acceptodds
Under review as a conference paper at ICLR 2027

MASS: Multi-mask Selective Self-Distillation for Leakage-Resilient Reasoning

Abstract

On-policy self-distillation enables the same model to act as student and teacher, with dense token-level supervision obtained by the teacher being given access to privileged information. However, conditioning the teacher on a single reference can concentrate its predictions on one reasoning trajectory and limit transfer to the student. We propose Multi-Mask Selective Self-Distillation (MASS), which approximates the unavailable marginal over diverse reference realizations using multiple independently masked views of one verified solution. MASS aggregates these complementary teacher predictions, selectively transfers reliable token-level corrections, and adaptively focuses masking on influential reference spans. Across mathematical reasoning benchmarks and model families, MASS consistently outperforms strong self-distillation baselines.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Self-distillation, on-policy distillation, multi-mask distillation, information leakage, language Model Reasoning, mathematical Reasoning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_hSvDOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies on-policy self-distillation (OPSD), where the same model acts as student and teacher and only the teacher also sees a verified reference solution. It argues that conditioning the teacher on one complete reference concentrates its predictions on that single trajectory, and proposes MASS, which builds several teacher views from the same reference by masking different spans.

On AIME24, AIME25 and HMMT25 (Avg@12), MASS is the best method in all 12 model-benchmark cells, and its three-benchmark average is 1.5-2.5 points above the strongest baseline, AVSD or RLSD. It also reports ablations and OOD results on GPQA and BBH.

Critical strengths
  1. MASS uses no more data or rollouts than OPSD and gives consistent gains at moderate cost. It is the best method in all 12 model-benchmark cells (two model families, 1.7B-8B), and its estimated per-step cost is 2.6x OPSD versus 6.4x for RLSD. On Qwen3-4B it keeps improving up to step 250, whereas OPSD falls below the base model.

  2. The token weight is derived explicitly, and its limits are stated. Its population version minimizes the mean-squared error of the weighted correction plus a small ridge term and a linear anchor bonus. Remark A.8 states where this plug-in weight departs from the optimum.

Critical weaknesses
  1. The margins are small, and no variance estimates are given. No seeds, confidence intervals or significance tests are reported, so all results appear to come from single training runs. On datasets such as AIME and HMMT, a difference of one or two problems can easily inflate the metrics. Each benchmark has 30 problems x 12 samples, so the per-benchmark margins over the strongest baseline in Table 1 (0.83-2.8 points) correspond to 3-10 of 360 responses, and many ablation gaps to 1-5 responses.

  2. The checkpoint-selection rule and the baseline settings are not stated. Appendix C gives no hyperparameters, RLSD group size or tuning budget for the baselines.

  3. The benefit of averaging several masks is not tested. Without the weight and adaptive masking, one random mask per step gives the same expected update as the average of K masks (Lemma A.1, Lemma A.4), so averaging helps only by lowering target variance. No K = 1 baseline is reported to check this (Fig. 5(a) covers K = 2, 4, 8).

  4. Fig. 1 does not distinguish broader support for other correct solutions from a teacher that has simply become less strict. Because the averaged teacher keeps any answer that at least one view supports, its PI Bias Score [1] for any continuation, correct or not, is never below the single-view average (Prop. A.3). An incorrect-solution control is therefore needed to show that the rise for a different correct solution is specific to correct solutions. The in-context score also falls when several masks are used, which seems inconsistent with this bound, and the number of problems and the uncertainty are not reported.

[1] S. Harne et al. Privileged, but biased: How PI-conditioned teachers break self-distillation. arXiv:2608.04794, 2026.

Questions for authors
  1. Checkpoints and baselines. Which checkpoint-selection rule was used for each method in Table 1, and were checkpoints selected based on AIME or HMMT results? Please report all methods under a single rule and give the baseline settings, including the RLSD group size and the AVSD configuration.

  2. Variance. Please report mean +/- std over training seeds for Qwen3-4B in Tables 1 and 3, together with paired per-problem bootstrap intervals for the differences between MASS and AVSD and RLSD.

