acceptodds
Under review as a conference paper at ICLR 2027

Beyond Selection: Recycling Noisy Real-World Preference Data for Post-Training

Abstract

Preference data based on real-user queries capture how people use language models in various tasks and contexts. Many examples, however, have limited value for post-training in their original form. Data-quality pipelines often approach this with selection: retain useful examples and discard the rest. We study a complementary question: can discarded examples instead be recycled to provide more useful supervision? Our framework, ReSource, leverages noisy queries as sources for new tasks, allowing the task to change while keeping meaningful connections to the underlying content. We train a source-grounded instruction generator and use the quality of induced responses as a proxy for instruction utility to guide both training and instruction selection. In comparisons of recycled supervision against raw supervision using the same source examples, we find that recycling consistently improves downstream performance across three benchmarks and different amounts of training data. We further investigate its value when mixed with retained data and when used to train other models. Overall, our findings suggest that the utility of an example for post-training in its original form does not fully determine its value as a source of supervision, motivating recycling as a complement to data selection.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Data Recycling, Preference Data, Synthetic Data, Data Selection, LLM Post-Training

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_W9YYOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether real-user queries discarded by preference-data filtering can still provide useful material for new training tasks. RESOURCE trains a source-grounded instruction generator with self-training and response-aware GRPO, searches over candidate instructions using response rewards, and constructs fresh preference pairs for downstream DPO. Source-matched experiments compare raw and recycled supervision across data scales, test mixtures with retained data, and examine instruction transfer to other target models.

Critical strengths
  1. The problem is practical. A discarded query can still contain useful code, context, or task material, so recycling is a reasonable complement to filtering. The paper also clearly distinguishes constructing a new task from recovering the user's original intention.

  2. The source-matched comparisons are a strong part of the design. Raw and recycled conditions use the same sources, training-set size, and preference-construction procedure. Table 1 shows consistent RESOURCE improvements across three benchmarks and all four tested data scales, making the main empirical result fairly convincing.

  3. The study goes beyond the raw-versus-recycled comparison. The mixture results and cross-model instruction transfer are useful, and Table 4 separates contributions from recycler training and instruction search. Together they support the method as a usable pipeline rather than just a few successful examples.

Critical weaknesses
  1. The comparisons match the final training data, but not the total construction cost. RESOURCE uses up to eight instruction candidates with four response probes each, followed by sixteen fresh responses, and also requires SFT and GRPO for the recycler. RAW only needs the final preference-construction stage. The reported gains are useful, but the practical value would be clearer with total generation/training costs and a stronger use of the same construction budget for the baseline.

  2. The necessity of the discarded source content is not isolated. The task can change substantially, and the validity checks mainly establish eligibility rather than semantic grounding or usefulness. There is no clear source-removed or source-shuffled control showing how much of the gain comes from meaningful source reuse rather than general synthetic-task generation and reward-based search. This limits the specific conclusion about recycling, even if the resulting data are useful.

  3. The evaluation leaves some uncertainty. The three benchmarks share a local Llama-3.1-70B judge, and the main construction pipeline also uses Llama models. This does not invalidate the comparisons, but independent judging would strengthen them. In addition, Tables 1–4 report standard errors without making repeated downstream-training variability clear enough. Evaluation uncertainty and training-seed uncertainty should be distinguished.

Questions for authors
  1. What are the measured construction and recycler-training costs, and over how much recycled data are these costs amortized? Please use existing measurements if available.

  2. Do you have an existing source-removed or source-shuffled control, or a systematic grounding audit? If not, can the claim be narrowed to the demonstrated utility of the generated data rather than the necessity of source reuse?

  3. Do the reported standard errors include independent downstream DPO runs, or only evaluation variation? Please also clarify whether any comparisons have already been checked with an independent judge.

1 comment
AuthorsOpenReview ↗

Response to Reviewer W9YY

We appreciate the reviewer’s thoughtful comments and feedback on our paper!

The comparisons match the final training data, but not the total construction cost. ...

What are the measured construction and recycler-training costs, and over how much recycled data are these costs amortized? …

We agree that recycling incurs additional construction cost. Our question is whether each discarded example can somehow become useful by investing in one-time recycler training and subsequent generation. We do not claim recycling is more compute-efficient than using these examples directly.

