acceptodds
Under review as a conference paper at ICLR 2027

Teach Yourself Where to Look: On-Policy Attention Self-Distillation for Reasoning

25 views

Abstract

On-policy self-distillation trains reasoning models on their own trajectories using dense token distribution guidance from a privileged teacher with access to a verified solution. This supervision transfers what the teacher predicts without directly transferring where it attends within the preceding context. We introduce On-Policy Attention Self-Distillation (OPASD), which complements token-level supervision with solution-conditioned attention distillation. Because the privileged teacher can attend to verified-solution tokens unavailable to the student, OPASD projects teacher attention onto student-visible positions and renormalizes the resulting distribution before alignment. Across three model sizes and four competition-level mathematics benchmarks, OPASD consistently outperforms token-only OPSD, improving average accuracy by 4.98 to 8.40 percentage points. OPASD also avoids the response-length inflation and performance degradation observed with token-only distillation, reducing generated rollout tokens by 73.9% and estimated model compute by 72.6% while training 1.53× faster. These results show that solution-conditioned attention provides a complementary supervision signal that makes on-policy self-distillation more accurate, stable, and compute-efficient.

Primary area:
foundation or frontier models, including LLMs
Keywords:
on-policy self-distillation, attention distillation, large language models, reasoning post-training, mathematical reasoning

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_wxJ2OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes On-Policy Attention Self-Distillation (OPASD), which extends OPSD by introducing attention-level distillation. The method uses attention distributions from a teacher conditioned on the reference solution to provide additional supervision for the student.

Experiments on Qwen3-1.7B, 4B, and 8B across AIME 2024, 2025, 2026, and HMMT 2025 show that OPASD improves mathematical reasoning accuracy over the original OPSD, while producing shorter responses, reducing training compute, and achieving more stable training.

Critical strengths

1. Good experimental results and ablation studies.

OPASD consistently outperforms the original OPSD across three Qwen3 model sizes. The paper also includes ablation studies on attention loss functions and distillation depth to examine the effects of different design choices.

2. Improved training stability and efficiency.

Compared with the original OPSD, OPASD alleviates the increase in reasoning length and the decline in accuracy during training, while also reducing training compute costs.

Critical weaknesses
  1. Insufficient comparisons with related methods.

The paper mainly compares OPASD with the original OPSD. Although the authors provide several ablation studies, comparisons with existing on-policy representation or attention distillation methods, such as OPRD and OISD, are missing. This makes it difficult to assess the advantages of the proposed attention distillation approach over related methods.

Questions for authors

Q1. Comparison with related methods.

Could the authors compare OPASD with existing on-policy representation or attention distillation methods, such as OPRD and OISD? If a direct comparison is not feasible due to differences in training setups, could the authors clarify how OPASD differs from these methods and what advantages it offers?

Q2. Generalizability of training efficiency.

The reported training speedup is interesting. Since the experiments use OpenThoughts-Math-30K for training, could the authors discuss whether the reduction in response length and training compute would also hold on other training datasets?

7 comments
AuthorsOpenReview ↗

Summary of Responses to Reviewer wxJ2's Comments (Part 1/3)

We thank the reviewer for the thoughtful and encouraging feedback. We appreciate their recognition of OPASD’s consistent accuracy gains over OPSD across three Qwen3 model sizes, the ablations on attention loss and distillation depth, and the improvements in training stability and efficiency through shorter responses, more stable accuracy, and lower training compute.

Summary of response

  1. We added comparisons with OPRD, SDFT, and OISD on four mathematical reasoning benchmarks.
  2. We extended the training efficiency evaluation to DeepMath-103K.
AuthorsOpenReview ↗

Response on Comparisons with Related Distillation Methods (Part 2/3)

Comparison with related methods. Could the authors compare OPASD with existing on-policy representation or attention distillation methods, such as OPRD and OISD? If a direct comparison is not feasible due to differences in training setups, could the authors clarify how OPASD differs from these methods and what advantages it offers?

We have expanded the evaluation to include SDFT [1], OPRD [2], and OISD [3]. We compare them with OPASD using Qwen3-4B on AIME24, AIME25, AIME26, and HMMT25 under comparable training and evaluation settings. As shown below, OPASD achieves the highest accuracy on all four benchmarks. On average, it improves over OPSD by 8.40 percentage points, SDFT by 2.29 points, OPRD by 3.05 points, and OISD by 4.96 points.

