acceptodds
Under review as a conference paper at ICLR 2027

Overcoming Scaling Limits in On-Policy Self-Distillation for LLM Reasoning

12 views

Abstract

On-policy self-distillation (OPSD) improves mathematical reasoning by training a student to match a privileged teacher distribution at each prefix of the student’s own sampled trajectory. In standard OPSD, the teacher is conditioned on a privileged context, typically a reference solution, but the distillation loss is applied along an unverified student rollout. This formulation couples two inputs with distinct roles: the sampled trajectory serves as a supervision scaffold, determining the prefixes at which supervision is applied, while the privileged context determines the teacher distribution at those prefixes. We separate these two factors in a factorial analysis and find that whether the scaffold reaches the correct final answer has a much stronger effect on downstream accuracy than whether the context is correct. When distillation uses unverified scaffolds, the teacher can rely on information unavailable to the student, creating an imitation gap. This gap decreases with larger model sizes, yet OPSD still allocates most supervision to unverified trajectories. In contrast, verified scaffolds provide supervision that the student can realistically imitate, even when the context comes from the student’s own unsuccessful attempt. Based on this finding, we introduce OASIS (On-policy Alignment via Scaffold-Isolated Supervision), which retains the OPSD objective but limits supervision to the student’s own verified trajectories and replaces the written reference solution with a failed attempt generated by the model. This method provides a scaffold within the student’s trajectory distribution, allowing OASIS to use only final-answer labels without requiring worked solutions. Across Qwen3-1.7B, 4B, and 8B on AIME 2024, AIME 2025, and HMMT 2025, OASIS improves over the base model by 3.2 to 3.8 points on average, while the benefit from OPSD drops from 3.05 points at 1.7B to 0.14 at 8B. At 8B, OASIS outperforms OPSD by 3.05 points, showing that restricting self-distillation to verified, on-policy scaffolds maintains its effectiveness as models scale.

Primary area:
foundation or frontier models, including LLMs
Keywords:
on-policy self-distillation, mathematical reasoning, large language models, verified trajectories, scaffold-context decoupling

Reviews

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

This paper separates the roles of trajectory scaffolds and teacher contexts in OPSD and introduces OASIS, which distills on short outcome-verified scaffolds using self-generated contexts. Experiments across three Qwen3 sizes show gains over OPSD without written reference solutions, at increased rollout-generation cost.

Critical strengths
  1. Separating trajectory scaffolds from teacher contexts provides a useful diagnostic perspective on OPSD. Comparisons at fixed prefixes show how privileged teacher advantages vary with trajectory outcomes, informing the choice of where to apply distillation supervision.

  2. OASIS removes the need for written reference solutions while improving reasoning accuracy across three Qwen3 sizes. These results support its practical value in settings with outcome verification but no reference solutions.

Critical weaknesses

OASIS combines additional sampling, verified-scaffold selection, and self-generated teacher contexts, but their contributions are not sufficiently separated. A control using the same selected scaffolds and training budget without the additional context is needed to establish its added value beyond successful-trajectory selection.

Questions for authors

Could you add an ablation without the self-generated teacher context, keeping the selected scaffolds and training budget unchanged? This would clarify whether the context contributes beyond successful-trajectory selection.

5 comments
AuthorsOpenReview ↗

Ablation of Self-Generated Teacher Context and Component Contributions

Thanks for your thoughtful feedback.

We conducted the requested ablation by removing the self-generated teacher context while keeping the selected verified scaffolds, training budget, and hyperparameters unchanged. We evaluated both settings on Qwen3-4B over three independent runs.

Table 1: Ablation of self-generated teacher context in OASIS

BenchmarkMetricOASIS w/contextOASIS w/o contextΔ (w/o context - w/context)
AIME24Avg@1276.674.4−2.2
Pass@1286.690.0+3.4
Maj@1281.180.0−1.1
AIME25Avg@1269.169.7+0.6
Pass@1286.683.3−3.3
Maj@1280.076.7−3.3
HMMT25Avg@1245.244.4−0.8
Pass@1263.363.30.0
Maj@1256.646.7−9.9

All values are percentages averaged over three independent runs. Negative Δ indicates lower performance without the context.