  3. To support the paper's central claim that MASS supports diverse reasoning paths, please also report Pass@k in addition to Avg@k.

  4. Multi-mask control. Please add an unweighted K = 1 random-mask baseline. If K = 8 averaging still clearly wins in this comparison, weakness 3 is largely resolved. Comparisons with suffix masking as in PAINT [1] under the same weight, or with a higher teacher temperature, would also help.

Minor comments:

  • The fonts in the figures are generally too small, which hurts readability.

[1] Z. Tan and Y. Hong. PAINT: Partial-solution adaptive interpolated training for self-distilled reasoners. arXiv:2604.26573, 2026.

4 comments

We thank the reviewer for the detailed and constructive review. Each concern is addressed below, with pointers to the revised PDF (changes in blue).

Weakness 1 & Question 2: small margins, seed variance, and bootstrap intervals. To test whether the margins could come from a lucky run, we trained all methods on Qwen3-4B with three seeds (Table 7, Appendix B.1):

MethodAIME24AIME25HMMT25
OPSD76.57±0.5876.57_{\pm 0.58}68.13±0.4268.13_{\pm 0.42}42.70±0.7042.70_{\pm 0.70}
AVSD78.61±0.7378.61_{\pm 0.73}73.50±0.8073.50_{\pm 0.80}45.65±0.8545.65_{\pm 0.85}
RLSD77.96±0.7077.96_{\pm 0.70}72.50±0.8372.50_{\pm 0.83}45.56±1.1145.56_{\pm 1.11}
MASS81.20±0.64\mathbf{81.20}_{\pm 0.64}73.89±0.73\mathbf{73.89}_{\pm 0.73}48.64±0.96\mathbf{48.64}_{\pm 0.96}

Standard deviations are smaller than the margin of MASS over the strongest baseline on AIME24 (2.6 points) and HMMT25 (3.0 points); on AIME25 the gap to AVSD is within one standard deviation, and we do not claim it as resolved. The component ablation is also reported over three seeds (Table 4, §3.3); each component's gain over the variant without either exceeds one standard deviation. To address the concern that one or two problems can drive the result, Table 2 (§3.2) reports paired bootstrap intervals that resample problems rather than responses:

BenchmarkMASS - AVSDMASS - RLSD
AIME24[+0.38,+4.91][+0.38, +4.91][+0.62,+5.88][+0.62, +5.88]
AIME25[−1.18,+1.96][-1.18, +1.96][+0.16,+2.74][+0.16, +2.74]
HMMT25[+0.41,+5.67][+0.41, +5.67][−0.42,+6.31][-0.42, +6.31]

Four of six intervals lie above zero, and the paper reports the remaining two as unresolved. We will extend Table 7 to all models in Table 1 before the end of the discussion period.

Weakness 2 & Question 1: checkpoint rule and baseline settings. All methods use the same rule (§3.1): checkpoints are evaluated every 50 steps and we report the best Avg@12 across them for each benchmark. We chose a common rule over a fixed final step because the methods peak at different times (Figure 3); a fixed step would penalize OPSD, which peaks early. The full AVSD and RLSD settings, including RLSD's group size G=8G=8 and AVSD's three privileged views, are in Table 15 (Appendix C) and follow each method's published main results. We kept AVSD at three views because its views are heterogeneous rather than interchangeable masks, and its authors report no gain from adding views on mathematics.

Weakness 3 & Question 4: K=1K=1 random-mask control. We added the unweighted K=1K=1 baseline you suggested to Figure 5(a) (§3.3). K=8K=8 clearly wins: K=1K=1 improves only until about step 100 and ends below its starting accuracy, while K=8K=8 keeps improving through step 250 and finishes more than 10 points higher. A single random mask with the same expected update therefore does not reproduce the benefit, consistent with the fact that averaging lowers the variance of the teacher target.

Weakness 4 & Question 3: Figure 1 control and Pass@k.

