acceptodds
Under review as a conference paper at ICLR 2027

Let Training Guide Selection: Online Synthetic Data Filtering via Real-Anchored Utility

Abstract

Synthetic data can scale training supervision when real-world data are limited, but noise and distribution mismatch can reduce its value. Existing synthetic data selection methods often emphasize fidelity or diversity rather than the learner's evolving needs. We propose FROST, an online framework that estimates synthetic-data utility through gradient feedback anchored in real training data. It calibrates batch utility against recent history to determine when filtering is needed and filters samples only in out-of-band batches to determine what to retain, without an external verifier or held-out validation set. Experiments on two public benchmarks for image classification and LLM fine-tuning for text-to-SQL show that FROST filters out around 20–30% of the synthetic data while improving real-task performance compared with training on the full synthetic data pool. We further apply FROST during training in a large-scale industrial ads re-ranking system, achieving significant performance gains over a highly optimized production baseline, demonstrating its effectiveness and generalizability.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
synthetic data, online data selection, data-centric learning

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_CPL1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes FROST, an online synthetic data filtering framework designed to improve model performance on real-world tasks by selecting useful synthetic training examples. The method estimates sample utility through gradient alignment with an exponential moving average of real-training gradients. It then employs a two-stage filtering strategy: batch-level utility calibration against recent training history determines whether filtering is necessary, while interquartile-range (IQR) thresholds select samples from batches with unusually low or high utility. The framework does not require an external verifier or a held-out validation set. Experiments on CIFAR-100 image classification, Text-to-SQL fine-tuning, and industrial ads re-ranking demonstrate that FROST can improve downstream performance while discarding a portion of the synthetic training data.

Critical strengths
  • The paper addresses an important challenge in synthetic data training, where noisy or distributionally mismatched samples may negatively affect downstream performance. By using real-training gradients to guide online data selection, FROST provides a practical approach to identifying useful synthetic supervision without requiring an external verifier or a separate validation set.

  • FROST combines gradient-based utility estimation with a conditional two-stage filtering strategy. The batch-level calibration determines when filtering is necessary, while sample-level IQR filtering selectively removes potentially harmful examples. The overall framework is straightforward to understand and implement, with a first-order approximation providing theoretical motivation for the utility estimation.

  • The authors evaluate FROST across image classification, LLM fine-tuning, and large-scale industrial ads re-ranking, demonstrating its applicability to different architectures and tasks. The method consistently improves downstream performance over training with all synthetic data while filtering a substantial portion of synthetic samples. The industrial results further highlight its potential practical value

Critical weaknesses
  • Although Appendix B.3 provides a theoretical analysis of the utility computation overhead and reports training throughput reductions of approximately 6.8% and 5.7% for image classification and industrial ads re-ranking, respectively, the paper does not report end-to-end wall-clock training time. More importantly, no runtime measurements are provided for the LLM fine-tuning experiment, where gradient-based utility estimation may introduce non-negligible overhead. Since FROST requires scoring synthetic candidates before filtering, reducing the number of retained samples does not necessarily translate into proportional computational savings. Without direct runtime comparisons against baselines such as GREATS, the practical computational efficiency of FROST remains unclear.

  • Although the authors conduct ablation studies to evaluate the two-stage filtering strategy, the compared variants retain substantially different amounts of synthetic data. In Figure 3, Sample IQR Only removes 32.2% of synthetic samples, whereas full FROST removes only 20.4%. A similar discrepancy appears in Figure 5, where the corresponding removal ratios are 54% and 36%. Since full FROST consistently retains more training data, the observed performance improvements cannot be attributed solely to the batch-level calibration mechanism. The current ablations therefore do not sufficiently isolate the contribution of this key component.

  • A central motivation for the bounded utility interval is the observation that extremely high-utility batches may lead to less stable training. However, the supporting experiment in Figure 2 is limited to 30 continuation steps from a single checkpoint after 20 training epochs. Although the moderate-utility group achieves higher accuracy at the final observed step, the high-utility group reaches a higher early peak. This limited experiment does not establish whether moderate-utility batches consistently outperform high-utility batches across different training stages or longer training horizons. Consequently, the empirical justification for routing extremely high-utility batches to additional filtering remains insufficient.

Questions for authors
  • Could the authors provide end-to-end wall-clock training time and training throughput comparisons between FROST, training with all synthetic data, and GREATS, particularly for the LLM fine-tuning experiment? This would help clarify whether the proposed filtering mechanism introduces additional computational overhead or offers practical training-time benefits despite requiring utility estimation for all candidate samples.

  • Could the authors adjust the filtering thresholds of Sample IQR Only to match the synthetic-data retention ratio of full FROST and report the corresponding performance? Such a controlled comparison would help determine whether batch-level calibration provides additional benefits beyond simply retaining more synthetic training examples.