Here are the measured costs for the one-time training and construction, in NVIDIA A40 GPU-hours.

StageGPU-hours
Recycler training, one-time (SFT + GRPO)143 (8 GPU x 17.9 hours)
Instruction generation and search, 2K62 (8 GPU x 7.8 hours)
Preference construction, 2K (also required for Raw)35 (8 GPU x 4.4 hours)

The same trained recycler is reused across our experiments without retraining. Its one-time training cost is therefore amortized over 8K sources, corresponding to approximately 18 GPU-hours per 1K sources. As such, recycling trades additional construction compute for better downstream performance from the same discarded sources (e.g., 46.20 to 54.86 in AlpacaEval2). We will make this trade-off explicit in the revision to help assess its practical value. Thank you!

The necessity of the discarded source content is not isolated. …

Do you have an existing source-removed or source-shuffled control, or a systematic grounding audit?

Following your suggestion, we additionally evaluated general synthetic data generation, including source-free synthesis and self-instruct-style generation using discarded queries as seeds. All methods use the same amount of 2K data, the same response generation configuration and preference construction.

Method (seed pool)AlpacaEval2 LCArena-HardWildBench
Synthetic (discarded)50.53 ± 0.4670.67 ± 1.2146.61 ± 0.94
Synthetic (source-free)43.29 ± 0.4769.83 ± 1.0343.95 ± 0.98
ReSource (discarded)54.86 ± 0.4573.93 ± 1.1049.12 ± 0.95

The results above clearly show the gap between using discarded queries as seeds/sources and source-free synthesis. This highlights the usefulness of discarded source content for recycling.

The three benchmarks share a local Llama-3.1-70B judge, and the main construction pipeline also uses Llama models. … independent judging would strengthen them.

Do the reported standard errors include independent downstream DPO runs, or only evaluation variation? …

Thank you for the suggestions. We additionally evaluated with an independent GPT-4o judge (gpt-4o-2024-11-20) using the same 2K-trained models. ReSource outperforms Raw across all three benchmarks under this judge as well. In addition, even with GPT-4.1 (gpt-4.1-2025-04-14), the same trend holds although the absolute scales differ across judges: AlpacaEval2 LC increases from 22.35 for Raw to 25.19 for ReSource.

MethodAlpacaEval2 LCArena-HardWildBench
Raw38.98 ± 0.1138.28 ± 1.2337.43 ± 0.92
ReSource41.62 ± 0.1243.77 ± 1.3041.75 ± 0.95

The SE in the paper reflects evaluation uncertainty. To further verify, we additionally ran five downstream training runs per setting using multiple seeds. The table below shows AlpacaEval2 LC under the original local judge, with mean ± SE across five runs. We will explicitly distinguish these two types of uncertainty in the revision.

ScaleRawReSource
1K38.76 ± 0.3743.53 ± 0.47
2K45.75 ± 0.3354.75 ± 0.31
8K49.92 ± 0.2955.54 ± 0.24

Thank you so much again for your time and valuable suggestions! We hope these additional results and clarifications address your concerns, and we would appreciate it if you could reconsider your assessment based on our response.

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

Data selection approaches has been thoroughly studied for enhancing the data-quality of the preference dataset, leading to higher downstream performance. Consequently, the filtered-out examples can not contribute to the optimization process when there still could be learnable values in them. To address this limitation, the authors propose ReSource, which leverages the discarded instructions as a context for new data generation. The authors leverage ArmoRM as a reward source to train a source-grounded instruction generator which learns to generate instruction from the base discarded instruction, that will have high learning value. The empirical performance of using such recycled data shows superior performance across multiple benchmarks.

Critical strengths

Recycling the discarded supervision seems like a good approach for a data augmentation, because it is highly aligned with the task the original dataset intended to teach. A data synthesis pipeline stemming from such task-coupled context seems like a good approach. Although minor, their experimental results are well-organized and easy to follow.