MethodAIME24AIME25AIME26HMMT25Avg.
OPSD51.1143.8846.1128.3342.36
OPRD58.6147.7852.5031.9447.71
SDFT60.8348.8953.0531.1148.47
OISD58.0646.6850.2728.1645.80
OPASD63.3350.2755.5633.8950.76

OPSD and SDFT distill next-token predictions from teachers conditioned on a reference solution and an expert demonstration, respectively. OPRD matches teacher and student hidden states along student-generated trajectories. The teacher’s attention provides direct guidance on where to attend, a signal that complements the prediction or representation targets used by these methods. OISD uses the model’s final layer to supervise next-token predictions and attention at an intermediate layer, using the problem and generated response as input. OPASD conditions the teacher on the privileged solution, projects its attention onto student-visible tokens, and distills that attention alongside next-token predictions.

References

[1] Shenfeld, Idan, et al. "Self-distillation enables continual learning." ICML 2026

[2] Yang, Shenzhi, et al. “OPRD: On-Policy Representation Distillation.” arXiv preprint arXiv:2606.06021 (2026).

[3] Liu, Xinyu, et al. “OISD: On-Policy Internal Self-Distillation of Language Models.” Findings of EMNLP 2026.

AuthorsOpenReview ↗

Response on Training Efficiency Gains Beyond OpenThoughts-Math-30K (Part 3/3)

Generalizability of training efficiency. The reported training speedup is interesting. Since the experiments use OpenThoughts-Math-30K for training, could the authors discuss whether the reduction in response length and training compute would also hold on other training datasets?

To assess whether the efficiency gains extend beyond OpenThoughts-Math-30K, we conducted additional experiments with Qwen3-4B using DeepMath-103K [1] as the training dataset. We compared OPSD and OPASD in response length, estimated model compute, wall-clock time, and reasoning accuracy.

On DeepMath-103K, OPASD maintains short responses and stable rollout accuracy throughout training, while OPSD’s responses grow longer and its rollout accuracy declines. OPASD guides the student’s attention toward the context emphasized by the solution-conditioned teacher, helping it preserve relevant information and avoid repeating earlier reasoning steps. OPASD generates 61.7% fewer response tokens than OPSD, reducing estimated model compute by 59.9% and cutting wall-clock training time from 5.20 to 2.83 hours, a 1.84× speedup. Peak allocated GPU memory also falls by 56.9%. OPASD improves accuracy on all four benchmarks, with an average gain of 5.56 percentage points over OPSD. These results show that OPASD’s training efficiency gains extend beyond OpenThoughts-Math-30K to DeepMath-103K.

MethodAIME24AIME25AIME26HMMT25Avg.
OPSD56.3847.5049.4429.1645.62
OPASD61.9455.0053.8933.8951.18

References

[1] He, Zhiwei, et al. "DeepMath-103K: A large-scale, challenging, decontaminated, and verifiable mathematical dataset for advancing reasoning." ICLR 2026

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

This paper proposes On-Policy Attention Self-Distillation (OPASD) which combines the attention distillation and on-policy self-distillation (OPSD). The method combines two learning branches, beginning with the standard OPSD where a model learns by setting itself as teacher with privileged solutions, and the first branch is token-level distillation and the second branch is attention-level distillation. The teacher's attention in the absence of solutions is defined as the normalized attention to the student's context. Experiments show that OPASD brings gains in accuracy as well as reduction in the amount of rollout tokens.

Critical strengths
  1. The idea of combining OPSD and attention distillation seems reasonable and effective.

  2. The method is well-presented and can be readily reproduced. The paper is clearly written and easy to follow.

  3. The performance gain as compared to the OPSD baseline seems significant over various model sizes.

Critical weaknesses
  1. Technical contribution. While the method is straightforward to understand and implement (a strength), the overall method is a simple combination of two well-known approaches: OPSD and attention distillation. The technical novelty and contribution appear somewhat limited from the viewpoint of ICLR standard.