  • Could the authors extend the analysis in Figure 2 to multiple training checkpoints and longer continuation horizons? Additionally, could the authors evaluate the effect of removing or relaxing the upper utility threshold while keeping the synthetic-data retention ratio comparable? These experiments would help establish whether filtering extremely high-utility batches is necessary for improving training stability and downstream performance.

  • Could the authors provide an experiment comparing the sample utility scores computed using the prediction head with those computed using full-model gradients? Reporting their ranking correlation would help assess the reliability of the partial-gradient approximation used in FROST.

6 comments
AuthorsOpenReview ↗

Response to Reviewer CPL1 (Part 1/4)

We sincerely thank you for the positive feedback and for recognizing the practical value of FROST's real-data-anchored utility estimation, conditional two-stage filtering, and evaluation across three domains. We appreciate your suggestions for strengthening the experimental evidence. We take this opportunity to address your comments in detail below.


1. End-to-end training time and computational overhead

Could the authors provide end-to-end wall-clock training time and training throughput comparisons between FROST, training with all synthetic data, and GREATS, particularly for the LLM fine-tuning experiment? This would help clarify whether the proposed filtering mechanism introduces additional computational overhead or offers practical training-time benefits despite requiring utility estimation for all candidate samples.

Here we report end-to-end training times for 0.5B and 7B text-to-SQL models. Regular training uses all real and synthetic data without filtering. Within each model setting, all methods use the same dataset and pre-filter batch size, with matched synthetic-data drop rates for FROST and GREATS. Parentheses show the time increase relative to regular training for the same model and GPU.

ModelGPURegular training (s)FROST (s; overhead)GREATS (s; overhead)
0.5BA100 80GB26102670 (+2.30%)2784 (+6.67%)
7BB20057185838 (+2.10%)6012 (+5.14%)

FROST adds 2.10–2.30% to the end-to-end training time, compared with 5.14–6.67% for GREATS. FROST is faster than GREATS in both configurations.

Selection overhead. Compared with FROST, GREATS [1] additionally computes pairwise candidate-gradient interactions, forming an m×mm\times m matrix, where mm is the number of synthetic candidates in a batch before filtering. For pp-dimensional scoring gradients, direct dense construction costs O(m2p)O(m^2p) once gradients are available. FROST aligns candidates with a single real-gradient reference in O(mp)O(mp), then calibrates the batch mean against recent history. The remaining operations use scalar scores: averaging and threshold comparisons take O(m)O(m), while sorting-based IQR computation takes O(mlog⁡m)O(m\log m) only for out-of-band batches.

Why both methods take longer than regular training. In our benchmark implementations of FROST and GREATS, filtering zeros the gradient contributions of excluded samples without skipping their dense forward/backward computation. Selection therefore adds computation without converting the drop rate into proportional training-time savings. Selected-sample execution could reduce this cost, but these optimizations are outside our current scope. Even with this dense implementation, FROST adds only around 2% to end-to-end training time in these LLM settings.


Reference

[1] Jiachen T. Wang, Tong Wu, Dawn Song, Prateek Mittal, and Ruoxi Jia. GREATS: Online Selection of High-Quality Data for LLM Training in Every Iteration. NeurIPS, 2024.

AuthorsOpenReview ↗

Response to Reviewer CPL1 (Part 2/4)

2. Matched-drop-ratio ablations

Could the authors adjust the filtering thresholds of Sample IQR Only to match the synthetic-data retention ratio of full FROST and report the corresponding performance? Such a controlled comparison would help determine whether batch-level calibration provides additional benefits beyond simply retaining more synthetic training examples.

Thank you for pointing this out. We adjusted the filtering thresholds to approximately match Full FROST's synthetic-data drop rate. Results are mean ± standard deviation over 3 seeds for CIFAR-100 and 5 seeds for Qwen2.5-Coder-0.5B text-to-SQL.

CIFAR-100.

GeneratorMethodSynthetic-data drop rate (%)Accuracy (%)
FLUXBatch only25.84 ± 0.8080.17 ± 0.17
FLUXSample only25.02 ± 0.6180.55 ± 0.08
FLUXFull FROST26.40 ± 0.5780.86 ± 0.13
SANABatch only26.02 ± 0.0979.26 ± 0.21
SANASample only25.58 ± 1.6779.82 ± 0.19
SANAFull FROST26.96 ± 0.8380.29 ± 0.29
SD1.4Batch only25.07 ± 0.8379.93 ± 0.12
SD1.4Sample only25.32 ± 0.0480.64 ± 0.21
SD1.4Full FROST25.32 ± 0.6480.80 ± 0.18

Text-to-SQL.

MethodSynthetic-data drop rate (%)EX (%)
Batch only21.71 ± 1.0364.42 ± 0.75
Sample only20.54 ± 0.0464.40 ± 0.47
Full FROST20.39 ± 0.3165.92 ± 0.59

Full FROST achieves higher mean performance than both single-stage variants in all four settings. In image classification, it achieves higher accuracy than Sample only and Batch only across all 3 generators at similar synthetic-data retention rates. In text-to-SQL, Full FROST reaches 65.92% EX versus 64.40% for Sample only at nearly identical drop ratios. These results support the benefit of the two-stage filtering design.