Critical weaknesses

The positioning of recycling is unclear

  1. Apart from data selection, recent research has been focusing of data synthesis, where they generate new data to enhance the model's downstream performance. I think that the recycling work can position itself as a data synthesis pipeline with the provided context the instruction of the rejected set, while still has meaningful task-related instruction. If the authors agree on such positioning, I think a data synthesis baseline (e.g. a data synthesis pipeline with the task description only) is needed to show that the training cost of the generator is justified.
  2. If the authors do not want to position their work with the data synthesis line of work, I think that the use of external reward model ArmoRM to synthesize reward for RL training should be reconsidered. For example, SSPO[1] trains a proxy reward model from only the labeled subset of the preference data to enhance the semi-supervised preference optimization performance. Likewise for this manuscript, I think that if that if the authors would like to isolate the effectiveness of the recycling itself, the authors should use an internally trained reward model. For example, like the InstQual baseline, the authors have trained a reward model solely from the retained / discarded grouping. The most naive baseline of such internal reward model could be trained via rewarding the selected response and penalizing the rejected one.

The successfulness of recycled examples need thorough evaluation

  1. To thoroughly assess the quality and the efficacy of ReSource, I ask the authors to run the recycling pipeline on the retained instruction set. If ReSource is also helpful in for such retained instruction, we have to question the source of the high performance of the recycled data. This weakness will also linked with my first critical weakness, as the effectiveness in the retained instruction too could mean the source of successfulness could be in the external knowledge ArmoRM.
  2. If the results actually show efficacy in the retained set, but the authors think that effectiveness in the retained set is acceptable, except for the ArmoRM's external knowledge. I ask the authors to justify such behavior.

[1] Semi-Supervised Preference Optimization with Limited Feedback, ICLR'26

Questions for authors
  1. What is the reason for using the mean response quality as the selection metric? I think that the instructions selected based on such score could lead to favoring easier task formulations, since instructions that have high response quality generally means that they are easy to respond. Have the authors considered using the variance of the response quality?
3 comments
AuthorsOpenReview ↗

Response to Reviewer qDoo

Thank you for the constructive feedback! We clarify our positioning and provide additional evidence below.

I think that the recycling work can position itself as a data synthesis ... If the authors agree on such positioning, I think a data synthesis baseline is needed …

We agree that ReSource is a source-grounded data synthesis method. In essence, our main question is whether real-user queries discarded by data selection can provide more useful training signal through recycling, and our source-matched comparisons are designed to address this.

Following your suggestion, we additionally evaluated data synthesis baselines, including self-instruct-style generation using retained or discarded queries as seeds, and also a source-free condition. All methods below use the same amount of 2K data, the same response generation configuration and an identical preference construction procedure.

Method (seed pool)AlpacaEval2 LCArena-HardWildBench
Raw46.20 ± 0.5168.55 ± 1.2444.29 ± 0.95
Synthetic (discarded)50.53 ± 0.4670.67 ± 1.2146.61 ± 0.94
Synthetic (retained)52.72 ± 0.4774.11 ± 1.1048.39 ± 0.94
Synthetic (source-free)43.29 ± 0.4769.83 ± 1.0343.95 ± 0.98
ReSource (discarded)54.86 ± 0.4573.93 ± 1.1049.12 ± 0.95

As it can be seen above, ReSource outperforms the discarded-seed and source-free baselines across all three benchmarks. Even compared with synthesis from higher-quality retained seeds, it achieves comparable/higher scores.

If the authors do not want to position their work with data synthesis … use an internally trained reward model.

As mentioned above, we agree with the data-synthesis positioning. We would also like to clarify that RAW and the synthesis baselines use the same ArmoRM-based preference construction. Our comparisons evaluate the additional benefit of ReSource’s source-grounded task generation and search for recycling under this shared procedure.

I ask the authors to run the recycling pipeline on the retained instruction set. …

We appreciate your insightful suggestion. We ran the same recycling pipeline on 2K retained instructions as requested.

SourceRawRecycled
Retained56.27 ± 0.4856.19 ± 0.44
Discarded46.20 ± 0.5154.86 ± 0.45

You can see that recycling on the retained set does not yield better performance, unlike recycling on discarded one. This weakens the concern that the gains simply reflect applying external reward supervision, rather than recovering value from discarded examples.

What is the reason for using the mean response quality as the selection metric? … Have the authors considered using the variance of the response quality?