These results show that OASIS retains most of its Avg@12 performance without self-generated teacher context, suggesting that verified-scaffold selection accounts for the majority of the overall performance gains. Nevertheless, the context provides additional benefits in majority-vote accuracy (Maj@12), with improvements of 1.1, 3.3, and 9.9 percentage points on AIME24, AIME25, and HMMT25, respectively. Thus, while verified scaffolds are the primary contributor to OASIS's performance, self-generated teacher context provides complementary benefits, particularly in improving the consistency of correct reasoning trajectories.

So, The ablation shows that removing the teacher context has limited impact on overall performance. We further analyzed the contribution of individual components. We found that the main performance gains come from using verified scaffolds (Table 8, Page 19-20). We also examined the effect of the teacher's thinking mode and found that it affects OPSD but has no noticeable impact on OASIS (Table 10-11, Page 20-21).

Thank you for adding the teacher-context ablation. My remaining concern is whether OASIS offers enough benefit to justify its sampling cost at scale. In the reported 1.7B setting, OASIS generates about 16 times as many rollout tokens as standard OPSD for a 0.59 percentage-point gain in average accuracy. Although this ratio does not measure total training cost, it shows substantial additional generation overhead for a modest accuracy gain. The reported results do not establish that this trade-off is worthwhile.

I would find it informative to compare OASIS with SFT on the same verified trajectories and report the total compute for both methods, including the cost of generating and filtering those trajectories. The results show that trajectory selection matters, but it remains unclear how much distillation adds once the correct trajectories have already been selected, or whether that added benefit would justify the cost for larger models.

AuthorsOpenReview ↗

Distillation Benefits and Computational Efficiency

Thank you for your thoughtful feedback and continued engagement. We have conducted additional experiments to address your remaining concerns.

1. Distillation improves performance beyond trajectory selection

Since OASIS selects the shortest verified trajectory for supervision, we investigate whether its gains come from trajectory selection alone or from the distillation objective. We compare standard SFT on written gold solutions with rejection-sampling SFT* , which uses the exact same shortest verified trajectories as OASIS. For both SFT* and OASIS, we generate 8 rollouts per problem and apply identical filtering and training steps.

Table 1. SFT and OASIS comparison on Qwen3-1.7B.

MethodTraining supervisionThinking AIME24 Avg@12Output tokensNon-thinking AIME24 Avg@12Non-thinking HMMT25 Avg@12
SFTWritten gold solutions48.4072911.675.83
SFT*Hard labels on verified trajectories48.6161011.395.83
OASISTeacher distributions on the same trajectories54.6161022.7810.28

SFT* performs almost identically to gold-solution SFT. This is consistent with the observation in Zhao et al. (2026) that SFT on concise reasoning traces can limit test-time generation length and reasoning performance. Therefore, SFT* is unlikely to provide substantial gains from these concise self-generated trajectories, consistent with findings in prior work. However, OASIS improves over SFT* by 6.00 points in thinking mode, around 11 points on AIME24 and 4.45 points on HMMT25 in non-thinking mode. Both methods generate the same average number of tokens for student during training. This shows that the improvement comes from token-level distillation.

These results are reported in Appendix A.4, Table 15, pages 21–23 of the revised paper.

2. Practical Rollout Cost, Training Efficiency, and Performance Gains on Smaller Models

OASIS generates more candidate tokens than OPSD, but as you also noted, these tokens are used to identify verified trajectories. Only the selected trajectories are used for student training. The additional generation also does not translate directly into the same increase in wall-clock time because rollouts are generated in parallel batches.

Table 2. Rollout generation cost and non-thinking AIME24 performance on Qwen3-1.7B.

MethodRollouts (KK)Generated tokensSupervised tokensLatency per batch (s)AIME24 Avg@12
SFT———0.2011.67
OPSD12.76M2.76M3.7118.61
OASIS422.8M2.01M4.1721.94
OASIS843.5M1.98M5.7022.78

Table 3: Comparison under a non-thinking evaluation on Qwen3-1.7B (Avg@12, %). The teacher operates with thinking mode enabled, while the student operates with thinking mode disabled. Best result per column in bold.