Industrial ads re-ranking. We also adjusted the drop rates of the single-stage variants for our commercial ads re-ranking model. Given the substantial training resources required, we report the available results below. Relative NE change is measured against real-only training; lower is better.

MethodSynthetic-data drop rate (%)Relative NE change (%) ↓
All Synthetic0+0.211
Batch only35+0.097
Sample only22+0.128
Sample only45+0.070
Sample only51+0.057
Full FROST36-0.096

At similar drop rates (35% versus 36%), Batch only yields a +0.097% relative NE change, while Full FROST achieves -0.096%, improving over real-only training. Sample only also underperforms Full FROST at all tested drop rates (22%, 45%, and 51%).

AuthorsOpenReview ↗

Response to Reviewer CPL1 (Part 3/4)

3. Multiple checkpoints and longer continuation horizons

Could the authors extend the analysis in Figure 2 to multiple training checkpoints and longer continuation horizons? Additionally, could the authors evaluate the effect of removing or relaxing the upper utility threshold while keeping the synthetic-data retention ratio comparable? These experiments would help establish whether filtering extremely high-utility batches is necessary for improving training stability and downstream performance.

Thank you for raising this question about the design of the upper utility bound. Prior work shows that excessively large stochastic gradients can push iterates away from the optimum and cause oscillations [1]. Merad and Gaïffas [2] use rolling gradient-norm quantiles to clip large stochastic gradients and improve stability with corrupted samples. Inspired by this principle, FROST uses an upper utility bound to route unusually high-utility batches to sample-level filtering. To further confirm our design choice, we extended the mid-band analysis to multiple checkpoints and 100 continuation steps, and tested removing the upper bound at comparable drop rates.

Multiple-checkpoint continuation analysis. We use image-model checkpoints at epochs 20, 90, and 160, and Qwen2.5-Coder-0.5B checkpoints at steps 10, 100, 300, and 500. From each checkpoint, we continue training for 100 steps using each of the Low, Mid, and High synthetic-batch utility bands, with 5 repeats per band. To reduce sensitivity to endpoint fluctuations, we first average accuracy or EX over the first 30, 50, or 100 steps within each repeat, then report the mean ± standard deviation of these window averages across the 5 repeats.

Image classification: accuracy (%).

Checkpoint (epoch)Averaging windowLow-band trainingMid-band trainingHigh-band training
20First 30 steps64.08 ± 0.9865.65 ± 0.6464.93 ± 0.97
20First 50 steps64.45 ± 0.7864.95 ± 0.5364.87 ± 0.57
20First 100 steps64.64 ± 0.4664.69 ± 0.5264.87 ± 0.50
90First 30 steps58.75 ± 0.6559.80 ± 0.5859.62 ± 0.35
90First 50 steps58.95 ± 1.0759.54 ± 0.2559.11 ± 0.53
90First 100 steps58.79 ± 0.5959.24 ± 0.4858.60 ± 0.56
160First 30 steps73.44 ± 0.3074.09 ± 0.4173.76 ± 0.52
160First 50 steps73.70 ± 0.1174.24 ± 0.1674.03 ± 0.39
160First 100 steps73.83 ± 0.2874.09 ± 0.1173.95 ± 0.19

Text-to-SQL: EX (%).

Checkpoint (step)Averaging windowLow-band trainingMid-band trainingHigh-band training
10First 30 steps48.83 ± 0.6852.20 ± 0.7550.77 ± 0.64
10First 50 steps48.75 ± 0.5451.05 ± 0.6150.47 ± 0.85
10First 100 steps49.59 ± 0.3151.22 ± 0.4350.32 ± 0.57
100First 30 steps51.84 ± 0.4252.33 ± 1.3152.07 ± 0.31
100First 50 steps52.13 ± 0.3052.51 ± 0.3352.94 ± 0.41
100First 100 steps53.25 ± 0.3653.90 ± 0.5454.26 ± 0.70
300First 30 steps60.86 ± 0.3461.18 ± 0.1760.69 ± 0.45
300First 50 steps61.26 ± 0.3261.27 ± 0.0661.49 ± 0.56
300First 100 steps62.65 ± 0.3262.56 ± 0.2362.41 ± 0.27
500First 30 steps64.66 ± 0.1765.16 ± 0.0863.50 ± 0.17
500First 50 steps64.78 ± 0.1465.14 ± 0.0663.51 ± 0.01
500First 100 steps64.94 ± 0.1665.08 ± 0.0963.66 ± 0.07

Mid has the highest mean in 8 of 9 image comparisons and exceeds High in 9 of 12 text-to-SQL comparisons. Overall, Mid-utility batches generally yield better average performance across the tested checkpoints and horizons, showing that higher current utility does not consistently translate into better subsequent performance.


