acceptodds
Under review as a conference paper at ICLR 2027

On-policy Instruction Tuning through Few-Shot Self-Distillation

Abstract

Pretrained large language models are typically adapted into assistants through an initial phase of supervised fine-tuning (SFT) on instruction-response demonstrations. Recent work suggests that the on-policy nature of reinforcement learning can help preserve capabilities acquired during previous post-training phases, motivating the question of whether instruction tuning can similarly benefit from on-policy training. We investigate this question by introducing Few-Shot Self-Distillation (FSSD), a method for on-policy instruction tuning of pretrained base models. FSSD conditions a teacher copy of the base model on a small set of relevant demonstrations and uses the teacher to provide token-level supervision on trajectories generated by the student. The teacher does not receive the reference response for the current instruction; instead, few-shot conditioning provides the supervision used to train the student. Across three model scales, FSSD helps mitigate regressions on general-capability benchmarks relative to standard SFT, while achieving stronger open-ended assistant performance on AlpacaEval 2.0 and Arena-Hard v2 and inducing substantially less distributional shift, as measured by the KL divergence from the pretrained policy.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Instruction Tuning, On-Policy Distillation, Self-Distillation, Few-Shot Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_HwDSOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Few-Shot Self-Distillation (FSSD) for on-policy instruction tuning of pretrained language models. A frozen copy of the base model, conditioned on task-relevant demonstrations, provides token-level supervision along student-generated responses without accessing the current instruction’s reference answer. Experiments on Qwen3 models from 1.7B to 8B show improved assistant performance, fewer regressions on general-capability benchmarks, and smaller distributional shifts from the pretrained model compared with standard SFT.

Critical strengths
  • FSSD uses the base model’s in-context learning capability to construct a teacher without requiring a separate, stronger model. Conditioning on related demonstrations provides a practical alternative to reference-answer conditioning.
  • The paper investigates demonstration count, category aware selection, reference-answer conditioning, self-generated SFT data, and fixed versus current policy trajectories. These comparisons help clarify the contributions of the method’s components.
Critical weaknesses
  • All experiments are restricted to Qwen3 models with at most 8B parameters and use LoRA. The constrained parameter updates under LoRA may themselves limit distributional changes, making it unclear whether the observed capability preservation would persist under full-parameter fine-tuning. The current results therefore provide insufficient evidence of generalization across model families or applicability to full-parameter fine-tuning of larger models.
  • Replacing reference responses with few-shot demonstrations may introduce additional practical constraints. Obtaining multiple relevant, high-quality examples from the same task category can be more demanding than obtaining a reference answer for a given instruction, particularly for specialized or novel tasks. Furthermore, a frozen base model may lack the capabilities needed to solve the task even with these demonstrations, making its token-level supervision potentially unreliable. The paper does not investigate these limitations sufficiently to establish whether the approach can scale to a broader range of tasks.
Questions for authors

Have you assessed the reliability of the few-shot-conditioned teacher, particularly on tasks where the base model has low accuracy? Is there evidence that FSSD remains beneficial in these cases rather than reinforcing the teacher’s errors, and how does its performance compare with SFT using the available reference answers?

3 comments

We thank the reviewer for their constructive feedback and for the opportunity to address their concerns.

W1: LoRA and model generalization. Our choice of Qwen3-1.7B, 4B, and 8B follows the experimental setup of [1], which evaluates the same model family and scales using LoRA. This allows us to investigate FSSD in an established experimental setting, while extending the study from instruction-tuned models to pretrained base models. To address the reviewer's concern regarding full-parameter fine-tuning, we additionally conducted experiments on Qwen3-1.7B using full-parameter fine-tuning instead of LoRA. As expected, the results are consistent with our main findings, and the main conclusions remain unchanged. In fact, without LoRA, we observe even more pronounced capability degradation. Thus, these results further support our decision to employ high-rank LoRA as a computationally efficient alternative to full-parameter fine-tuning.

[1] Zhao, Siyan, et al. "Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models." arXiv preprint arXiv:2601.18734 (2026).

Table A. Open-ended assistant performance with Full-FT. Scores (%) are mean ± SEM.