MethodAIME24HMMT25Average
OPSD18.618.3313.47
OASIS (ours)22.7810.2816.53

On a single NVIDIA H100, increasing the rollout budget from 1 to 8 increases batch latency by only 1.54× (3.71s to 5.70s), despite the much larger generated-token count. Reducing the budget from 8 to 4 nearly halves the generated tokens and lowers latency to 4.17s, while retaining similar accuracy (21.94% compared with 22.78% on non-thinking AIME24). The 4-rollout setting also achieves 52.77% on thinking-mode AIME24. We also notice that although the average gap between OPSD and OASIS is only 0.59% in thinking mode for 1.7b model, it jumps to 3.06% (13.47% vs. 16.53%) on the same model when thinking is turned off during evaluation

There is also a difference in how the rollouts are used. GRPO generates eight rollouts per problem and uses all of them for policy optimization. In our setup, GRPO requires 500 training steps, whereas OASIS requires only 100. OASIS uses only the selected verified trajectory for student distillation. The other candidates are discarded after selection, so they do not need to be stored across iterations or used for student backpropagation.

Reviewer_5XZWOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper identified a weakness in on-policy self-distillation (OPSD) where the privileged teacher’s advantage, which comes from seeing a reference solution the student lacks, is concentrated on the student’s failed or truncated trajectories, yet about 70% of OPSD’s updates fall on exactly those trajectories, creating an imitation gap the student can’t close and contributing to OPSD’s benefit shrinking from 3.05 points at Qwen3-1.7B to just 0.14 at 8B. The authors propose OASIS, which keeps the OPSD loss unchanged but samples K rollouts per problem, supervises only along the shortest verified-correct one, and replaces the written reference solution with another of the student’s own same-problem rollouts, usually a failed attempt, so only final-answer labels are needed. Across AIME 2024, AIME 2025, and HMMT 2025, OASIS improves over the base model by roughly 3.2 to 3.8 points at all three model sizes and beats OPSD by 3.05 points at 8B.

Critical strengths
  • The paper separates two roles that standard OPSD bundles together: the scaffold (which prefixes get supervised) and the context (what the teacher sees). This framing is simple, clarifying, and leads directly to the method.
  • The recovery and signal-replay experiments are inference-only, so they cost little, and they are well controlled.
  • Proposition 1 shows the forward-KL optimum under hidden context is a context-averaged mixture with irreducible loss equal to the conditional mutual information. This ties OPSD to the learning-with-privileged-information and imitation-gap literature, and it predicts the rising student entropy seen in training.
Critical weaknesses
  • The teacher signal may be mostly thinking-mode distillation. The student template has thinking disabled and the teacher has it enabled, and Figure 2a attributes 55–64% of the KL gap to the template alone, with only 17–30% from reference content. Figure 2b also shows template-only and empty contexts sit about as close to the gold teacher as the student’s own failed rollout. This raises the possibility that OASIS is largely distilling thinking-mode behavior into non-thinking mode on correct trajectories.
  • Checkpoints are evaluated every 20 steps with “the same checkpoint selection rule,” but the paper doesn’t say whether selection uses the test benchmarks themselves.
  • The experiments are a bit narrow. The experiments use one model family (Qwen3), one domain (competition math), and one training dataset.
Questions for authors

Please refer to the weaknesses

4 comments
AuthorsOpenReview ↗

Experiments on Thinking-Mode Independence, Checkpoint Selection

Thank you for the thoughtful and constructive review. Your insightful comments helped us strengthen our experimental evaluation and improve the clarity and completeness of the paper. We address the reviewer's three concerns with additional controlled experiments.

1. Is OASIS mainly benefiting from thinking-mode distillation?

To answer this, we conducted two ablations on Qwen3-4B.

Ablation A: Disabling teacher thinking during training. We kept the student training configuration unchanged (thinking OFF) and evaluated with thinking ON. Disabling teacher thinking consistently reduces OPSD performance, but has little effect on OASIS. OASIS even improves slightly on AIME25 and HMMT25.

Table 1A: OPSD with and without teacher thinking

BenchmarkTeacher Think ONTeacher Think OFFDifference
AIME2474.473.4−1.0
AIME2567.966.4−1.5
HMMT2543.141.1−2.0

Table 1B: OASIS with and without teacher thinking

BenchmarkTeacher Think ONTeacher Think OFFDifference
AIME2476.675.6−1.0
AIME2569.169.7+0.6
HMMT2545.245.3+0.1

Ablation B: Removing thinking mode from the entire pipeline. We disabled thinking for both teacher and student during training and evaluation. OASIS still outperforms OPSD across all three benchmarks and all three metrics.