Reference

[1] Eduard Gorbunov, Marina Danilova, and Alexander Gasnikov. Stochastic Optimization with Heavy-Tailed Noise via Accelerated Gradient Clipping. NeurIPS, 2020.

[2] Ibrahim Merad and Stéphane Gaïffas. Robust Stochastic Optimization via Gradient Quantile Clipping. Transactions on Machine Learning Research, 2024.

Reviewer_Ba9bOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper targets key drawbacks of existing synthetic‑data selection algorithms: reliance on pre‑trained feature extractors, static selection scores, and failure to capture the dynamically‑changing data needs of models throughout training. To tackle these issues, the authors propose FROST (Filtering with Real data for Online Synthetic‑data Training), an online filtering framework that selects synthetic samples beneficial for real‑world task performance. The effectiveness of FROST is verified across multiple experimental settings.

Critical strengths

The key contribution is FROST, a model‑ and task‑agnostic online synthetic‑data filtering framework. It estimates sample utility using real‑data‑anchored gradient alignment and leverages conditional two‑level filtering at both batch and sample granularity, without offline pre‑processing of synthetic data or an external validation set.

Critical weaknesses

(1)In the LLM text‑to‑SQL experiment, only the tiny 0.5B‑parameter Qwen2.5‑Coder model is tested. Experiments on larger‑size LLMs are absent, leaving open the question whether the observed gains transfer to bigger models. (2)The industrial‑scale ad re‑ranking experiment does not report the exact underlying recommendation model architecture. Note that this may stem from confidentiality constraints of production systems. (3)Generalization across more diverse model architectures within each task is not sufficiently validated. While the method is evaluated across three different application domains, each task is demonstrated with only one primary model.

Questions for authors

The third experiment is conducted on real‑world production data for industrial ad re‑ranking. Many industrial applications suffer from severe data scarcity and thus heavily rely on synthetic data for model training. By contrast, the ad re‑ranking scenario comes with relatively abundant real‑world samples. It would be valuable to see results over additional distinct industrial use‑cases, especially settings where real data is extremely limited and synthetic data constitutes the major source of supervision. Could the authors comment on why such evaluations are missing?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Ba9b (Part 1/2)

We really appreciate your detailed and constructive feedback, as well as your recognition of FROST's practical design. We have conducted additional experiments to address your concerns about model scale, generalizability, and industrial applicability.

We believe our response has addressed most of your concerns, and we respectfully ask whether you would be willing to reconsider your rating in light of our efforts and dedication. Below, we provide detailed responses to all your comments:


1. Additional industrial model, a larger LLM, and model generalization

In the LLM text‑to‑SQL experiment, only the tiny 0.5B‑parameter Qwen2.5‑Coder model is tested. Experiments on larger‑size LLMs are absent, leaving open the question whether the observed gains transfer to bigger models.

Generalization across more diverse model architectures within each task is not sufficiently validated. While the method is evaluated across three different application domains, each task is demonstrated with only one primary model.

Thank you for these suggestions. Our original evaluation spans three distinct architectures: a convolutional ResNet-18 for image classification, a Transformer-based LLM for text-to-SQL, and a deep recommendation model for ads re-ranking. These results demonstrate FROST's generalizability across architectures and domains. We further extend the evaluation to another industrial model, an encoder–decoder LLM, a vision Transformer, and a larger LLM below.

Additional industrial ads model.

In the original paper, we applied FROST to a commercial ads re-ranking model for conversion rate (CVR) prediction. We have now also applied FROST to a commercial ads model for click-through rate (CTR) prediction. It uses a substantially different model architecture, including its feature embeddings and user-history sequence modeling. We report relative NE change against the CTR model trained on real data only; lower is better.

MethodSynthetic-data drop rate (%)Relative NE change (%) ↓
All Synthetic0+0.124
Full FROST25-0.031

FROST turns the +0.124% relative NE regression from All Synthetic into a -0.031% relative NE change while filtering 25% of the synthetic data. These results extend the industrial evidence from conversion prediction to click prediction.

This 0.031% relative NE improvement is significant in practice, exceeding the approximately 0.02% gain considered significant in prior studies of well-optimized production recommenders [1, 2].


Additional LLM architecture: Google FLAN-T5-Large.

We also evaluate Google's FLAN-T5-Large on text-to-SQL, extending the evaluation from decoder-only Qwen models to an encoder–decoder architecture. Results are mean ± standard deviation over 5 seeds. FROST improves mean EX by 0.46 percentage points over All Synthetic, 1.06 over Random, and 0.98 over GREATS, while filtering 20.52% of the synthetic data.

MethodSynthetic-data drop rate (%)EX (%)
Real only10049.30 ± 1.19
All Synthetic058.44 ± 0.90
Random20.5257.84 ± 0.61
GREATS20.82 ± 0.6357.92 ± 0.78
Full FROST20.52 ± 0.3158.90 ± 0.38

Additional image classification architecture: TinyViT-5M.