ModelMethodAlpacaEval 2.0Arena-Hard v2
Qwen3-1.7BBase5.77±0.7928.8±1.1
SFT3.01±0.6830.6±1.0
FSSD7.99±0.9535.3±1.2

Table B. General-capability performance under full-FT. Scores (%) are mean ± SEM; Avg is the mean across eight benchmarks. ARC-C = ARC Challenge, Math-H = Math-Hard, MMLU-P = MMLU-Pro, WG = WinoGrande, HE = HumanEval.

ModelMethodARC-CBBHGPQAMath-HMuSRMMLU-PWGHE
Qwen3-1.7BBase37.6±1.434.2±0.628.6±1.330.3±1.239.9±1.715.4±0.358.7±1.452.4±3.9
SFT39.0±1.434.7±0.628.5±1.313.1±1.240.1±1.718.3±0.361.2±1.464.0±3.9
FSSD40.6±1.434.1±0.626.3±1.331.5±1.242.5±1.716.1±0.362.4±1.463.4±3.8

W2: Demonstration requirements. FSSD does not require additional annotated data beyond standard SFT. It reuses existing instruction-response pairs as demonstrations, sampled from the training dataset. Moreover, our ablations show that even a single demonstration yields substantial improvements, and that FSSD remains effective with randomly selected demonstrations (36.0 on Arena-Hard v2 versus 39.2 with category-aware selection; Table 4). This suggests that category-matched demonstrations are beneficial but not strictly necessary.

Q1: Teacher reliability and low-accuracy tasks. While we do not directly measure teacher accuracy conditional on base-model failures, our results provide evidence that low base-model accuracy does not preclude benefits from FSSD. For example, on Math-Hard, Qwen3-4B improves from 39.5 (base) to 45.6 with FSSD, whereas SFT decreases performance to 33.9. Similar improvements are observed at 1.7B and 8B.

Thank you for the clarifications and additional experiments. The full-parameter fine-tuning results partially address my concern that the benefits might depend on LoRA. However, in Table B, the clearest evidence for capability preservation is concentrated on Math-Hard. On the other benchmarks, the differences between FSSD and SFT are generally modest or mixed, with SFT performing comparably or better in several cases. These results therefore do not yet establish a broadly consistent advantage, and generalization beyond the evaluated Qwen3 family remains uncertain.

Nevertheless, I appreciate the authors’ efforts to address the concerns. The clarification that demonstrations reuse existing SFT data, together with the ablations on demonstration selection, helps alleviate the concern about additional data requirements. Considering these responses and the new full-parameter results, I am willing to modestly increase my score, while retaining the above reservations.

We thank the reviewer for their follow-up and for acknowledging the additional experimental evidence.

To directly address the reviewer's remaining concern regarding generalization beyond the Qwen3 model family, we conducted additional experiments using Ministral-3-3B. The results show that FSSD's improvements in open-ended assistant performance and reduced distributional shift extend to this model family as well.

ModelMethodAlpacaEval 2.0Arena-Hard v2KL div
Ministral-3-3BBase5.91±0.8730.8±1.10.000±0.000
SFT4.97±0.8033.8±1.20.083±0.002
FSSD7.86±0.9738.7±1.10.023±0.001

These results provide further evidence that FSSD's benefits extend beyond the Qwen3 model family.

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

This paper proposes Few-Shot Self-Distillation (FSSD), an on-policy instruction-tuning method that conditions a frozen teacher on task-relevant few-shot demonstrations and distills its token-level supervision into a student model. Experiments across three Qwen3 model scales show that FSSD improves open-ended assistant performance over SFT and existing on-policy self-distillation methods, while reducing capability degradation and distributional shifts from the pretrained model.

Critical strengths

The paper addresses an important and practical problem of improving instruction tuning while preserving pretrained capabilities.

The proposed method is simple and intuitive, combining few-shot in-context learning with on-policy self-distillation.

The experiments demonstrate consistent improvements across different model scales and multiple benchmarks.

The ablation studies on teacher conditioning, demonstration selection, and on-policy training provide useful supporting evidence.

Critical weaknesses

FSSD may primarily distill behavioral preferences induced by few-shot demonstrations into model parameters. However, the paper does not sufficiently investigate whether these preferences generalize to unrelated tasks or introduce negative transfer.