  2. Limited baselines. There have been a number of recent contributions associated with OPSD and its improvements (e.g., SDFT [1] or SD-zero [2] or other works cited in this paper). However, this work compares OPASD only with vanilla OPSD. The lack of comparison makes it difficult to judge whether OPASD stands out from recent variants or parallels of OPSD.

  3. Theoretical justification. The paper proposes the student-support projection (target attention to learn from teacher) to be renormalizing the attention mass to the context available to student. At first, this approach sounds reasonable and is justified by minimizing KL to teachers attention (Theorem O.2), but is somewhat questionable with further thinking. When the teacher generates the attention weights from the answer, it is likely that the most of attention mass are distributed over the privileged solution. Thus, it is questionable that if the "residual" (remaining) attention to the context outside the privileged solutions contain meaningful information. It is likely that the overall mass to outside the answer is small, and its attention distribution may be noisy, but the student tries learn from this signal. I see from ablation study that there indeed is improvement in performance with the proposed attention distillation, However, I think there could be a better justification, e.g., why minimizing KL to "residual" attention in the absence of golden answer necessarily leads to a desirable target attention. It would be good if such justification is provided.

[1] Shenfeld, Idan, et al. "Self-distillation enables continual learning." ICML 2026

[2] He, Yinghui, et al. "Self-distillation zero: Self-revision turns binary rewards into dense supervision." arXiv preprint arXiv:2604.12002 (2026).

Questions for authors

Please see weaknesses. In addition,

  1. model architectures. The model used for evaluation are somewhat small (<10B). The work can benefit from examining whether the effectiveness of OPASD can be extended to larger models. Also the evaluation is limited only to Qwen models: although the authors have acknowledged it in the Limitation, but it would be good to provide reasons why other architectures are not evaluated.

  2. downstream tasks. It appears that the proposed method can be applied to various reasoning tasks, however in the paper only the math tasks are evaluated. Can the method be evaluated on other tasks like coding?

10 comments
AuthorsOpenReview ↗

Summary of Responses to Reviewer PxYC's Comments (Part 1/6)

We thank the reviewer for the thoughtful and encouraging feedback. We appreciate their recognition of the effectiveness of combining OPSD with attention distillation, the paper’s clarity, the method’s reproducibility, and OPASD’s accuracy gains over OPSD across model sizes alongside fewer rollout tokens.

Summary of response

  1. We clarify OPASD’s technical novelty and contributions.
  2. We compare OPASD with OPRD, SDFT, and OISD on four mathematical reasoning benchmarks.
  3. We extend the evaluation to a larger model, Qwen3-14B, and two other model families, DeepSeek-R1-Distill-Llama-8B and Olmo-3-7B.
  4. We evaluate code generation on the HumanEval+ and LiveCodeBench benchmarks.
  5. We measure how much teacher attention remains on student-visible tokens and explain why the projected attention provides a useful training signal.
AuthorsOpenReview ↗

Response on OPASD’s Technical Novelty and Contributions (Part 2/6)

Technical contribution. While the method is straightforward to understand and implement (a strength), the overall method is a simple combination of two well-known approaches: OPSD and attention distillation. The technical novelty and contribution appear somewhat limited from the viewpoint of ICLR standard.

OPSD distills the teacher’s next-token distribution along the student’s own reasoning trajectory. At each step, this distribution assigns probabilities to possible next tokens and tells the student what the teacher predicts. The teacher’s attention distribution assigns weights to preceding token positions and provides a complementary dense signal about where the teacher attends when forming that prediction. This matters in multi-step reasoning because later steps often depend on constraints in the problem and intermediate results from earlier steps. OPSD leaves this attention signal unused as a distillation target.

Transferring the teacher’s attention requires resolving a support mismatch. The teacher and student predict over the same vocabulary, but their attention distributions cover different sets of token positions. Both see the problem and the student-generated prefix, and the teacher also sees the privileged solution. OPASD uses student-support projection to remove solution-only positions and renormalize the remaining attention. This preserves the teacher’s relative attention over student-visible tokens. OPASD distills this solution-conditioned attention alongside next-token predictions on the same student-generated trajectory.

Our contributions are listed below.