We also evaluate TinyViT-5M, a vision Transformer, on CIFAR-100 with FLUX-generated images. This extends the evaluation beyond ResNet-18. FROST improves accuracy by 1.27 percentage points over All Synthetic, while filtering 23.71% of the synthetic data. Results are mean ± standard deviation over 3 seeds.

MethodSynthetic-data drop rate (%)Accuracy (%)
Real only—62.71 ± 0.52
All Synthetic065.28 ± 0.30
Full FROST23.71 ± 0.0966.55 ± 0.27

Larger LLM: Qwen2.5-Coder-7B.

We extend the text-to-SQL evaluation from the original 0.5B model to Qwen2.5-Coder-7B, which already achieves 79.10% EX without fine-tuning. Even with this strong starting performance, FROST is the only evaluated selection method that improves mean EX over All Synthetic, while filtering 19.89% of the synthetic data. Results for All Synthetic and the selection methods are reported as mean ± standard deviation over 5 seeds.

MethodSynthetic-data drop rate (%)EX (%)
Zero-shot—79.10
Real only10080.63 ± 0.34
All Synthetic081.64 ± 0.63
Random20.01 ± 0.0381.58 ± 0.73
GREATS19.53 ± 0.0081.56 ± 0.70
Full FROST19.89 ± 0.5181.92 ± 0.88

References

[1] Yan Li, Dhruv Choudhary, Xiaohan Wei, Baichuan Yuan, Bhargav Bhushanam, Tuo Zhao, and Guanghui Lan. Frequency-aware SGD for Efficient Embedding Learning with Provable Benefits. ICLR, 2022.

[2] Fan Lai et al. AdaEmbed: Adaptive Embedding for Large-Scale Recommendation Models. OSDI, 2023.

AuthorsOpenReview ↗

Response to Reviewer Ba9b (Part 2/2)

2. Industrial ads re-ranking model architecture

The industrial‑scale ad re‑ranking experiment does not report the exact underlying recommendation model architecture. Note that this may stem from confidentiality constraints of production systems.

Thank you for your understanding. As you noted, confidentiality constraints prevent us from disclosing the full architecture of this proprietary commercial production model. As described in the paper, it is a deep learning recommendation model for conversion prediction, and FROST computes utility using gradients from the final shared-backbone layer.


3. Additional industrial applications and sparse supervision

The third experiment is conducted on real‑world production data for industrial ad re‑ranking. Many industrial applications suffer from severe data scarcity and thus heavily rely on synthetic data for model training. By contrast, the ad re‑ranking scenario comes with relatively abundant real‑world samples. It would be valuable to see results over additional distinct industrial use‑cases, especially settings where real data is extremely limited and synthetic data constitutes the major source of supervision. Could the authors comment on why such evaluations are missing?

Thank you for this valuable suggestion. We would like to clarify that our original industrial CVR experiment already represents a setting with sparse real-world task supervision. Although real-world advertising systems generate large volumes of interaction logs, conversions that can be reliably attributed to ad impressions remain sparse. For example, the public recommendation dataset released with ESMM contains approximately 84 million impressions but only 18,000 conversions (0.02%) [1]. Similarly, in our scenario, real conversion supervision remains limited even with large-scale logs. In our setting, synthetic examples help us expand training supervision, and FROST selects useful synthetic examples to improve model performance.

Following your suggestion, we have also evaluated an additional distinct industrial use case: commercial click-through rate (CTR) prediction, as reported in Part 1. FROST improves relative NE by 0.031% over real-only training while filtering 25% of the synthetic data, extending our industrial evaluation beyond conversion prediction.


Thank you again for your constructive feedback. We hope these additional results and clarifications address your concerns about model generalization and industrial applicability. If you find that your concerns have been resolved, we would be grateful if you could reconsider your rating.

Reference

[1] Xiao Ma, Liqin Zhao, Guan Huang, Zhi Wang, Zelin Hu, Xiaoqiang Zhu, and Kun Gai. Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate. SIGIR, 2018, Table 1.

Dear Reviewer Ba9b,

We hope this message finds you well. We have provided additional results on Qwen2.5-Coder-7B, FLAN-T5-Large, TinyViT-5M, and another commercial ads model for CTR prediction. These experiments extend our evaluation across model sizes and architectures in all 3 domains, showing that FROST can consistently and effectively improve downstream task performance while using fewer synthetic training samples.

As the discussion period is nearing its end, we would like to ensure that we have addressed all your concerns satisfactorily. If there are any additional points or feedback you would like us to consider, please let us know. We greatly value your insights and would be happy to address any remaining questions.

If your concerns have been resolved, we would be grateful if you could reconsider your rating in light of the additional evidence.

Thank you again for your time and consideration.

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

This paper proposes FROST, an online framework for selecting synthetic data. It uses gradient alignment between real data and synthetic data to estimate selection utility. It contains two stages: first identify whether current batch utility falls within the history calibrated utility band, and then conduct sample-level filtering based on IQR if the batch is out-of-band. The framework is evaluated on image classification, LLM text-to-SQL, and industrial ads re-ranking. Experiments show FROST's improvements on different downstream tasks.