Our choice is based on our observation that retained queries elicit higher mean rewards than discarded queries (0.1821 vs. 0.1379). In fact, higher response variability is not necessarily preferable. Prior work, RIP [1] argues that large reward gaps can reflect prompt ambiguity and they actually filter out prompts with such gaps. We also noted that high-quality retained queries showed lower within-query reward variance than discarded ones. These motivate our use of mean response quality as a practical selection metric. Although we do not claim it is necessarily optimal, our empirical results support its effectiveness.

[1] R.I.P. : Better Models by Survival of the Fittest Prompts

Thank you so much again for your time and valuable suggestions! We hope these additional results and clarifications address your concerns, and we would appreciate it if you could reconsider your assessment based on our response.

I thank the authors for the detailed rebuttal on my concerns.

  • The data synthesis baseline with the matched external reward model ArmoRM is less performant compared to ReSource, meaning that the source-grounded task generation is the source of the performance increase
  • Recycling the retained set does not increase the performance, meaning that the use of external feedback is not the source of high performance.
  • I appreciate the authors' explanation on the choice of metric selection.

As all of my concerns are resolved, I am raising my score.

Thank you for raising your score. We are glad to hear that your concerns have been addressed. Your valuable insights and feedback have helped us improve our paper!

Reviewer_VcM4OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a data recycling method that extracts useful information from data with limited training value and uses it to synthesize new training data. The experiments show that training on the recycled and synthesized dataset improves model performance.

Critical strengths
  • The proposed method shows empirical effectiveness. Under source-matched comparisons, RESOURCE improves downstream performance over using the original instructions across AlpacaEval2, Arena-Hard, and WildBench.
  • Recycling discarded data has practical significance. The method gives data with limited value in its original form another use by extracting reusable content and constructing new tasks and preference supervision.
Critical weaknesses
  • The baseline set is limited. The paper compares against RAW, INSTQUAL, and several ablations, but it does not include simple alternatives such as rewriting rejected queries with strong LLMs, using standard synthetic-instruction pipelines with rejected queries as seeds or using a faithful task-preserving recycle baseline. This makes it hard to tell how much of the gain comes from the full RESOURCE pipeline.
  • The preference-pair construction relies heavily on the same ArmoRM best/worst selection procedure. Additional response-selection and reward-model baselines would help isolate where the gains come from.
  • The paper does not provide a sufficient comparison with retained data. The paper shows that recycled data can complement retained data, but it does not fully isolate the marginal value of recycled rejected examples against adding the same amount of retained high-quality data under a matched total budget.
  • The paper uses Llama-3.1-70B-Instruct as the evaluation model, but it does not report the accuracy or reliability of this evaluator in the paper. The paper should provide an experiment measuring the evaluator's agreement with human judgments.
  • The choice of Llama-3.1-8B-Instruct for constructing the synthetic data is unclear. Since the method relies heavily on synthetic instruction generation, a strong no-training generator may already perform well. The paper does not compare current setup with larger open models or commercial LLMs used directly for recycling. Therefore it is hard to tell whether SFT/GRPO training is essential
Questions for authors

In addition to the weaknesses noted above, please improve the paper’s layout and typesetting. Figure 2 contains substantial blank space, which reduces the amount of information presented. Adding a case study could make the figure more informative. The letter spacing in Figures 1 and 2 also appears unnatural. Please revise the typography and clarify whether generative AI tools were used in preparing these figures

4 comments
AuthorsOpenReview ↗

Response to Reviewer VcM4 (Part 1)

Thank you for the detailed review and suggestions for strengthening the empirical comparisons. We provide additional experiments and clarifications below to address your concerns.

The baseline set is limited. … rewriting rejected queries with strong LLMs … standard synthetic-instruction pipelines … a faithful task-preserving recycle baseline.

Thank you for your suggestion. We added the following baselines: faithful rewriting of discarded queries with stronger Qwen3-32B, which preserves the original task and improves clarity, general self-instruct style synthetic data generation using discarded queries as seeds, and also source-free synthesis. All methods below use 2K training examples, the same Llama-3.1-8B response generation configuration, and an identical preference construction procedure.