  1. We add solution-conditioned attention supervision to OPSD. OPASD gives the student a direct attention target over the preceding context along the same trajectories used for next-token distillation. On Qwen3-1.7B, combining token and attention distillation achieves an average accuracy of 34.02, compared with 29.04 for token-only distillation and 30.97 for attention-only distillation. These results show that the two signals complement each other (Table 8).
  2. We resolve the support mismatch with student-support projection. We remove solution-only positions and renormalize the remaining attention, preserving the teacher’s relative attention over student-visible tokens. The target remains solution-conditioned because the teacher computes attention with access to the privileged solution. Theorem O.2 shows that the projection uniquely minimizes generalized Jensen-Shannon divergence to the full teacher attention among distributions supported on student-visible positions. Projection achieves an average accuracy of 34.02, compared with 31.39 for redistribution and 32.15 for shuffled redistribution (Table 6).
  3. We demonstrate improvements in accuracy, stability, and efficiency. Across three Qwen3 model sizes and four benchmarks, OPASD improves macro-average accuracy over OPSD by 4.98–8.40 percentage points. In the efficiency experiment, it generates 73.9% fewer rollout tokens, reduces estimated compute by 72.6%, and trains 1.53× faster (Figure 5). The training and validation curves show shorter responses and more stable accuracy. Smaller increases in epistemic markers such as “wait,” “maybe,” and “alternatively” are consistent with fewer repetitive uncertainty loops (Figure 4 and Appendix D).
AuthorsOpenReview ↗

Response on Comparisons with Recent Distillation Baselines (Part 3/6)

Limited baselines. There have been a number of recent contributions associated with OPSD and its improvements (e.g., SDFT [1] or SD-zero [2] or other works cited in this paper). However, this work compares OPASD only with vanilla OPSD. The lack of comparison makes it difficult to judge whether OPASD stands out from recent variants or parallels of OPSD.

We have expanded the evaluation to include SDFT [1], OPRD [2], and OISD [3]. We compare them with OPASD using Qwen3-4B on AIME24, AIME25, AIME26, and HMMT25 under comparable training and evaluation settings. As shown below, OPASD achieves the highest accuracy on all four benchmarks. On average, it improves over OPSD by 8.40 percentage points, SDFT by 2.29 points, OPRD by 3.05 points, and OISD by 4.96 points.

MethodAIME24AIME25AIME26HMMT25Avg.
OPSD51.1143.8846.1128.3342.36
OPRD58.6147.7852.5031.9447.71
SDFT60.8348.8953.0531.1148.47
OISD58.0646.6850.2728.1645.80
OPASD63.3350.2755.5633.8950.76

OPSD and SDFT distill next-token predictions from teachers conditioned on a reference solution and an expert demonstration, respectively. OPRD matches teacher and student hidden states along student-generated trajectories. The teacher’s attention provides direct guidance on where to attend, a signal that complements the prediction or representation targets used by these methods. OISD uses the model’s final layer to supervise next-token predictions and attention at an intermediate layer, using the problem and generated response as input. OPASD conditions the teacher on the privileged solution, projects its attention onto student-visible tokens, and distills that attention alongside next-token predictions.

References

[1] Shenfeld, Idan, et al. "Self-distillation enables continual learning." ICML 2026

[2] Yang, Shenzhi, et al. “OPRD: On-Policy Representation Distillation.” arXiv preprint arXiv:2606.06021 (2026)

[3] Liu, Xinyu, et al. “OISD: On-Policy Internal Self-Distillation of Language Models.” Findings of EMNLP 2026

Reviewer_iNu9OpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces On-Policy Attention Self-Distillation (OPASD), a novel post-training framework designed to transfer a privileged teacher's contextual focus to a student model during multi-step mathematical reasoning. Standard On-Policy Self-Distillation (OPSD) aligns next-token probability distributions, guiding what a model predicts but failing to guide where it allocates attention across its preceding context. This lack of attention alignment leads to severe response-length inflation, long repetitive loops of uncertainty, and training instability. To bridge this gap, OPASD integrates solution-conditioned attention distillation along the student’s on-policy trajectories. Because the privileged teacher handles reference solution text hidden from the student, the paper develops a student-support projection mechanism to resolve the resulting support mismatch. This projection strips away solution-only text positions and renormalizes the remaining attention mass over student-visible tokens, maintaining the teacher's exact relative attention layouts. Tested across three Qwen3 configurations (1.7B, 4B, 8B) on competitive math benchmarks (AIME 2024–2026, HMMT 2025), OPASD consistently outpaces token-only baselines. Crucially, it improves macro-average accuracy by 4.98 to 8.40 percentage points while slashing generated rollout tokens by 73.9%, dropping compute by 72.6%, and accelerating wall-clock training speed by 1.53×.

