoint, before separately evaluating whether either direction improves held-out reward? This would distinguish objective misalignment from train-test distribution shift.

  • We add new gradient comparisons on matched problems at the same checkpoint using Qwen3-8B (Appendix D.6). With a thinking teacher and the reference solution, OPSD remains negatively aligned with GRPO at initialization (cosine similarity −0.128), despite using the same problems and checkpoint
  • We have not directly measured held-out reward changes after controlled updates: small-step effects are difficult to resolve reliably within our evaluation budget, while larger steps introduce higher-order terms that prevent isolating the first-order effect measured by gradient alignment
  • This difficulty is consistent with Armandpour et al., who report needing ~45–200K rollouts per problem for their gradient-alignment diagnostic (Sections 2.5–3). While their alignment metric differs in implementation, we cite this as context for the scale of the computational cost needed

Which conclusions are intended to extend to methods whose teacher observes the student's actual attempt and verifier outcome, rather than a privileged reference solution or thinking-mode prompt?

  • On LCBv6, the SDPO teacher is already constructed from verifier feedback on student generated trajectories (expected outputs vs received inputs on failing test cases)
  • In general, all our SDPO experiments (LCBv6/SciKnowEval/ToolAlpaca) use no dataset reference solution: the teacher sees verifier feedback, and, when one exists, a passing sibling rollout.
  • Our analysis does not extend to self-reflection methods. We attempted to reproduce Self-Distillation Zero (He et al.): base-model accuracy closely matched the paper (49.5% vs. 48.1% across six math benchmarks), but its initial self-revision training (SRT) stage reduced accuracy to 42.0%, versus the reported increase to 55.7%. Unable to resolve this discrepancy or differences between the paper’s and released code’s training procedures, we excluded the method from our reported results due to difficulty in reproducing their results

We hope our responses above have resolved your concerns about our analyses, and we would be glad to address any additional questions you may have.

References:

  • Yinghui He, Simran Kaur, Adithya Bhaskar, Yongjin Yang, Jiarui Liu, Narutatsu Ri, Liam Fowl, Abhishek Panigrahi, Danqi Chen, and Sanjeev Arora. Self-distillation zero: Self-revision turns binary rewards into dense supervision, 2026. URL https://arxiv.org/abs/2604.12002
  • Armandpour, Mohammadreza, et al. "Unmasking on-policy distillation: Where it helps, where it hurts, and why." arXiv preprint arXiv:2605.10889 (2026).

Dear Reviewer Fvvk,

The authors have provided a detailed reply to your review. Could you please take a look, acknowledge the rebuttal, and give your response?

Thank you for serving as a reviewer for ICLR.

Best,

AC

Most of my concerns have been addressed and I remain my positive judgement to the work.

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

This paper audits whether privileged-information self-distillation improves reasoning. It separates held-out coding problems from held-out unit tests, compares SDPO with final-answer SFT on science/tool tasks, factorially varies OPSD teacher mode and reference information, and measures gradient alignment. Experiments cover Qwen3 models from 1.7B to 14B and Olmo-3-7B, with four training seeds. The strongest findings are poor SDPO transfer to new coding problems and the ability of irrelevant-reference controls to recover OPSD gains. The stronger interpretation of gradient misalignment as a test of teacher optimality requires correction.

Critical strengths
  1. The coding audit makes a consequential evaluation distinction: passing new unit tests for a trained-on problem is different from solving a new problem. At step 200, held-out-problem changes are -2.5/-0.5/-9.9 points for SDPO versus +3.7/+4.9/+2.6 for GRPO across the three models (Tables 4–6). Four seeds, two model families, and separate seen/unseen-test-case curves make this a useful and comparatively robust empirical contribution. The original SDPO paper explicitly describes its within-problem test split; this submission usefully evaluates an additional generalization question.

  2. The teacher-mode/reference factorial controls are more informative than an unablated performance comparison. Matched, absent, and shuffled references, both evaluation modes, response lengths, and orphan closing tags jointly support the narrower conclusion that a target-matched solution is unnecessary in the tested recipes. The final-answer SFT baseline is also practically valuable: it matches or exceeds SDPO point estimates in four of five science/tool domains (Table 7). These controls help separate available base-model behavior from the claimed benefit of target-specific privileged information.