  • Less strict teacher? Figure 1 (§1) now includes PS incorrect. If multi-mask averaging only made the teacher less strict, the incorrect-solution score would rise with the different-correct score. Instead, moving from a single mask to multiple masks raises it by only 0.002 to 0.003, against 0.040 to 0.063 for the different correct solution, and it stays below the full-reference teacher. The increase for correct continuations is thus more than an order of magnitude larger than for incorrect continuations.
  • Consistency with Proposition A.3. The original single-mask bar came from one random draw, whereas Proposition A.3 bounds the KK-mask score by the average single-view score over the same masks. The revised single-mask bar averages the 8 single-mask views, and all three continuation types now satisfy the proposition.
  • Pass@k. Table 11 (Appendix B.7) reports Pass@12 on Qwen3-1.7B. MASS is highest on all three benchmarks, and its average lead over the strongest baseline grows from 1.5 points under Avg@12 (Table 1) to 7.2 points under Pass@12, indicating that it solves a broader set of problems rather than only answering the same problems more reliably.

We hope these additions resolve the reviewer's concerns, particularly Weakness 3, and would be glad to clarify anything further.

Following up on Weakness 1 and Question 2: Table 1 now reports mean ± std over three training seeds for all four models, not only Qwen3-4B. The table below gives the margin of MASS over the strongest baseline in each cell. ✓ marks a margin larger than both methods' standard deviations; for ✗, the larger standard deviation is shown in parentheses.

ModelAIME24AIME25HMMT25Avg.
Qwen3-1.7B1.67 ✓0.84 ✗ (0.89)1.11 ✓1.66 ✓
Qwen3-4B2.59 ✓0.39 ✗ (0.80)2.99 ✓1.99 ✓
Qwen3-8B1.67 ✓1.20 ✓1.67 ✓1.72 ✓
OLMo-7B1.48 ✓3.15 ✓1.76 ✓2.19 ✓

MASS has the highest mean in all 12 model–benchmark cells, and the margin exceeds one standard deviation in 10 of them. The two exceptions, AIME25 on Qwen3-1.7B and Qwen3-4B, are consistent with the paired bootstrap in Table 2, and we do not claim them as resolved. On the three-benchmark average, the margin is more than three times the standard deviation for every model.

With the updated Table 1, the Avg@12 lead on Qwen3-1.7B used in our Pass@k comparison is 1.7 points (previously 1.5), against 7.2 points under Pass@12 (Table 11).

I thank the authors for their response and the revision. The three-seed results with paired bootstrap intervals, the same checkpoint-selection rule for all methods, the baseline settings, the K = 1 control, the incorrect-solution control and Pass@12 resolve most of the concerns raised in my original review, so I am raising my score to 3.

Thank you for the careful review and for raising your score. If any of your original concerns remain open, we would be glad to address them before the discussion closes.

Reviewer_G1sPOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MASS addresses reference-specific supervision in on-policy self-distillation. The student generates a reasoning trajectory without a reference solution, while the same model evaluates its prefixes under multiple independently masked versions of one verified reference. Arithmetic averaging forms the teacher target. A selective token weight combines estimated correction strength, finite-mask variance, and alignment with a full-reference anchor. Approximate Shapley attribution then adjusts masking probabilities on later visits to each example. Experiments across Qwen3 scales and an OLMo reasoning model report improvements over OPSD, AVSD, and RLSD on three mathematics benchmarks, along with component ablations and GPQA/BBH transfer results.

Critical strengths
  1. Teacher-context construction is a useful intervention, and the reported gains are consistent across the evaluated settings (§§2.1–2.2; Table 1). Producing complementary supervision from one reference avoids requiring several verified solutions. The method improves over both the single-reference baseline and a multi-view baseline across the reported model settings, making the empirical pattern more persuasive than a gain on one model or benchmark.

  2. The separation of view diversity from uncertainty in the finite-view average is well motivated (§2.2; Appendix A.3–A.4). The variance correction and scalar-shrinkage derivation explain why disagreement need not imply unusable supervision. The appendix appropriately notes that unbiased component estimates do not make the clipped, shared-sample weighted gradient unbiased. The factorial weighting/masking ablation and anchor-target comparison provide useful initial evidence for the design.