Critical strengths

Efficient and practical framework. The framework requires gradient calculation only on a selected model block when estimating utilities, and does not require external verifiers or selection models.

Broad coverage of evaluation. The method is evaluated across tasks in three different domains, including image classification, LLM downstream task, and industrial ads re-ranking. FROST outperforms multiple baselines on these tasks.

Critical weaknesses

Uncontrolled drop ratio in ablation studies. In Figure 3 and Table 5, the drop ratios vary substantially, from 20.4% to 80.0%. Across three ablation variants, lower drop ratios correlate with better performance. Therefore, this experiment cannot distinguish whether the improvement mainly comes from more training data or from higher selection quality.

Sensitivity to synthetic dataset quality. Both batch-level routing and sample-level filtering focus on relative utilities (compared with history or within the batch), but do not explicitly ensure that the retained samples have positive utility. For example, a low quality data batch can be retained if buffer scores are comparably low, even if its optimization contribution might be negative. Although Figure 2 shows the effectiveness of moderate selection on one dataset, it does not show the selection stability given history data batches with different quality.

Questions for authors

Drop-ratio-matched ablations. Could the authors provide comparisons among different variants under the same drop ratio, to isolate the effects of more training data and selection strategies?

Moderate selection strategy. Appendix A.1 suggests that high utility might come from large gradient norms, which can also increase the higher-order approximation error. So could using normalization in utility computation better align utility with actual training benefits?

Batch routing and buffer update. In Line 225, does “During filtering, the current U_t enters the buffer only after its routing decision” mean that out-of-band batches are not added to buffers, or all batches are added? If out-of-band batches are not added, will an initialization of low utilities in buffer continue to retain low utility batches?

5 comments
AuthorsOpenReview ↗

Response to Reviewer JvVM (Part 1/3)

We sincerely thank you for the detailed and constructive feedback. We appreciate your recognition of FROST's practical design, its use of selected-block gradients without external verifiers, and its broad evaluation across three domains. Your suggestions have helped us strengthen the experimental evaluation and clarify our design choices.

We have carefully considered each of your comments and provide additional experimental results and detailed clarifications below, organized around the following four points:

  1. Matched-drop-ratio ablations: Controlled comparisons of Full FROST, Batch only, and Sample only.
  2. Benefits of unnormalized utility: Comparison with normalization-based selection.
  3. Clarification of buffer updates.
  4. Clarification of sensitivity to synthetic-data quality.

1. Matched-drop-ratio ablations

Drop-ratio-matched ablations. Could the authors provide comparisons among different variants under the same drop ratio, to isolate the effects of more training data and selection strategies?

We adjusted the filtering thresholds to approximately match Full FROST's synthetic-data drop rate. Results are mean ± standard deviation over 3 seeds for CIFAR-100 and 5 seeds for Qwen2.5-Coder-0.5B text-to-SQL.

CIFAR-100.

GeneratorMethodSynthetic-data drop rate (%)Accuracy (%)
FLUXBatch only25.84 ± 0.8080.17 ± 0.17
FLUXSample only25.02 ± 0.6180.55 ± 0.08
FLUXFull FROST26.40 ± 0.5780.86 ± 0.13
SANABatch only26.02 ± 0.0979.26 ± 0.21
SANASample only25.58 ± 1.6779.82 ± 0.19
SANAFull FROST26.96 ± 0.8380.29 ± 0.29
SD1.4Batch only25.07 ± 0.8379.93 ± 0.12
SD1.4Sample only25.32 ± 0.0480.64 ± 0.21
SD1.4Full FROST25.32 ± 0.6480.80 ± 0.18

Text-to-SQL.

MethodSynthetic-data drop rate (%)EX (%)
Batch only21.71 ± 1.0364.42 ± 0.75
Sample only20.54 ± 0.0464.40 ± 0.47
Full FROST20.39 ± 0.3165.92 ± 0.59

Full FROST achieves higher mean performance than both single-stage variants in all four settings. In image classification, it achieves higher accuracy than Sample only and Batch only across all 3 generators at similar synthetic-data retention rates. In text-to-SQL, Full FROST reaches 65.92% EX versus 64.40% for Sample only at nearly identical drop ratios. These results support the benefit of the two-stage filtering design.

Industrial ads re-ranking. We also adjusted the drop rates of the single-stage variants for our commercial ads re-ranking model. Given the substantial training resources required, we report the available results below. Relative NE change is measured against real-only training; lower is better.

MethodSynthetic-data drop rate (%)Relative NE change (%) ↓
All Synthetic0+0.211
Batch only35+0.097
Sample only22+0.128
Sample only45+0.070
Sample only51+0.057
Full FROST36-0.096

At similar drop rates (35% versus 36%), Batch only yields a +0.097% relative NE change, while Full FROST achieves -0.096%, improving over real-only training. Sample only also underperforms Full FROST at all tested drop rates (22%, 45%, and 51%).