Critical weaknesses
  1. The inference from Proposition 3.1 to the teacher-optimality diagnosis is not established. The proposition and Appendix B proof concern complete-sequence KL gradients for one fixed prompt and appear correct. The authors explicitly acknowledge practical differences on p.8, lines 419–422, but their claimed directional consequence still need not hold. Equation (3) stops gradients through student-sampled prefixes; equal reverse-KL values do not imply equal gradients. In a two-step binary-reward policy with an exactly optimal frozen teacher, this token loss can have derivative -0.01503 while the reward derivative is -0.015, so minimizing the token loss moves opposite to reward ascent; the full-sequence KL derivative is correctly +0.015. Thus even an ideal teacher need not pass the proposed test. Comparing OpenThoughts gradients with benchmark gradients adds a distribution change, and Appendix E.3 additionally replaces LoRA with full-parameter gradients. These are interesting empirical similarities, but cannot by themselves establish the teacher claim on p.9, lines 461–466. The original SDPO Appendix A.1 also distinguishes token and sequence estimators.

  2. The evaluation supports narrower quantitative and causal conclusions than several headlines. Tables 1 and 8–11 select the best checkpoint using the same reported benchmark averages; Appendix D states this explicitly, but neither D nor E identifies an independent checkpoint-selection set or selection-adjusted uncertainty. Four-seed standard deviations do not resolve this issue, particularly for small thinking-on gains. Also, final-answer-only supervision does not establish that a task or model performs no reasoning. Correct versus absent references change thinking-off accuracy by 7.2 and 11.4 points at 1.7B and 4B (Tables 8–9), although shuffled references recover the gains. This supports lack of necessity of target matching, rather than reference content being irrelevant or an exclusive causal explanation by thinking transfer. Appendix E.4 retains different teacher instructions even with empty reference content. Existing response-length statistics are useful, but fixed-token-budget comparisons would better separate longer computation from improved efficiency.

Questions for authors
  1. Can the teacher-optimality conclusion be withdrawn or supported by a diagnostic valid for Equation (3), with matched prompts and sampling? Please distinguish the valid trajectory-KL proposition from its proposed practical consequence; calibration against an exactly optimal teacher on an enumerable example would be decisive.

  2. Tables 1 and 8–11 select checkpoints using the reported benchmark averages. How sensitive are the main conclusions to this selection rule? Would fixed-final-checkpoint or independently selected results support the same comparisons? Please clarify how training, checkpoint selection, and final evaluation are separated, and how benchmark overlap is controlled.

  3. Can the claims be narrowed to absence of evidence for a necessary target-matched reference and no need for explicit chain-of-thought supervision? Would the key comparisons persist at a fixed inference token budget, and which reported convergence/comparison claims are intended to hold beyond the inherited training budgets?

5 comments

Thank you for the constructive review and detailed feedback. We appreciate your positive assessment of the robustness and practical value of our SDPO and OPSD experiments across coding, science reasoning, and math settings. Please note that for the revision, all changes are in red and new material is Section 4.3, Section 4.5, and Appendices D.3.2, D.3.4, and D.6.1.

We address the specific weaknesses and questions mentioned below.

The inference from Proposition 3.1 to the teacher-optimality diagnosis is not established.

  • We agree with your analysis, including the two-step example: Eq. 3 does not differentiate through sampled prefixes, so an exactly optimal teacher need not pass the test. We withdraw the teacher-optimality conclusion and remove it from the abstract, Section 3, Section 4.5, and the conclusion. Revised Section 4.5 instead examines the alignment of practical self-distillation gradients with an estimated held-out reward gradient

Can the teacher-optimality conclusion be withdrawn or supported by a diagnostic valid for Equation (3), with matched prompts and sampling? Please distinguish the valid trajectory-KL proposition from its proposed practical consequence