Critical weaknesses
  1. The adaptive-masking mechanism is not sufficiently justified under the stated implementation (§2.3; Appendix A.5; Tables 8–9). Proposition A.9 applies a fixed-game importance-sampling argument, but Equation (7) constructs the game using the empirical teacher average from the same masks reused in the regression. Detaching that average removes gradients, not statistical dependence; the displayed unbiased-objective argument therefore does not directly apply to the implemented reuse. Separately, adaptation is said to act on later visits to the same example, yet 250 steps with effective batch size 32 amount to 8,000 prompt draws from a 30,000-example pool. Under sampling without replacement, no example would receive an adaptive revisit. Repetition is possible under other schedules, but that schedule and the extent of reuse need to be specified. These issues do not invalidate multi-mask averaging, but they materially weaken the explanation and reproducibility of the adaptive component.

  2. The reported comparisons do not establish the reliability or computational value of the smaller gains (§3; Table 3; Appendix B.4). Each mathematics benchmark contains 30 problems, with 12 sampled answers per problem, and no problem-level uncertainty or training-seed variation is reported. Multiple answers to the same question do not supply 360 independent problem instances. Several component and baseline differences are only a few percentage points. Meanwhile, MASS is estimated to cost 2.60 times OPSD per step, and the computational comparison omits the closest multi-view baseline, AVSD. The consistent improvements are encouraging, but equal-step results do not isolate a better learning method at equal compute, and the precision of the smaller differences remains unclear.

Questions for authors
  1. How are training examples sampled over the 250 updates? Please report how many prompts were revisited and actually used nonuniform masking probabilities, and confirm whether each example's probabilities and attribution history persist between visits.
  2. For Proposition A.9, is the correction direction estimated independently of the regression masks in the actual implementation? If it uses the same masks, how should the guarantee be revised to account for the resulting dependence?
  3. How is the all-masked coalition value required by Equation (9) obtained, and are that evaluation and the full-reference anchor included in the reported FLOPs?
  4. What were AVSD's teacher-route count and training budget, and how were checkpoints and hyperparameters selected for Table 1? Please also clarify whether the reported values aggregate multiple training runs or only repeated decoding samples.
3 comments

We thank the reviewer for the careful technical reading and constructive comments. We address each point below, with pointers to the revised PDF (changes in blue).

Weakness 1 & Question 1: prompt revisits and persistence of masking states. The reviewer is right that sampling without replacement would give no revisits. We have now clarified (in §2.3) that our schedule uses 50% unseen and 50% previously visited prompts in every update after the first. Each example's qq and zz persist across visits: on a revisit they are restored, a fresh on-policy trajectory is generated, and masks are sampled from the restored qq. Table 6 (§3.3) answers the question on extent:

Statistic (250 steps)Count
Total prompt draws8,000
Unique prompts seen4,016
Unique prompts revisited2,001
Total revisit events3,984
Draws using non-uniform qq3,602

About 45% of draws use masking probabilities inherited from earlier visits, and Figure 8 (Appendix B.6) shows these probabilities becoming increasingly non-uniform over training, so the updated states actually change the teacher context. The resulting performance contribution is measured in the three-seed ablation (Table 4, §3.3).

Weakness 1 & Question 2: shared masks in Proposition A.9. We agree that stop-gradient removes gradients but not the statistical dependence between P^K\widehat P_K and the regression masks, so the original fixed-game argument did not cover the implementation. We reuse the masks to avoid doubling teacher passes. The revised Proposition A.9 (Appendix A.5) analyzes the implemented game vKv_K with this dependence included: for fixed β\beta, E[L^K(β)]=L∞(β)+O(K−1)\mathbb{E}[\widehat L_K(\beta)] = L_\infty(\beta) + O(K^{-1}) uniformly over bounded β\beta, and the full constrained estimator is consistent as K→∞K\to\infty. Mask reuse therefore introduces only an O(K−1)O(K^{-1}) objective bias.

Question 3: all-masked coalition value and FLOP accounting. The all-masked distribution p[J],tp_{[J],t} replaces every reference span with [MASK] and scores the same student prefixes as the anchor p0,tp_{0,t}; v([J])v([J]) is then computed from Eq. (7). Both passes are included in the reported FLOPs: each step counts KK masked teacher passes, one full-reference pass, one all-masked pass, the student rollout, and the backward pass.