The methodological novelty is relatively limited, as FSSD mainly extends existing on-policy self-distillation by replacing reference-response conditioning with few-shot demonstrations.

FSSD requires approximately 2.5× the FLOPs of standard SFT, but compute-matched comparisons are missing, making its practical advantages less convincing.

Questions for authors

1.Do the behavioral preferences induced by few-shot demonstrations transfer to unrelated tasks, and could they negatively affect downstream performance?

2.Can the authors provide compute-matched or KL-matched SFT baselines to better isolate the benefits of FSSD?

3.To what extent does FSSD improve underlying model capabilities rather than simply internalizing behaviors elicited through in-context learning?

2 comments

We thank the reviewer for their constructive feedback. We appreciate the opportunity to address your concerns and clarify our work.

W1 & Q1: Cross-Task Generalization and Negative Transfer. We agree that negative transfer is an important issue, and indeed one of our central motivations for developing an on-policy methods to replace SFT. We believe that the catastrophic forgetting ablation in Table 3 already answers the reviewer concern in this regard, as it shows evidence of positive transfer to unrelated tasks. Specifically, when training exclusively on Information Seeking or Coding & Debugging, FSSD improves Math-Hard performance from 30.3 (base) to 32.5 and 32.1, respectively. In contrast, SFT reduces performance to 24.0 and 24.7, which suggest negative transfer that our method can successfully mitigate. While we cannot rule out negative transfer in all settings, our experiments show no substantial negative transfer.

W2: Methodological Novelty. Our novelty consists in adapting on-policy distillation for on-policy instruction tuning of pretrained base models with a new method meant to replace standard off-policy SFT. Few-shot demonstrations allow us to solve a concrete limitation of existing self-distillation approaches, without needing to rely on a separately trained teacher. We have further emphasized these novelty aspects in the revised manuscript.

W3 & Q2: Compute-Matched Comparison We agree that controlling for training compute is important. To address this, we conducted additional experiments with a compute-matched FSSD variant (FSSD-MC), trained by reducing the number of training steps to approximately match the computational cost of SFT, while keeping all other hyperparameters unchanged. As shown below, the main conclusions of the paper remain unchanged. These results indicate that FSSD's advantages cannot be attributed solely to its additional computational cost.

Table A. Open-ended assistant performance under matched training compute. Scores (%) are mean ± SEM.

ModelMethodAlpacaEval 2.0Arena-Hard v2
Qwen3-1.7BBase5.77±0.7928.8±1.1
SFT5.23±0.7932.7±1.2
FSSD-MC7.07±0.8835.3±0.9
FSSD8.72±0.9939.2±1.1
Qwen3-4BBase11.27±1.0561.4±1.3
SFT8.37±0.9962.4±1.0
FSSD-MC13.04±1.1467.0±1.2
FSSD16.39±1.2770.3±1.1
Qwen3-8BBase11.70±1.0466.3±1.4
SFT12.99±1.1570.2±1.2
FSSD-MC14.93±1.1971.5±0.9
FSSD15.31±1.1872.2±1.0

Table B. General-capability performance under matched training compute. Scores (%) are mean ± SEM; Avg is the mean across eight benchmarks. ARC-C = ARC Challenge, Math-H = Math-Hard, MMLU-P = MMLU-Pro, WG = WinoGrande, HE = HumanEval.