Table 2: Comparison in the non-thinking setting

BenchmarkMethodAvg@12Pass@12Maj@12
AIME24OPSD23.156.726.7
OASIS36.473.353.3
Gain+13.3+16.6+26.6
AIME25OPSD20.660.030.0
OASIS31.166.746.7
Gain+10.5+6.7+16.7
HMMT25OPSD10.626.716.7
OASIS17.843.330.0
Gain+7.2+16.6+13.3

These results show that OASIS's gains do not depend on distilling thinking-mode behavior. OASIS also outperforms OPSD when thinking mode is disabled throughout the pipeline. We have added these results to Section 4 of the paper.

2. Checkpoint selection

In the original experiments, we evaluated checkpoints at Steps 20, 40, 60, 80, and 100, then selected the checkpoint with the highest macro-averaged Avg@12 across the evaluation benchmarks. We followed the same protocol as OPSD (Zhao et al., 2026) to ensure a consistent comparison.

To check whether our results depend on this selection, we evaluated all newly added experiments using only the final checkpoint, without any checkpoint selection. These include the non-thinking evaluation (Table 2), Olmo-3-7B-Instruct (Table 3), and the Biology and Medical benchmarks (Table 4). OASIS outperforms OPSD in all these evaluations.

We have clarified the checkpoint-selection protocol in Section 5 (page 8) and added the new selection-free results to the revised manuscript.

AuthorsOpenReview ↗

Generalization

3. Generalization across model families and domains

We extended our experiments beyond Qwen3 and competition math. Specifically, we evaluated OASIS on Olmo-3-7B-Instruct, Qwen2.5-7B-Instruct, and two additional domains: Biology and Medicine.

Cross-family generalization. On Olmo-3-7B-Instruct, OASIS achieves higher Avg@12 than OPSD on all three math benchmarks.

Table 4: OASIS vs. OPSD on Olmo-3-7B-Instruct

BenchmarkMetricOPSDOASISGain
AIME24Avg@1243.3350.83+7.50
Pass@1283.3383.330.00
Maj@1256.6760.00+3.33
AIME25Avg@1237.7840.00+2.22
Pass@1270.0070.000.00
Maj@1246.6753.33+6.66
HMMT25Avg@1220.0023.89+3.89
Pass@1240.0050.00+10.00
Maj@1230.0030.000.00

Cross-domain generalization. We evaluated OASIS on SciKnowEval and MMLU-Pro for biology, and MedQA and MedMCQA for medical reasoning. OASIS outperforms OPSD across all eight model--benchmark settings. Notably, it achieves these gains using fewer training samples in every setting.

Table 5: Generalization to Biology and Medical Reasoning

DomainModelBenchmarkBase Accuracy (%)OPSD SamplesOASIS SamplesOPSD Accuracy (%)OASIS Accuracy (%)
BiologyQwen2.5-7B-InstructSciKnowEval20.012,8003,92041.043.0
MMLU-Pro71.312,8003,92071.672.8
Qwen3-4B-InstructSciKnowEval36.012,8001,92044.045.0
MMLU-Pro84.512,8001,92084.986.8
MedicalQwen2.5-7B-InstructMedQA46.53,2002,33660.763.0
MedMCQA44.23,2002,33655.756.4
Qwen3-4B-InstructMedQA72.79,6002,62470.673.8
MedMCQA60.29,6002,62460.160.4

All experiments use the final checkpoint without benchmark-based checkpoint selection. Both teacher and student operate in non-thinking mode.

Overall, these experiments show that OASIS's gains are not dependent on thinking-mode distillation, are also observed without checkpoint selection in the new evaluations, and extend beyond Qwen3 and competition math. We have incorporated these results and clarifications into the revised manuscript.

I thank the authors for the detailed reply. I've increased my score.

We are glad that the additional clarifications and experiments have addressed your concerns. We appreciate your updated rating and your recognition of the contributions of our work.

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

This paper studies on-policy self-distillation (OPSD) for mathematical reasoning by separating two factors coupled in the standard formulation: the student trajectory scaffold, which determines where supervision is applied, and the additional teacher context, which determines the target distribution. The authors propose OASIS, which generates multiple on-policy candidates per problem, selects the shortest successful trajectory as the supervision scaffold, and uses another self-generated trajectory for the same problem as teacher context. The method requires only final-answer labels and outperforms OPSD across three Qwen3 model sizes and three competition-mathematics benchmarks.