AuthorsOpenReview ↗

Response to Reviewer JvVM (Part 2/3)

2. Benefits of unnormalized utility

Moderate selection strategy. Appendix A.1 suggests that high utility might come from large gradient norms, which can also increase the higher-order approximation error. So could using normalization in utility computation better align utility with actual training benefits?

Normalizing our inner-product utility by the product of the two gradient norms gives a cosine similarity score:

ui,tcos=R^t⊤gi,t∥R^t∥ ∥gi,t∥=cos⁡(R^t,gi,t).u_{i,t}^{\mathrm{cos}}=\frac{\hat R_t^\top g_{i,t}}{\|\hat R_t\|\,\|g_{i,t}\|}=\cos(\hat R_t,g_{i,t}).

This normalized score measures directional alignment while removing magnitude information. Our original utility preserves both directional alignment and gradient magnitude:

ui,t=R^t⊤gi,t=∥R^t∥ ∥gi,t∥cos⁡(R^t,gi,t).u_{i,t}=\hat R_t^\top g_{i,t}=\|\hat R_t\|\,\|g_{i,t}\|\cos(\hat R_t,g_{i,t}).

The inner product estimates first-order real-loss reduction under a hypothetical SGD update, where gradient magnitude also affects the predicted benefit. FROST retains this information and routes extreme batch scores to sample-level filtering.

PCGrad [1] also finds that both gradient direction and magnitude are important for good multi-task learning performance.

Moreover, we directly tested normalization with two alternatives: Cosine greedy, which selects the highest-cosine samples in every batch, and Cosine two-stage, which uses cosine scores within FROST's batch-routing and sample-filtering structure. Results are mean ± standard deviation over 3 seeds on CIFAR-100, with similar synthetic-data drop rates.

GeneratorMethodSynthetic-data drop rate (%)Accuracy (%)
FLUXCosine greedy26.40 ± 0.0080.15 ± 0.21
FLUXCosine two-stage26.64 ± 0.0680.42 ± 0.13
FLUXFROST (inner product)26.40 ± 0.5780.86 ± 0.13
SANACosine greedy26.96 ± 0.0079.28 ± 0.21
SANACosine two-stage26.93 ± 0.0779.62 ± 0.37
SANAFROST (inner product)26.96 ± 0.8380.29 ± 0.29
SD1.4Cosine greedy25.35 ± 0.0079.45 ± 0.53
SD1.4Cosine two-stage25.22 ± 0.0779.77 ± 0.29
SD1.4FROST (inner product)25.32 ± 0.6480.80 ± 0.18

As shown in the table above, FROST achieves higher mean accuracy than both cosine variants across all 3 generators. The comparison between the two-stage variants supports retaining gradient magnitude alongside directional alignment in the utility score.

Reference

[1] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient Surgery for Multi-Task Learning. NeurIPS, 2020.

AuthorsOpenReview ↗

Response to Reviewer JvVM (Part 3/3)

3. Clarification of buffer updates

Batch routing and buffer update. In Line 225, does “During filtering, the current U_t enters the buffer only after its routing decision” mean that out-of-band batches are not added to buffers, or all batches are added? If out-of-band batches are not added, will an initialization of low utilities in buffer continue to retain low utility batches?

Thank you for raising this question. We would like to clarify that all candidate batches enter the buffer, including out-of-band batches. We first compute the raw, pre-filter batch utility UtU_t, make the routing/filtering decision using the preceding buffer, and then append this original UtU_t. We do not replace it with the retained subset's utility. Thus, “only after” specifies when the score enters the buffer, not which batches are admitted.

The buffer therefore captures the utility distribution of all recent candidate batches, rather than only accepted batches. We have revised Section 3.2 to make this update rule explicit.


4. Clarification of sensitivity to synthetic-data quality

Sensitivity to synthetic dataset quality. Both batch-level routing and sample-level filtering focus on relative utilities (compared with history or within the batch), but do not explicitly ensure that the retained samples have positive utility. For example, a low quality data batch can be retained if buffer scores are comparably low, even if its optimization contribution might be negative. Although Figure 2 shows the effectiveness of moderate selection on one dataset, it does not show the selection stability given history data batches with different quality.

Thank you for raising this important question. We would like to clarify that including every candidate batch's pre-filter utility, including negative scores, in the buffer is by design. If we kept only positive scores, the buffer would be biased toward high-utility batches.

First, our synthetic batches are formed by uniformly sampling examples from the candidate pool. As discussed in Section 3.2, the central limit theorem (CLT) provides a normal approximation for the mean utility of sufficiently large batches. We use recent batch utilities as a calibration baseline and compute the current batch's z-score, which measures its standardized deviation from the recent mean. FROST retains batches within the moderate-utility band and applies sample-level IQR filtering to out-of-band batches to remove utility outliers. This rule is relative and does not require every retained sample to have positive utility.