ModelMethodARC-CBBHGPQAMath-HMuSRMMLU-PWGHEAvg
Qwen3-1.7BBase37.6±1.434.2±0.628.6±1.330.3±1.239.9±1.715.4±0.358.7±1.452.4±3.937.1
SFT37.7±1.431.3±0.627.2±1.332.1±1.238.6±1.715.5±0.359.9±1.453.6±3.937.0
FSSD-MC39.2±1.434.4±0.628.8±1.332.0±1.240.3±1.715.7±0.359.8±1.454.3±3.838.1
FSSD41.2±1.434.8±0.627.3±1.333.3±1.240.3±1.716.6±0.361.8±1.462.2±3.839.7
Qwen3-4BBase41.8±1.442.5±0.630.9±1.339.5±1.241.2±1.816.3±0.364.9±1.363.4±3.842.6
SFT39.8±1.441.5±0.628.1±1.333.9±1.238.3±1.715.1±0.365.5±1.373.7±3.442.0
FSSD-MC41.5±1.443.6±0.630.4±1.343.2±1.241.5±1.716.2±0.365.8±1.368.3±3.443.8
FSSD42.3±1.443.4±0.630.6±1.345.6±1.241.0±1.816.2±0.366.2±1.371.9±3.544.6
Qwen3-8BBase42.9±1.444.4±0.632.7±1.437.7±1.239.6±1.817.8±0.371.5±1.373.7±3.445.0
SFT43.0±1.444.6±0.629.7±1.340.2±1.237.0±1.717.2±0.371.1±1.378.6±3.245.2
FSSD-MC43.8±1.444.6±0.632.4±1.346.5±1.238.9±1.717.7±0.372.1±1.376.8±3.246.6
FSSD43.9±1.444.6±0.632.8±1.448.8±1.239.0±1.717.8±0.371.9±1.379.2±3.247.2

Q3: Eliciting vs. Learning Capabilities. We would argue that internalizing capabilities elicited through in-context learning is precisely the objective of FSSD, rather than a limitation. Our goal is not to introduce new knowledge during instruction tuning, but to enable pretrained models to reliably express their existing capabilities without relying on off-policy demonstrations or requiring few-shot examples at inference time.

Thank you for the compute-matched comparison. It addresses my concern about the role of additional training compute. I also appreciate the clarification that Table 3 shows FSSD preserves Math-Hard performance after training on information-seeking or coding tasks. I still have concerns about how broadly this finding generalizes, particularly whether behaviors learned from demonstrations could hurt performance on other tasks. I therefore maintain my weak reject rating.

Reviewer_6GRGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Few-Shot Self-Distillation (FSSD) for instruction tuning pretrained language models. A frozen copy of the base model serves as a teacher, conditioned on demonstrations sampled from the same task category as the current instruction. The student generates responses on-policy and learns through token-level forward-KL distillation. Experiments use Qwen3 base models at 1.7B, 4B, and 8B, trained with LoRA on 50k instruction-response examples. FSSD improves reported assistant benchmark scores over SFT and OPSD, and preserves more performance on general-capability benchmarks and produces smaller KL divergence from the base model.

Critical strengths
  1. The topic is worth investigating. The approach shows consistent improvements across the three evaluated model sizes on most of the benchmarks. The experiment covarage are worth commending. The results show that the method could indeed improve assistant behavior against SFT across different tasks.
  2. FSSD does not require a separately trained, stronger teacher. It turns few-shot contextual guidance into a training signal for instruction tuning. The resulting student internalizes this guidance and operates without demonstrations at inference time, providing a practical way to establish persistent assistant behavior.
Critical weaknesses
  1. The methodological contribution is incremental. Internalizing contextual demonstrations has been explored in lot of works. Distilling a context-conditioned teacher on student trajectories is also established by [1]. The framework modification compared to OPSD is also limited. So the main contribution could only be regarded as adapting and validating those ideas for initial instruction tuning, limiting the novelty of this work.
  2. FSSD selects demonstrations by sampling from pre-annotated task categories. While category-based sampling itself is cheap, processing the additional context and generating on-policy student responses introduce extra training cost. Appendix A also reports approximately 2.5× the FLOPs of SFT. Table 5 reports nearly identical KL divergence for Off-FSSD and FSSD, but general-capability results for Off-FSSD are not provided. It therefore remains unclear whether on-policy sampling offers additional capability-retention benefits and whether the observed performance gains justify its extra training cost.
  3. Although Algorithm 1 samples demonstrations from the same task category and excludes the current example, it does not specify whether sampling is uniform or uses additional selection criteria, or how the selected demonstrations are ordered. These details are needed to reproduce the teacher’s conditioning context, which directly determines FSSD’s supervision signal.

[1] Ye, Tianzhu, et al. "On-policy context distillation for language models." arXiv preprint arXiv:2602.12275 (2026).