MethodAlpacaEval2 LCArena-HardWildBench
Raw46.20 ± 0.5168.55 ± 1.2444.29 ± 0.95
Qwen3-32B faithful rewriting46.78 ± 0.5269.20 ± 1.1545.48 ± 0.93
Synthetic (discarded-seed)50.53 ± 0.4670.67 ± 1.2146.61 ± 0.94
Synthetic (source-free)43.29 ± 0.4769.83 ± 1.0343.95 ± 0.98
ReSource54.86 ± 0.4573.93 ± 1.1049.12 ± 0.95

We observed that faithful rewriting shows only modest improvements over Raw even with a strong LLM, which suggests that simply rewriting may not be sufficient for recycling. It can also be seen that standard synthesis using discarded queries as seeds improves over Raw, but ReSource further improves performance. These highlight the usefulness of discarded source content and the effectiveness of explicit source-grounded recycling.

The preference-pair construction relies heavily on the same ArmoRM best/worst selection procedure. …

To clarify, Raw and ReSource use the same ArmoRM-based best/worst preference construction, so their gap cannot be attributed to using different preference-construction procedures. We nevertheless tested an alternative RM, internally trained using original preference pairs from the retained and discarded raw data.

RMRawReSource
ArmoRM46.20 ± 0.5154.86 ± 0.45
Internal RM47.13 ± 0.5054.78 ± 0.45

With the only change of the RM component, we found that ReSource similarly improves over Raw with either RM.

The paper shows that recycled data can complement retained data, but it does not fully isolate the marginal value of recycled rejected examples against adding the same amount of retained high-quality data under a matched total budget.

Our mixture experiments (Table 2) in the paper already include comparisons at matched total training-set sizes. Here are summarized results below.

RetainedRecycledTotalAlpacaEval2 LC
4K-4K55.56 ± 0.49
2K2K4K57.56 ± 0.47
4K4K8K59.01 ± 0.43

At the same total training-set size of 4k, replacing half of the retained examples with recycled examples improves from 55.56 to 57.56. Our claim is not that recycled data are necessarily better than retained ones. Importantly, the retained pool contains only 4k examples, as data selection filtered out many other examples. Recycling allows us to expand beyond this limited size. Adding more 4K recycled examples to 4K retained examples improves the score from 55.56 to 59.01. This demonstrates the practical value of recovering additional supervision from discarded data.

AuthorsOpenReview ↗

Response to Reviewer VcM4 (Part 2)

The paper uses Llama-3.1-70B-Instruct as the evaluation model, but it does not report the accuracy or reliability of this evaluator in the paper. The paper should provide an experiment measuring the evaluator's agreement with human judgments.

As stated in the Evaluation section, Llama-3.1-70B-Instruct shows strong agreement with human judgments according to [1]. In their evaluation, it achieves 66.2% agreement with human judgments, compared with 64.8% for GPT-4-Turbo and 67.0% inter-human agreement.

To further assess the reliability, we additionally evaluated the same 2K-trained models using GPT-4o judge (gpt-4o-2024-11-20). ReSource outperforms Raw across all three benchmarks under this judge as well. In addition, even with GPT-4.1 (gpt-4.1-2025-04-14), the same trend holds although the absolute scales differ across judges: AlpacaEval2 LC increases from 22.35 for Raw to 25.19 for ReSource.

MethodAlpacaEval2 LCArena-HardWildBench
Raw38.98 ± 0.1138.28 ± 1.2337.43 ± 0.92
ReSource41.62 ± 0.1243.77 ± 1.3041.75 ± 0.95

[1] HREF: Human Response-Guided Evaluation of Instruction Following in Language Models

The choice of Llama-3.1-8B-Instruct for constructing the synthetic data is unclear. Since the method relies heavily on synthetic instruction generation, a strong no-training generator may already perform well. … Therefore it is hard to tell whether SFT/GRPO training is essential.

We additionally evaluated stronger and more capable Qwen3-32B as an untrained generator within the same recycling pipeline. All settings use 2K training examples, the same instruction-search configuration, Llama-3.1-8B response generation, and preference construction.