  • We replace the teacher-optimality conclusion with a Taylor expansion argument compatible with Equation 3 self-distillation gradient estimators in our revised Section 4.5. This is motivated by the following equation, which does not depend on Proposition 3.1: Jheld(θ0+ηg)−Jheld(θ0)=η⟨∇θJheld(θ0),g⟩+O(η2),\small J_{\mathrm{held}}(\theta_0 + \eta g) - J_{\mathrm{held}}(\theta_0) = \eta \langle \nabla_\theta J_{\mathrm{held}}(\theta_0), g \rangle + O(\eta^2),
  • We find negative alignment at initialization in the tested Qwen3 OPSD settings and weak, sometimes positive alignment later in training and that SDPO is less aligned than GRPO on LiveCodeBench v6 at initialization. Appendix D.6 repeats the analysis on matched problems and on LoRA parameters. Each method keeps its own sampling parameters, since the Taylor approximation does not depend on matched sampling

Also, final-answer-only supervision does not establish that a task or model performs no reasoning.

  • Our claim is that these benchmark gains do not establish improved reasoning, rather than that SDPO performs no reasoning
  • Moreover, we add new experiments showing that SDPO degrades aggregate performance on domains not trained on by up to 16.6 percentage points whereas GRPO increases average performance in 4 of 5 training domains (revision Section 4.3)
  • Together with the coding overfitting results (Section 4.1), the new experiments demonstrate limited generalizability under the tested SDPO recipe

Tables 1 and 8–11 select the best checkpoint using the same reported benchmark averages; Appendix D states this explicitly, but neither D nor E identifies an independent checkpoint-selection set or selection-adjusted uncertainty.

  • We select one checkpoint based on average performance across four seeds, rather than each seed’s best checkpoint separately. This reduces sensitivity to seed-specific fluctuations, although some bias remains. However, we now show that our conclusions regarding OPSD performance are robust towards checkpoint-selection schemes
  • We report our numbers under fixed-checkpoints (step 200/500) in Appendix D.3.2 (Tables 13-20) of our revision. Our main conclusions remain unchanged: OPSD achieves higher thinking-off accuracies with a thinking teacher, and at every model size, at least one of the shuffled or absent variants matches or exceeds the matched reference
  • We keep our original best-checkpoint-selection results in the main body to maintain parity with OPSD's reported results

Can the claims be narrowed to absence of evidence for a necessary target-matched reference and no need for explicit chain-of-thought supervision?

  • Yes, these were our intended claims, and we have reworded our introduction, revision Section 4.4, and conclusion for clarity

Would the key comparisons persist at a fixed inference token budget, and which reported convergence/comparison claims are intended to hold beyond the inherited training budgets?

  • We add new analyses in Appendix D.3.4 that supports our interpretation that OPSD’s thinking-off gains depend on internalizing token-hungry thinking behavior
  • Specifically, at 4,096 tokens, all thinking-teacher OPSD variants underperform the base model, with large gains being shown only at higher maximum token lengths
  • All our previous evaluations were done under a fixed inference budget (8192 tokens for SDPO, 38912 for OPSD) and all comparisons are at the stated budgets (200 steps for SDPO, OPSD and SFT; 500 for GRPO on math); we make no claims beyond them

Appendix E.4 retains different teacher instructions even with empty reference content

  • We note that student and teacher prompts typically differ in self-distillation, as the goal is to construct a stronger teacher through additional context.
  • Furthermore, our teacher prompts are identical across matched/empty/shuffled reference solutions, ablating only the solution content

[On OPSD] ... how benchmark overlap is controlled

  • The OpenThoughts subset that we train on has already been decontaminated against AIME24 and predates AIME25 and HMMT25. See OpenThoughts' original implementation here and here (Guha et al.)
  • We also did independent investigations into contamination across training and evaluation problems, finding no evidence for dataset contamination

We hope the responses above address your concerns about our analyses and welcome any follow-up questions you may have.