Questions for authors
  1. Please clarify the methodological differences between FSSD and prior on-policy context distillation methods such as [1]. An explicit comparison identifying which components are new and what they contribute beyond adapting existing techniques to initial instruction tuning would be great.
  2. Please report Off-FSSD’s results on the general-capability benchmarks and compare its training cost with FSSD. Given their nearly identical KL divergence in Table 5, I would like to see whether on-policy sampling provides additional capability-retention benefits and whether its performance gains justify the extra computation under matched budgets.
  3. Within each task category, are demonstrations sampled uniformly or according to additional criteria? How are the selected demonstrations ordered in the teacher’s context?
3 comments

We thank the reviewer for their constructive feedback. We appreciate the opportunity to address your concerns, clarify some points, and make the case for the paper's significance beyond an incremental contribution.

W1 & Q1: Methodological Novelty. We agree and acknowledged in the introduction that FSSD builds on established on-policy context-distillation techniques, including Ye et al. [1]. Our novelty consists in adapting these techniques for on-policy instruction tuning of pretrained base models with a new method that replaces standard off-policy SFT. Unlike Ye et al. [1], who focus on internalizing accumulated experience or optimized system prompts, FSSD leverages the few-shot learning capabilities of pretrained base models to induce assistant behavior. Few-shot demonstrations allow us to solve a concrete limitation of existing self-distillation approaches, without needing to rely on a separately trained teacher. We have further emphasized these novelty aspects in the revised manuscript.

W2: Computational Cost and Matched-Compute Comparison. We agree that computational cost should be considered. To address this concern, we conducted additional experiments with a compute-matched variant (FSSD-MC), trained by reducing the number of training steps to approximately match the computational cost of SFT, while keeping all other hyperparameters unchanged. As shown below, the main conclusions of the paper remain unchanged.

Table A. Open-ended assistant performance under matched training compute. Scores (%) are mean ± SEM.

ModelMethodAlpacaEval 2.0Arena-Hard v2
Qwen3-1.7BBase5.77±0.7928.8±1.1
SFT5.23±0.7932.7±1.2
FSSD-MC7.07±0.8835.3±0.9
FSSD8.72±0.9939.2±1.1
Qwen3-4BBase11.27±1.0561.4±1.3
SFT8.37±0.9962.4±1.0
FSSD-MC13.04±1.1467.0±1.2
FSSD16.39±1.2770.3±1.1
Qwen3-8BBase11.70±1.0466.3±1.4
SFT12.99±1.1570.2±1.2
FSSD-MC14.93±1.1971.5±0.9
FSSD15.31±1.1872.2±1.0

Table B. General-capability performance under matched training compute. Scores (%) are mean ± SEM; Avg is the mean across eight benchmarks. ARC-C = ARC Challenge, Math-H = Math-Hard, MMLU-P = MMLU-Pro, WG = WinoGrande, HE = HumanEval.

ModelMethodARC-CBBHGPQAMath-HMuSRMMLU-PWGHEAvg
Qwen3-1.7BBase37.6±1.434.2±0.628.6±1.330.3±1.239.9±1.715.4±0.358.7±1.452.4±3.937.1
SFT37.7±1.431.3±0.627.2±1.332.1±1.238.6±1.715.5±0.359.9±1.453.6±3.937.0
FSSD-MC39.2±1.434.4±0.628.8±1.332.0±1.240.3±1.715.7±0.359.8±1.454.3±3.838.1
FSSD41.2±1.434.8±0.627.3±1.333.3±1.240.3±1.716.6±0.361.8±1.462.2±3.839.7
Qwen3-4BBase41.8±1.442.5±0.630.9±1.339.5±1.241.2±1.816.3±0.364.9±1.363.4±3.842.6
SFT39.8±1.441.5±0.628.1±1.333.9±1.238.3±1.715.1±0.365.5±1.373.7±3.442.0
FSSD-MC41.5±1.443.6±0.630.4±1.343.2±1.241.5±1.716.2±0.365.8±1.368.3±3.443.8
FSSD42.3±1.443.4±0.630.6±1.345.6±1.241.0±1.816.2±0.366.2±1.371.9±3.544.6
Qwen3-8BBase42.9±1.444.4±0.632.7±1.437.7±1.239.6±1.817.8±0.371.5±1.373.7±3.445.0
SFT43.0±1.444.6±0.629.7±1.340.2±1.237.0±1.717.2±0.371.1±1.378.6±3.245.2
FSSD-MC43.8±1.444.6±0.632.4±1.346.5±1.238.9±1.717.7±0.372.1±1.376.8±3.246.6
FSSD43.9±1.444.6±0.632.8±1.448.8±1.239.0±1.717.8±0.371.9±1.379.2±3.247.2