Critical strengths
  • Dual-Perspective Guidance: Combining standard token-level probability matching with explicit attention map alignment creates a robust supervisory footprint that targets the fundamental causal dependencies of multi-step thinking.
  • Pareto-Dominant Performance & Efficiency: Unlike traditional reasoning frameworks where increased accuracy demands longer generation steps, OPASD scales accuracy while drastically compacting generations (73.9% token reduction) and dropping estimated FLOP costs by 72.6%.
  • Mitigation of Uncertainty Loops: The behavioral analysis neatly proves that training the student's attention anchors prevents it from falling into repetitive, long-winded self-doubt traps, as evidenced by a massive reduction in epistemic markers like wait, maybe, and alternatively.
  • Mathematical Rigor: The mathematical appendix provides a complete closed-form proof (Theorem O.2) verifying that the student-support projection is the unique distribution that minimizes the generalized Jensen-Shannon divergence to the full teacher attention matrix.
  • High Structural Stability: The ablation experiments demonstrate exceptionally tight confidence intervals across random seeds (±0.14\pm 0.14 standard deviation) and stable accuracy plateaus across expanding test-time generation token limits.
Critical weaknesses
  • Final-Layer Truncation: The attention distillation objective is restricted entirely to the final transformer layer. While multi-layer objectives were ablated and found to degrade performance, matching only the final layer overlooks hierarchical context-refinement tracking in intermediate blocks.
  • Static Teacher Baseline: The privileged teacher model is completely frozen at its starting checkpoint throughout the 200 training steps. This lack of online teacher progression limits the fluid evolution of target representations as the student's sampled trajectories shift.
  • Homogeneous Domain and Family: The evaluation is bound entirely to competition mathematics within a single model architecture branch (Qwen3) . The cross-domain performance of OPASD remains unverified on soft-reasoning or language-heavy contexts (e.g., source code comment synthesis, planning, or legal instruction following).
Questions for authors
  1. Renormalization Under Extreme Overlap (ρt→1\rho_t \to 1): When the teacher concentrates almost all of its attention mass on the hidden privileged solution tokens, the remaining mass distributed over student-visible tokens shrinks significantly . In these high-privilege conditions, does the projected attention distribution target (A˜tT\~{A}_{t}^{T}) become highly unstable or vulnerable to random floating-point noise?
  2. Causal Factors Behind Multi-Layer Degradation: Why does scaling the attention loss depth across the final 2 or 3 layers systematically degrade average performance (dropping from 34.02 to 31.04) ? Is this decrease due to structural semantic drift, or are intermediate layers fundamentally more susceptible to teacher-student capacity mismatches?
  3. Cross-Tokenizer Applicability: Because attention support maps onto ordinal context position indices rather than literal vocabulary strings , can this exact structural projection layer be adapted to cross-distillation frameworks where the teacher and student utilize entirely separate tokenizers?
7 comments
AuthorsOpenReview ↗

Summary of Responses to Reviewer iNu9's Comments (Part 1/6)

We thank the reviewer for the thoughtful and encouraging feedback. We appreciate their recognition of the complementary roles of token and attention supervision, the gains in reasoning accuracy and training efficiency, and the behavioral evidence of fewer repetitive uncertainty loops. We also appreciate their recognition of the theoretical justification for student-support projection and the consistent results across random seeds and test-time generation budgets.

Summary of our response

  1. We extended the evaluation to code generation on HumanEval+ and LiveCodeBench and to two additional model families using DeepSeek-R1-Distill-Llama-8B and Olmo-3-7B.
  2. We added an EMA teacher comparison for OPSD and OPASD.
  3. We explained how final-layer supervision guides learning in earlier layers and why matching additional layers may constrain the student’s intermediate processing.
  4. We measured how much attention the teacher assigned to the privileged solution to assess whether the remaining attention mass was small enough to raise concerns about numerical stability.
  5. We explained why the current projection cannot be applied when the teacher and student use different tokenizers.
AuthorsOpenReview ↗

Response on Generalization Across Tasks and Model Families (Part 2/6)

Homogeneous Domain and Family: The evaluation is bound entirely to competition mathematics within a single model architecture branch (Qwen3). The cross-domain performance of OPASD remains unverified on soft-reasoning or language-heavy contexts (e.g., source code comment synthesis, planning, or legal instruction following).