Weakness 2: equal-compute comparison including AVSD, and precision of the gains. Table 10 (Appendix B.5) now includes AVSD and measured wall-clock time on Qwen3-4B:

MethodFLOPs / stepRelative FLOPsWall-clock time
OPSD1.31×10151.31\times10^{15}1.00×2 h 12 min
AVSD1.84×10151.84\times10^{15}1.40×2 h 37 min
MASS3.41×10153.41\times10^{15}2.60×2 h 49 min
RLSD8.39×10158.39\times10^{15}6.40×15 h 43 min

MASS's extra teacher evaluations are parallel prefill passes over one trajectory, so it trains only 12 minutes longer than AVSD and over 5× faster than RLSD. Figure 7 (Appendix B.5) places each method's accuracy gain against total training time: MASS obtains the largest gain at nearly the same cost as AVSD, so the improvement is not explained by additional compute. On precision, three training seeds (Table 7, Appendix B.1) give standard deviations of 0.42 to 1.11 points, and the paired per-problem bootstrap (Table 2, §3.2) places four of six MASS-versus-baseline intervals above zero, with the two smallest gains reported as unresolved.

Question 4: AVSD configuration, checkpoint selection, and aggregation. AVSD uses its published three-view configuration and training budget (Table 15, Appendix C); we did not add views because its views are heterogeneous and its authors report no gain from more views on mathematics. All methods follow one checkpoint rule (§3.1): checkpoints every 50 steps, reporting the best Avg@12, since a fixed final step would penalize early-peaking methods such as OPSD (Figure 3). On aggregation, each Table 1 entry is one training run averaged over 12 decoding samples per problem, whereas Table 7 and Table 4 aggregate three training runs. We will extend Table 7 to all models in Table 1 before the end of the discussion period.

We thank the reviewer again and are happy to provide further details.

Reviewer_G1sPOpenReview ↗

Improve the paper to weak accpetance

Thank you for the detailed response and substantial revisions. The clarified revisit schedule, persistence of qq and zz, and new usage statistics resolve my concern about whether adaptive masking is exercised in practice. The revised Proposition A.9 also directly acknowledges and analyzes the dependence caused by mask reuse. The FLOP accounting and AVSD comparison address my computational-cost questions.

The added three-seed and paired-bootstrap analyses materially improve the empirical support, while appropriately identifying the two smallest gains as unresolved. One clarification remains: are the “best Avg@12” checkpoints selected using a held-out validation set or the reported evaluation problems? Selection on the evaluation set could introduce optimism, particularly with single-run Table 1 results. Extending the multi-seed analysis to all Table 1 models, as proposed, would address the remaining uncertainty.

Overall, the rebuttal resolves my principal technical concerns and significantly strengthens the paper.

Thank you for the follow-up and for raising your assessment. Checkpoints are selected on the reported evaluation problems, not on a held-out set. However, to keep the comparison fair, the same selection rule is applied identically to every method, including all baselines, so no method receives a selection advantage over another. The multi-seed analysis has now been extended to all models in Table 1: the current revision now reports mean ± std over three training seeds for all four models: MASS has the highest mean in all 12 model–benchmark cells, and its lead on each model's three-benchmark average exceeds twice the standard deviation of the difference. Please let us know if there is anything else we can clarify.

Reviewer_E1JqOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces Multi-Mask Selective Self-Distillation (MASS), a new variant of OPSD that aims to stabilize long-term training and performance. In order to do, the paper constructs different variants of the masked reference answer before feeding a single answer to the teacher. Further, the paper proposes heuristics to filter out bad Empirical valuation show stabilized long-horizon training across multiple math benchmarks and models.