Q2: On-Policy Sampling and Capability Retention. Thank you for the opportunity to clarify the relation between the Off-FSSD baseline and FSSD. As requested, we ran general capabilities benchmark of Off-FFSD. Analogously to FSSD, Off-FSSD requires both generating student trajectories. Off-FSSD is therefore not strictly off-policy as it samples from the initial student policy; but it's distinct from FSSD in that it does not sample new rollouts as the student is updated. Thus, the two methods share the costs of generating initial responses and computing teacher supervision, although "Off-FSSD" avoids repeated rollout generation. In other words, Off-FSSD is an ablation study, whose purpose is to evaluate the benefit of few-shot teacher conditioning without the online component of continually refreshing student trajectories. As expected, we do not observe substantial difference on general capability performance with respect to FSSD, which is consistent with our previous observation of nearly identical KL divergence from the pretrained policy. Yet, FSSD shows improved open-ended generation, which confirms the benefits of on-policy sampling extending beyond simply maintaining proximity to the pretrained policy.

To make sure that the subtleties of the distinction between Off-FSSD and FSSD are not overlooked, we opted for renaming Off-FSSD to Fixed-FSSD in the revised manuscript.

Table A. General-capability performance. Scores (%) are mean ± SEM; Avg is the mean across eight benchmarks. ARC-C = ARC Challenge, Math-H = Math-Hard, MMLU-P = MMLU-Pro, WG = WinoGrande, HE = HumanEval.

ModelMethodARC-CBBHGPQAMath-HMuSRMMLU-PWGHE
Qwen3-4BBase41.8±1.442.5±0.630.9±1.339.5±1.241.2±1.816.3±0.364.9±1.363.4±3.8
Off-FSSD42.2±1.443.6±0.630.7±1.346.2±1.240.6±1.716.4±0.367.0±1.471.9±3.9
FSSD42.3±1.443.4±0.630.6±1.345.6±1.241.0±1.716.2±0.366.2±1.471.9±3.9

W3 & Q3: Demonstration Sampling and Ordering. In the revision, we clarify that demonstrations are sampled uniformly at random without replacement from the same pre-annotated task category, excluding the current example. They are presented to the teacher in the order sampled, without additional ranking or ordering criteria.

We thank the reviewer again for their feedback, which has helped us highlight the methodological contribution, and we hope that the additional experimental results and clarifications address their concerns.

Reviewer_6GRGOpenReview ↗

Post-Rebuttal Acknowledgment

Thank you to the authors for the thoughtful rebuttal, the revised framing of the paper's claims, and the additional experiments (including the compute-matched FSSD-MC comparison, the Fixed-FSSD ablation, and the additional model-family evaluation). These additions clarify several aspects of the empirical behavior and better align the claims with the results. Nonetheless, my core reservations regarding the incremental methodological novelty relative to prior on-policy context/self-distillation approaches and the overall cost-benefit trade-off remain. Therefore, I will maintain my current rating of 2 (Weak Rejection).

Comments on the paper

1 comment

Dear AC and Reviewers,

We sincerely thank all reviewers for their time, thoughtful feedback, and valuable insights. Their comments have helped us clarify our methodological contributions and improve the overall presentation of the paper.

We have revised the manuscript to sharpen its motivation and better align our claims with the empirical evidence. In particular, we (1) reframed the motivation around leveraging pretrained models' few-shot learning capabilities for on-policy instruction tuning, (2) softened the claims regarding capability retention, and (3) reorganized the presentation to emphasize improved open-ended assistant performance and smaller distributional shift relative to SFT.

We also plan to incorporate the additional experimental results presented in our rebuttal into the camera-ready version.

We believe these changes considerably strengthen the presentation of our work, and highlight its significance in the context of the related literature.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.