Critical strengths
  1. Separating supervision locations from teacher context provides a valuable analytical perspective. Rather than relying solely on final accuracy comparisons, the paper combines distribution replay and truncated-continuation experiments to investigate the sources and effects of teacher signals across trajectories. This provides useful insight into why a stronger context-conditioned teacher does not necessarily yield more effective training supervision.
  2. The method is simple and reduces dependence on worked solutions. Without redesigning the distillation objective, OASIS achieves improvements across three model sizes through final-answer verification and self-generated context, with larger gains at 4B and 8B. This has practical value for datasets with verifiable answers but no complete reasoning annotations.
Critical weaknesses
  1. The paper’s motivation does not sufficiently distinguish imitation difficulties caused by privileged information from alternative explanations, including trajectory completeness, stopping behavior, length preferences, and changes in the difficulty distribution of training problems.
  2. Proposition 1 establishes that, under its assumptions, the optimal student matches the teacher distribution averaged across contexts. However, this does not directly establish that failed trajectories are less valuable for distillation. An inability to perfectly match the teacher distribution does not imply an inability to learn useful error-correction capabilities.
Questions for authors
  1. Are there controlled experiments distinguishing completed but incorrect trajectories from trajectories truncated at the generation limit? When training problems and teacher context are held fixed, and trajectory lengths are matched as closely as possible, do successful trajectories retain their advantage? This would help determine whether OASIS benefits primarily from complete solutions and stopping signals or from answer correctness itself.
  2. In Table 1, on Failed prefixes, Gold reference improves continuation success over Answer only from 61% to 77% at 1.7B and from 59% to 67% at 4B. This suggests that the additional reasoning steps provide substantial assistance and could constitute valuable error-correction signals for distillation. Why interpret this observation as evidence of supervision that is difficult to imitate rather than an opportunity to learn error correction? Could the smaller gains on Verified prefixes reflect a ceiling effect, while failed trajectories offer greater room for learning through distillation?
  3. Related-work suggestion (not a basis for my rating): The subsequently released work Not Every Rollout Needs Distillation: Correctness-Filtered On-Policy Self-Distillation (arXiv:2610.25842 ) also investigates trajectory selection in OPSD, but reports that distilling a subset of incorrect responses can maintain performance while reducing training cost. This work became available after the present paper, so its absence is not a citation-related concern. If accepted, I encourage the authors to cite and discuss it in the camera-ready version.
1 comment
AuthorsOpenReview ↗

Controlled Experiments on Scaffold Selection and Error Recovery

Thank you for the thoughtful review. We appreciate your positive assessment of separating the student scaffold from the teacher context. We address the two concerns with additional experiments.

1. Does scaffold correctness matter beyond completeness and length? (Weakness 1, Question 1)

We agree that our original analysis did not separate correctness from trajectory completeness, stopping behavior, length, and changes in training-problem difficulty. We therefore conducted a controlled experiment on AIME24 using Qwen3-1.7B (thinking) and Qwen3-4B (non-thinking). We fixed the training problems, teacher context, and training budget, varying only the scaffold selection strategy.

ModelScaffold RuleAvg. Supervised LengthAvg@12 (%)Pass@12 (%)Maj@12 (%)
Qwen3-1.7B (Thinking)Truncated (limit)102434.17 ± 0.4270.00 ± 1.6753.33 ± 1.92
Shortest wrong (completed)68238.06 ± 0.2878.33 ± 1.6763.00 ± 0.00
Longest correct (completed)90952.33 ± 0.5173.33 ± 1.9260.00 ± 1.92
Shortest correct (OASIS)60354.17 ± 0.5880.00 ± 0.0068.88 ± 1.57
Qwen3-4B (Non-thinking)Truncated (limit)102427.7856.6736.67
Longest correct (completed)81535.8373.3356.67

The results show that completed but incorrect trajectories perform substantially worse than correct trajectories of comparable length. For Qwen3-1.7B, shortest correct achieves 54.17 Avg@12, compared with 38.06 for shortest wrong. Even longest correct achieves 52.33 despite being considerably longer. We observe the same pattern with Qwen3-4B. These results indicate that scaffold correctness plays an important role beyond trajectory completeness and length. In particular, completed but incorrect trajectories perform substantially worse than verified-correct trajectories, even when their supervised lengths are comparable.