Second, uniform sampling mixes examples without grouping them by quality, reducing systematic quality clustering across consecutive batches. Our industrial ads results further demonstrate effectiveness in a noisy setting where using the full synthetic pool hurts performance: All Synthetic yields a +0.211% relative NE change, whereas FROST achieves -0.096%, both relative to real-only training. These results show that FROST can recover useful supervision even when unfiltered synthetic augmentation is detrimental.

We appreciate your suggestion and have clarified the sampling procedure and the rationale for historical calibration in the revised manuscript.


Thank you again for your constructive feedback. We hope these results and clarifications address your concerns. If you find that your concerns have been resolved, we would be grateful if you could consider improving your score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their thoughtful and constructive feedback. We are encouraged by your recognition of FROST's practical design, real-data-anchored utility estimation, and evaluation across multiple domains. We have carefully considered the comments and conducted additional experiments to address the main concerns. Below, we summarize our core contributions and the new evidence and clarifications presented in this rebuttal.

For clarity and simplicity, we refer to Reviewers JvVM, Ba9b, and CPL1 as R1, R2, and R3, respectively.


Core contributions and recognized strengths

  1. Efficient and practical online filtering. FROST uses a real-data-anchored gradient reference to estimate synthetic-data utility, without requiring an external verifier. Its lightweight scoring design and straightforward implementation support practical online selection.
  2. Real-data-anchored, conditional filtering. FROST provides a model- and task-agnostic framework that combines real-training gradient feedback with conditional batch-to-sample filtering, without offline preprocessing or a held-out validation set.
  3. Cross-domain effectiveness and industrial impact. Experiments across image classification, LLM fine-tuning, and industrial ads re-ranking demonstrate improved downstream performance with fewer synthetic training samples, highlighting the framework's practical value.

Additional experimental evidence

  1. Controlled ablations (R1, R3). We approximately match the synthetic-data drop rates of Full FROST, Batch only, and Sample only on CIFAR-100 and text-to-SQL. Full FROST achieves higher mean performance across all 4 public settings. We also adjust the single-stage variants' drop rates in industrial ads re-ranking, where Full FROST outperforms both variants at the tested drop rates.
  2. Comparison with normalized utility (R1). We compare inner-product utility with cosine-greedy and cosine-two-stage selection. FROST achieves higher mean accuracy than both alternatives across all 3 image generators at similar drop rates.
  3. Larger LLM (R2). We extend the text-to-SQL evaluation to Qwen2.5-Coder-7B. FROST is the only selection method among FROST, Random, and GREATS to improve mean EX over All Synthetic, while filtering 19.89% of the synthetic data.
  4. Additional industrial application (R2). We extend the original conversion-prediction study to a commercial click-through-rate model. FROST turns the +0.124% NE regression with All Synthetic into a -0.031% NE change while filtering 25% of the synthetic data.
  5. Additional model architectures (R2). We apply FROST to more model architectures across 3 domains, including TinyViT-5M, FLAN-T5-Large, and an additional industrial CTR model. In all 3 domains, FROST improves downstream performance while using fewer synthetic training samples.
  6. End-to-end training time (R3). FROST adds only 2.10–2.30% to training time across 0.5B and 7B LLMs, compared with 5.14–6.67% for GREATS.
  7. Effectiveness of the upper utility bound (R3). We extend continuation experiments to 3 image-model checkpoints and 4 text-to-SQL checkpoints, up to 100 steps, with 5 repeats. We also remove the upper routing bound and adjust filtering thresholds toward matched drop rates. Full FROST achieves higher mean performance than this variant in all 4 settings; text-to-SQL EX is 65.92% versus 65.00% at nearly identical drop rates.
  8. Effectiveness of partial-gradient scoring (R3). On Qwen2.5-Coder-0.5B text-to-SQL, LM-head gradient-based scoring achieves 65.92% mean EX versus 65.18% with full-model gradients at similar drop rates over 5 seeds. Using full-model gradients increases training time by 59.78% but gives lower mean EX.

Updates to the manuscript during rebuttal

  • Section 3.2 (R1, R3): Clarified uniform sampling, historical calibration, and buffer updates using every batch's original pre-filter utility; added prior-work motivation for the upper utility bound.
  • Sections 4.1–4.3 (R1–R3): Added two-stage ablations, cosine comparisons, upper-bound results, and summaries of the 7B and CTR experiments; clarified industrial-model confidentiality.
  • Appendices B.4–B.5 (R3): Added end-to-end training times, selection-computation and dense-masking details, and the head/full-model scoring comparison.
  • Appendices D.1–D.4 (R1–R3): Added multi-checkpoint continuation results, the text-to-SQL upper-bound ablation, and full 7B and CTR results.

We have incorporated these results and clarifications into the revised manuscript, strengthening the evidence for FROST's effectiveness, generalizability, and practical value. All additions and revised passages are highlighted in blue for ease of reference. The manuscript was revised on November 14, 2026 (AoE).

We thank the reviewers again for their valuable feedback and respectfully ask that these additions be considered in the final assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.