Following OPSD’s experimental setup, our original experiments focused on mathematical reasoning using Qwen3-1.7B, Qwen3-4B, and Qwen3-8B. We extended the comparison to an additional domain and two other model families, using the same training and evaluation protocol for OPSD and OPASD within each setting.

Additional domain. We evaluated code generation on HumanEval+ and LiveCodeBench using Qwen3-4B. OPASD improves over OPSD by 4.27 percentage points on HumanEval+ and 2.27 points on LiveCodeBench. These gains show that attention-level supervision complements the token-level objective in code generation as well as in mathematical reasoning.

MethodHumanEval+LiveCodeBench
OPSD84.7561.80
OPASD89.0264.07

Other model families. We compared OPASD with OPSD on DeepSeek-R1-Distill-Llama-8B and OLMo-3-7B. Across four mathematical reasoning benchmarks, OPASD improves average accuracy over OPSD by 2.99 points on DeepSeek-R1-Distill-Llama-8B and 5.40 points on Olmo-3-7B. The gains on both models show that adding attention distillation to next-token supervision improves reasoning accuracy beyond the Qwen family.

ModelMethodAIME24AIME25AIME26HMMT25Avg
DeepSeek-R1-Distill-Llama-8BOPSD34.1629.1629.1716.6727.29
DeepSeek-R1-Distill-Llama-8BOPASD40.5631.9428.8919.7230.28
Olmo-3-7BOPSD56.6144.4450.8330.9445.71
Olmo-3-7BOPASD61.9449.1758.8934.4451.11
AuthorsOpenReview ↗

Response on the Static Teacher Baseline (Part 3/6)

Static Teacher Baseline: The privileged teacher model is completely frozen at its starting checkpoint throughout the 200 training steps. This lack of online teacher progression limits the fluid evolution of target representations as the student's sampled trajectories shift.

We added an EMA-teacher comparison for OPSD and OPASD using Qwen3-4B. Both methods used the same training and evaluation settings and an EMA decay of 0.95. With an EMA teacher, OPASD outperforms OPSD on all four benchmarks, improving average accuracy by 8.59 points. This shows that projected attention continues to provide useful solution-conditioned guidance alongside next-token supervision as the teacher evolves during training.

MethodAIME24AIME25AIME26HMMT25Avg
OPSD50.3844.5247.7328.1542.70
OPASD62.6750.5656.6135.3151.29

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal Responses

We thank the Area Chair and reviewers iNu9, PxYC, and wxJ2 for their careful and constructive reviews. We addressed all questions and concerns in our individual replies and updated the manuscript with additional experiments and analysis in Appendices P–T.

For reviewer iNu9, we extended the evaluation to DeepSeek-R1-Distill-Llama-8B and Olmo-3-7B (Appendix P), evaluated code generation on HumanEval+ and LiveCodeBench (Appendix Q), and added an EMA-teacher comparison (Appendix T). We also explained final-layer supervision and multi-layer degradation, examined the extreme-overlap concern through an attention-mass analysis, and clarified cross-tokenizer applicability.

For reviewer PxYC, we clarified OPASD’s technical novelty and contributions, added comparisons with OPRD, SDFT, and OISD (Appendix R), evaluated one larger model (Qwen3-14B) and two additional model families (DeepSeek-R1-Distill-Llama-8B and Olmo-3-7B) (Appendix P), and evaluated code generation (Appendix Q). We also analyzed the retained teacher attention and clarified the justification for student-support projection.

For reviewer wxJ2, we added comparisons with OPRD, SDFT, and OISD (Appendix R) and extended the training-efficiency evaluation to DeepMath-103K (Appendix S).

Reviewers acknowledged in the discussion that their concerns had been addressed. Their feedback has made the paper clearer and stronger, and we thank them and the Area Chair for their 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.