References:

  • Siyan Zhao, Zhihui Xie, Mengchen Liu, Jing Huang, Guan Pang, Feiyu Chen, and Aditya Grover. Self-distilled reasoner: On-policy self-distillation for large language models. In Forty-third International Conference on Machine Learning, 2026a. URL https://openreview.net/forum?id=Jpxfof0EaS.
  • Etash Guha, Ryan Marten, Sedrick Keh, et al. Openthoughts: Data recipes for reasoning models, 2025. URL https://arxiv.org/abs/2506.04178.

Dear Reviewer hjia,

The authors have provided a detailed reply to your review. Could you please take a look, acknowledge the rebuttal, and give your response?

Thank you for serving as a reviewer for ICLR.

Best,

AC

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

This paper critically reexamines whether the gains reported by recent on-policy self-distillation methods reflect genuine improvements in reasoning or more limited effects. Through reproductions and controlled experiments on SDPO and OPSD, the authors identify several potential confounders: overfitting to training problems in code generation, strong performance from answer-only supervision on science and tool-use tasks, and the transfer of thinking-mode behavior rather than privileged information in mathematical reasoning. The paper also derives a relationship between distillation toward an ideal KL-regularized teacher and policy gradients, then measures gradient alignment in practical methods. The results question whether existing self-conditioned teachers provide reward-aligned supervision.

Critical strengths
  • The paper raises an important and timely question about whether reported self-distillation gains reflect genuine reasoning improvements. The overall analysis is well organized, and the controlled experiments provide a useful examination of commonly assumed mechanisms behind self-distillation.
  • The code-generation experiments identify a consequential evaluation issue by distinguishing generalization to unseen test cases from generalization to entirely unseen problems. The finding that SDPO improves performance on training problems but fails to generalize to held-out problems is particularly important for interpreting previous results.
  • The controlled ablations separating thinking-mode behavior from privileged information are convincing and informative.
Critical weaknesses
  • Some conclusions about reasoning improvements are stronger than what the experiments directly establish. In particular, the fact that answer-only SFT matches or outperforms SDPO does not necessarily imply that SDPO fails to improve reasoning capabilities. Answer-only SFT receives direct supervision on the evaluation target, and the selected benchmarks may not adequately distinguish answer memorization from compositional reasoning. The results convincingly question whether existing evaluations demonstrate reasoning improvement, but do not fully establish the absence of such improvement.
  • The gradient-alignment analysis is an interesting diagnostic, but its connection to actual optimization outcomes is not sufficiently established. Proposition 3.1 relies on an ideal KL-regularized teacher, whereas practical self-distillation methods use approximations and evolving policies. Moreover, negative cosine similarity with an estimated GRPO gradient at initialization does not necessarily imply worse optimization or generalization under adaptive optimizers and changing rollout distributions. The paper would benefit from directly relating gradient alignment to measured changes in held-out reward after controlled policy updates.
Questions for authors
  • Can the gradient-alignment analysis be connected to actual reward changes? Some empirical verification would be helpful.
  • The gradient comparison is performed at the initial policy. How does alignment evolve during training? Is there evidence that the self-distillation direction becomes more or less aligned with the reward gradient as the student improves?
  • The experiments focus on SDPO and OPSD. Which of the identified failure modes do the authors expect to persist under selective or contrastive self-distillation, where teacher supervision is explicitly filtered using reward or verifier information? What would be the most informative test of that hypothesis?
2 comments

Thank you for the thoughtful review. We appreciate your recognition of the importance of our generalization analysis and thinking-mode ablations, and of the timeliness of the question. Please note that for the revision, all changes are in red and new material is Section 4.3, Section 4.5, and Appendices D.3.2, D.3.4, and D.6.1.

Below we address the specific concerns raised.

The [Answer-only SFT] results convincingly question whether existing evaluations demonstrate reasoning improvement, but do not fully establish the absence of such improvement.

  • Our claim is that these benchmark gains do not establish improved reasoning, rather than that SDPO performs no reasoning: final-answer SFT recovers comparable gains without explicit chain-of-thought supervision.
  • We add new experiments showing that SDPO degrades aggregate performance on domains not trained on by up to 16.6 percentage points whereas GRPO increases average performance in 4 of 5 training domains (revision Section 4.3).
  • Together with the coding overfitting results (Section 4.1), the new experiments demonstrate limited transfer under the tested SDPO recipe.