RecyclerTrainingAlpacaEval2 LC
Qwen3-32BNone51.16 ± 0.47
Llama-3.1-8BNone48.85 ± 0.48
Llama-3.1-8BSFT50.72 ± 0.49
Llama-3.1-8BSFT + GRPO54.86 ± 0.45

As expected, the larger Qwen3-32B performs better than the untrained 8B model. We do not claim that recycling necessarily requires Llama-3.1-8B or that stronger generators cannot perform well directly. Our goal however is to improve a given generator to better leverage its potential for recycling. As shown in the paper, within the same 8B model, SFT and subsequent GRPO progressively improve performance. This shows that training actually helps and that a trained smaller recycler can achieve better performance than a larger untrained model.

please improve the paper’s layout and typesetting. … The letter spacing in Figures 1 and 2 also appears unnatural. Please revise the typography and clarify whether generative AI tools were used in preparing these figures

Thank you for the presentation suggestions. AI tools were only used to explore preliminary layout and structural ideas, but the figures themselves were manually created using PowerPoint by the author. We will review the typography and spacing and refine the layout to improve readability.

Thank you so much again for your time and valuable suggestions! We hope these additional results and clarifications address your concerns, and we would appreciate it if you could reconsider your assessment based on our response.

I thank the authors for the detailed clarification. As most of my concerns have been addressed, I will raise my score. I encourage the authors to incorporate the rebuttal content into the next version of the paper.

Thank you for raising your score! We have revised our paper to incorporate the rebuttal content, and we believe your constructive suggestions have really helped us improve the paper. We sincerely appreciate it!

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to Reviewers

We thank all the reviewers for taking their valuable time to provide insightful comments and suggestions to improve our paper.

We are pleased to note that the reviewers recognized the practical motivation, empirical effectiveness, and controlled experimental design of our study, as evidenced by the following positive comments:

  • “The proposed method shows empirical effectiveness. …” (VcM4)
  • “Recycling discarded data has practical significance. …” (VcM4)
  • “Recycling the discarded supervision seems like a good approach …” (qDoo)
  • “their experimental results are well-organized and easy to follow” (qDoo)
  • “The problem is practical. …” (W9YY)
  • “The source-matched comparisons are a strong part of the design. …” (W9YY)
  • “Table 1 shows consistent RESOURCE improvements across three benchmarks and all four tested data scales, making the main empirical result fairly convincing.” (W9YY)
  • “The mixture results and cross-model instruction transfer are useful, and Table 4 separates contributions from recycler training and instruction search.” (W9YY)

We believe the thoughtful reviews and the suggestions made by the reviewers have substantially improved the quality of the paper. In particular, following our additional experiments and clarifications, Reviewer qDoo confirmed that “all of my concerns are resolved” and raised their score from 2 to 3. Based on the reviewers' suggestions, we have updated the paper with the following major changes. Changes are highlighted in purple throughout the updated paper.

  • More baseline comparisons: We included more baselines to compare, such as self-instruct-style synthesis, source-free generation, faithful rewriting, and direct recycling with Qwen3-32B. (Section 4.1, Table 2; VcM4, qDoo, W9YY)
  • Other reward model: We tested an internally trained reward model for final preference-pair selection. (Appendix C.4, Table 9; VcM4)
  • Recycling retained data: We applied the same pipeline to retained queries to compare its effects on retained and discarded data. (Appendix C.3, Table 8; qDoo)
  • Additional evaluation judges: We added GPT-4o and GPT-4.1 judges for evaluations, and clarified the published human-agreement evidence for our local judge. (Section 3, Appendix C.1, Table 6; VcM4, W9YY)
  • Multi-seed training runs: We added five-run training results across data scales, distinguishing across-run standard errors and evaluation uncertainty. (Appendix C.2, Table 7; W9YY)
  • Construction budgets: We reported training/recycling costs and discussed the compute-performance trade-off. (Appendix C.5, Table 10, Conclusion; W9YY)
  • Clarified existing comparisons and design choices: We highlighted matched-size retained-data comparisons and explained the motivation for mean response quality as the selection metric. (Sections 4.2 and 2.4; VcM4, qDoo)
  • Improved presentation: Revised the typography and spacing of Figures 1 and 2. (VcM4)

We hope these additional results and revisions address the reviewers’ concerns and improve the clarity and completeness of the paper. Thank you again for your time, valuable suggestions, and continued consideration of our work!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.