2. Does Training on Failed Trajectories Improve Error Recovery? (Weakness 2, Question 2)

We agree that failed trajectories may provide useful error-correction signals. However, their usefulness may also depend on other factors, such as trajectory correctness, length, and completeness, which our controlled experiments show can affect training performance.

To examine whether training on failed trajectories improves downstream recovery, we conducted a post-training recoverability experiment. OPSD is trained on approximately 70% failed trajectories with gold solutions as privileged teacher context, whereas OASIS uses exclusively verified scaffolds and mostly self-generated, potentially incorrect privileged contexts (approximately 80%). If training on failed trajectories provides a stronger recovery capability, we would expect OPSD to perform better in this test.

We collected failed rollouts from base Qwen3-4B on AIME24, truncated them at 25% and 75% of their original lengths, and evaluated both trained models on the same prefixes.

Failed prefix lengthOPSDOASISOASIS gain
25%55.4361.96+6.53
75%58.7063.29+4.59

OASIS achieves higher recovery rates at both prefix lengths, despite not using failed trajectories as training scaffolds. This suggests that exposure to failed trajectories alone does not guarantee better recovery. Their effectiveness may depend on how the trajectories are selected and how the teacher's privileged context interacts with them. We will clarify this distinction in the revised discussion.

3. Related Work (Question 3)

Thank you for bringing this relevant work to our attention. We will cite and discuss Not Every Rollout Needs Distillation (arXiv:2610.25842) in the camera-ready version.

Comments on the paper

1 comment

We sincerely thank all reviewers for their thorough and constructive feedback, which has significantly improved our manuscript. We have carefully addressed all the concerns raised and substantially revised our paper accordingly. All changes are highlighted in blue text in the revised manuscript. Below, we summarize the main additions and revisions.

Summary of the Revisions

We have isolated the contributions of scaffold selection and teacher context by:

  • Comparing truncated, completed-incorrect, and completed-correct scaffolds under controlled settings, showing the effect of scaffold selection on downstream reasoning performance (Appendix A.4, Table 8, page 19-20) [Suggested by Reviewer sTgD]
  • Measuring the contribution of self-generated teacher context by removing it while keeping verified-scaffold selection, training budget, and optimization settings fixed (Appendix A.4, Table 9, page 20) [Suggested by Reviewer CpUn]

We have tested the generalization of OASIS across model families, domains, and thinking configurations by:

  • Evaluating OASIS on Olmo-3-7B-Instruct to examine cross-family generalization (Section 5.2, Table 3, page 9) [Suggested by Reviewer 5XZW]
  • Extending evaluation to Biology and Medical Reasoning using Qwen2.5-7B-Instruct and Qwen3-4B-Instruct across SciKnowEval, MMLU-Pro, MedQA, and MedMCQA (Section 5.2, Table 4, page 9) [Suggested by Reviewer 5XZW]
  • Investigating the effect of teacher thinking mode during training and evaluating a fully non-thinking pipeline with both teacher and student operating without thinking mode (Appendix A.4, Tables 10–11, page 21) [Suggested by Reviewer 5XZW]

We have strengthened the empirical analysis and methodological clarity of OASIS by:

  • Comparing post-training recovery from identical failed reasoning prefixes at 25% and 75% truncation levels, showing improved recovery with OASIS (Appendix A.4, Figure 8, page 21) [Suggested by Reviewer sTgD]
  • Clarifying the checkpoint-selection protocol and reporting cross-family and non-thinking ablation results using final step-100 checkpoints without benchmark-based checkpoint selection (Section 5.1, page 8; Tables 3–4, page 9) [Suggested by Reviewer 5XZW]
  • Expanding the related-work discussion to include recent correctness-filtered on-policy self-distillation work and clarifying how OASIS differs from C-OPSD (Section 2, page 3) [Suggested by Reviewer sTgD]

Together, these revisions provide a more detailed understanding of the individual contributions of scaffold selection and teacher context, while strengthening the empirical validation of OASIS across model families, reasoning settings, and domains.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.