Proposition 3.1 relies on an ideal KL-regularized teacher, whereas practical self-distillation methods use approximations and evolving policies.

  • We agree. We have moved Proposition 3.1 to Appendix B and use it for motivation only, and withdraw the teacher-optimality conclusion, as gradient misalignment is not a test of teacher optimality.
  • Our revised Section 4.5 instead examines the alignment of practical self-distillation gradients with an estimated held-out reward gradient. This diagnostic is motivated by the following Taylor expansion (that does not depend on Proposition 3.1): Jheld(θ0+ηg)−Jheld(θ0)=η⟨∇θJheld(θ0),g⟩+O(η2);\small J_{\mathrm{held}}(\theta_0 + \eta g) - J_{\mathrm{held}}(\theta_0) = \eta \langle \nabla_\theta J_{\mathrm{held}}(\theta_0), g \rangle + O(\eta^2); Eq. 4 holds for any direction g, so it needs no assumption on the teacher

The paper would benefit from directly relating gradient alignment to measured changes in held-out reward after controlled policy updates

  • We have not directly measured held-out reward changes after controlled updates: small-step effects are difficult to resolve reliably within our evaluation budget, while larger steps introduce higher-order terms that prevent isolating the first-order effect measured by gradient alignment
  • This difficulty is consistent with Armandpour et al., who report needing ~45–200K rollouts per problem for their gradient-alignment diagnostic (Sections 2.5–3). While their alignment metric differs in implementation, we cite this as context for the scale of the computational cost needed
  • The goal of our gradient alignment analysis is complementary to our empirical diagnostics, which study domain-overfitting, thinking-mode leakage, and reward curves throughout training and incorporate optimizer dynamics (Sections 4.1–4.4). We interpret alignment as a local diagnostic relative to an estimated held-out GRPO direction

How does alignment evolve during training?

  • We add a Qwen3-8B OPSD training-trajectory case study: alignment with held-out GRPO is negative at initialization and potentially weakly positive later in training (revised Appendix D.6).
  • For example, at step 200, OPSD with the reference solution and teacher thinking enabled achieves a cosine of 0.022, compared with 0.375 for GRPO at the same checkpoint.

The experiments focus on SDPO and OPSD. Which of the identified failure modes do the authors expect to persist under selective or contrastive self-distillation, where teacher supervision is explicitly filtered using reward or verifier information?

  • SDPO already filters supervision by verifier outcomes: it supervises only incorrect trajectories, using correct sibling rollouts and verifier feedback to construct the teacher. Problem-level overfitting (Section 4.1) and OOD degradation (Section 4.3) persist under this filtering, suggesting that filtering alone may not prevent these failures.
  • The most informative test would compare self-distillation variants against SDPO and GRPO on the same problem-disjoint and cross-domain evaluations, while controlling teacher thinking mode. Whether these variants mitigate the observed failures remains an open question that we leave to future work.

We hope the responses we detail above resolve any concerns you may have regarding our analyses. We are more than happy to address any follow-up questions.

References:

  • Armandpour, Mohammadreza, et al. "Unmasking on-policy distillation: Where it helps, where it hurts, and why." arXiv preprint arXiv:2605.10889 (2026).

Dear Reviewer Nmmp,

The authors have provided a detailed reply to your review. Could you please take a look, acknowledge the rebuttal, and give your response?

Thank you for serving as a reviewer for ICLR.

Best,

AC

Comments on the paper

2 comments
AuthorsOpenReview ↗

Request for reviewer follow-up

Dear Reviewers Nmmp, hjia, Fvvk

A gentle reminder that we have responded to your reviews. We would be happy to follow up on any remaining concerns, and would welcome any additional feedback. We thank you for your valuable time and consideration.

AuthorsOpenReview ↗

Further request for reviewer follow-up

Dear Reviewers Nmmp, hjia, Fvvk

Another gentle reminder that we have responded to your reviews. The end of the discussion period is soon; we would welcome additional feedback. Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.