Critical strengths
  1. Novel and elegant method to control overfitting on a single reference answer.
  2. The method achieves remarkable stability and high overall performance, as evident in Figure 3 and Table 1.
  3. The method does not require additional sampling cost, improving its real world efficiency.
Critical weaknesses
  1. The approach, while improving stability and intuitively appealing, introduces too many hyperparameters in practice, such as mask ratio, span length, and confidence thresholds. It is unclear how sensitive the algorithm is to these hyperparameters.
  2. Empirical validation outside of math domains would further strengthen claims around generality of the method. Only OOD generalization results are shown on a few domains. What would be interesting is to also measure training these models on non-math domains (or both math + non-math)
Questions for authors

See weaknesses.

1 comment

We thank the reviewer for the positive assessment and helpful suggestions.

Weakness 1: number of hyperparameters and sensitivity. All models and families use one shared configuration (Table 12, Appendix C.2) without per-model tuning. For the hyperparameters you named:

  • Mask ratio (Figure 5(b), §3.3): accuracy follows a smooth inverted U peaking at ρ=0.3\rho=0.3 on both Qwen3-4B and OLMo-7B, and every tested value still outperforms OPSD.
  • Span length is not tuned: spans follow fixed structural boundaries such as punctuation and math delimiters (Appendix C.1).
  • Number of routes KK (Figure 5(a), §3.3): accuracy improves steadily with KK, so larger KK is a compute choice rather than a sensitive setting.
  • Confidence thresholds: MASS has none. The token weight in Eq. (5) is computed in closed form from mask statistics. The only bound, the masking-probability clip ϵ\epsilon, changes AIME25 accuracy by at most about one point across a 10× range (Table 8, Appendix B.2):
Modelϵ=0.01\epsilon=0.01ϵ=0.03\epsilon=0.03ϵ=0.05\epsilon=0.05ϵ=0.10\epsilon=0.10
Qwen3-8B79.0879.4479.1778.61
OLMo-7B74.6275.0074.7273.89

Weakness 2: training beyond mathematics. We agree this is valuable. Training on domains such as code would require domain-specific span segmentation, which we note as future work in §5. Within the current scope, Table 5 (§3.3) addresses generality from the evaluation side: after math-only training, OPSD and AVSD fall below the base model on GPQA and BBH for both model families, whereas MASS preserves or slightly improves base accuracy, so its gains do not come at the cost of general reasoning.

We thank the reviewer again for the encouraging evaluation.

Comments on the paper

4 comments
AuthorsOpenReview ↗

General Response to All Reviewers

We thank all reviewers for their careful and constructive feedback. We are encouraged that the reviewers found the method novel and elegant, the gains consistent across all evaluated model–benchmark settings without additional rollouts, and the separation of view diversity from finite-mask uncertainty well motivated. The manuscript has been revised in accordance with the reviewers' feedback. All revisions in the updated PDF are highlighted in blue. Below we summarize key points addressed in our response and revision; full details are provided in the responses to individual reviewers.

G1. Statistical reliability (hSvD W1, Q2; G1sP W2)

  • Three training seeds for all methods on Qwen3-4B (Table 7, Appendix B.1) and for the component ablation (Table 4, §3.3). Standard deviations (0.42 to 1.11 points) are smaller than the margin of MASS over the strongest baseline on AIME24 and HMMT25, and the ablation ordering holds across seeds. We will extend Table 7 to all models in Table 1 before the end of the discussion period.
  • Paired per-problem bootstrap intervals against AVSD and RLSD (Table 2, §3.2): four of six lie above zero; the two smaller gains remain unresolved at the problem level, and we state this in the paper.

G2. Checkpoint selection and baseline configurations (hSvD W2, Q1; G1sP Q4)

  • One reporting rule for all methods (§3.1): best Avg@12 across checkpoints evaluated every 50 steps.
  • Full AVSD and RLSD configurations, taken from each method's published main results (Table 15, Appendix C).

G3. Other concerns

  • K=1K=1 single-mask control: K=8K=8 clearly outperforms it (Figure 5(a), §3.3) (hSvD W3, Q4).
  • Incorrect-solution control: the incorrect-solution score stays near zero, verifying that multi-mask averaging does not make the teacher uniformly less strict (Figure 1, §1) (hSvD W4).
  • Pass@12: MASS leads by a wider margin than under Avg@12 (Table 11, Appendix B.7) (hSvD Q3).
  • Prompt revisits: the sampling schedule is specified (§2.3), and 45% of training draws use masking probabilities updated on earlier visits (Table 6, §3.3; Figure 8, Appendix B.6) (G1sP W1, Q1).
  • Proposition A.9 now analyzes the implemented shared-mask estimator (Appendix A.5) (G1sP W1, Q2).
  • Compute: AVSD added; MASS trains only 12 minutes longer than AVSD despite more teacher passes (Table 10, Figure 7, Appendix B.5) (G1sP W2, Q3).
  • Hyperparameters: one shared configuration for all models (Table 12, Appendix C.2), with smooth sensitivity to KK, ρ\rho (Figure 5, §3.3) and ϵ\epsilon (Table 8, Appendix B.2) (E1Jq W1).
  • Beyond mathematics: MASS preserves out-of-distribution accuracy on GPQA and BBH where OPSD and AVSD lose it (Table 5, §3.3); training on other domains is noted as future work (§5) (E1Jq W2).
AuthorsOpenReview ↗

request for feedback from reviewers

Dear reviewers,

Thank you again for your efforts in providing thoughtful reviews for our paper. With only a day left in the discussion period, we would welcome any questions or comments you might have on our revised paper and response. We would be happy to clarify any questions or concerns that remain.

We look forward to any feedback you might have.

Update: three-seed results for all models (Table 1)

As committed in G1, Table 1 now reports mean ± std over three training seeds for every method on all four models (Qwen3-1.7B, Qwen3-4B, Qwen3-8B, and OLMo-7B).

  • MASS has the highest mean in all 12 model–benchmark cells.
  • On the three-benchmark average, MASS leads the strongest baseline by 1.66 to 2.19 points on every model, more than three times the larger of the two standard deviations in each case.

Since Table 1 now covers the Qwen3-4B seeds, Table 7 (Appendix B.1) duplicates those rows. We keep it during the discussion so that the table numbers cited in our replies remain valid, and will remove it in the final version. This update also supersedes our earlier statement (reply to G1sP, Question 4) that Table 1 entries are single training runs.

AuthorsOpenReview ↗

Summary of the revision

We thank the reviewers again. Below we summarize, at a high level, how the revision changes the paper; details and pointers are in our individual replies.

The results are now shown to be robust, not single-run. All main results (Table 1) are now means over three training seeds for every model and baseline. MASS has the highest mean on all 12 model–benchmark pairs, and on each model's three-benchmark average its lead over the strongest baseline is more than twice the seed-to-seed variation. A problem-level bootstrap further shows which individual gains are clearly resolved. We state openly that the smallest ones (mainly AIME25) are not.

The evaluation is now fully transparent. All methods follow one checkpoint-reporting rule, and the complete baseline configurations are listed, so the comparison can be checked directly.

We now isolate why averaging several masked views helps. A single random mask improves briefly and then degrades, while averaging eight masks keeps improving throughout training. The multi-mask teacher is not simply less strict: its support for a different correct solution rises sharply while support for an incorrect solution stays near zero. Consistent with this, the advantage of MASS widens under Pass@12, meaning it solves a broader set of problems rather than the same ones more reliably.

The adaptive masking demonstrably operates. The revisit schedule is now specified, and about 45% of training draws use masking probabilities learned on earlier visits, which become increasingly non-uniform over training.

The theory now matches the implementation. The Shapley-attribution result is restated for the estimator we actually use, where the same masks serve both purposes; it shows only an O(1/K) bias and asymptotic consistency.

The method remains practical and general. MASS trains in nearly the same wall-clock time as AVSD and more than five times faster than RLSD, uses a single hyperparameter configuration for all models, and preserves out-of-distribution accuracy on GPQA and BBH where other self-distillation methods lose it.

Note: since Table 1 now includes the Qwen3-4B seeds, Table 7 (Appendix B.1) duplicates those rows. We keep it during the discussion so that table numbers cited in our replies stay valid, and will remove it in the final version. This also supersedes our earlier statement that Table 1 entries are single training runs.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.