acceptodds
Under review as a conference paper at ICLR 2027

Valuing Data Before Training: Geometric Coverage and Topological Fidelity Outperform Gradient-Based Selection

Abstract

Estimating the value of training data often requires training the target model, adding cost before useful subsets can be selected. We introduce C-GeoVal (Coverage- Geometric Valuation), which values samples before target-model training using frozen foundation-model embeddings. Sample value is defined as the expected marginal gain of a set utility that rewards density-weighted coverage and penalizes topological distortion. A scalable coreset procedure combines density filtering and greedy coverage maximization with persistence-guided subset repair. At an 80% data budget, C-GeoVal improves CIFAR-10N accuracy by 4.93 percentage points over full-data cross-entropy training and GoEmotions Macro-F1 by 2.06 points over full-data BERT fine-tuning. Its valuation scores detect human label errors with AUROC 0.937. Ablations show gains from topological repair in both tasks. For mathematical instruction tuning, the density-and-coverage variant outperforms DDCF at all three tested budgets on shared embeddings.

Primary area:
foundation or frontier models, including LLMs
Keywords:
data valuation, coreset selection, training-free selection, label noise, topological data analysis

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_RzA3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper targets the problem of coreset selection, where the goal is to select a subset of the training dataset while retaining as much performance as possible. The paper targets this via a frozen pretrained foundation model such as CLIP ViT-L/14 for vision, BERT's [CLS] text classification and Qwen3-Embedding-4B for LLM fine-tuning and reasoning. The approach involves extracting the embedding for the target dataset using the pretrained model. Then the proposed algorithm C-GeoVal is applied which is a 3 phase algorithm. In the first phase, the local density weight is calculated using the cosine similarity to k nearest neighbors which the paper dubs "density-weight-coverage" F~(A)\tilde{F}(A). Then the bottom 12% of the density quantile is removed. In next phase the goal is to obtain the topology component of the score for the subset AA. The second phase effectively builds the coreset using the density part of the coreset valuation score. Phase 3 looks for rare topological structure and swaps samples/points to retain structure. The paper then presents experimental results on vision (CIFAR10N), GoEmotions for text classification and AIME-24 and MATH-500 for supervised fine tuning.

Critical strengths

The idea of preserving structure in the coreset is something that can add value and is missing broadly in existing work (a few recent contemporary works attack the same problem [1, 2]).

The paper also perform the evaluation on two modes, vision and language to show the applicabilityof the method

References:
[1] Roy, Arjun, et al. "TopoPrune: Robust Data Pruning via Unified Latent Space Topology." arXiv preprint arXiv:2602.02739 (2026).
[2] Yehuda, Ofer, et al. "Active learning through a covering lens." Advances in Neural Information Processing Systems 35 (2022)

Critical weaknesses
  1. One of the major concerns is that the gains are likely coming from the base frozen embedding model. If this model has already seen the downstream eval dataset, the performance gains are hardly surprising. There is evidence in the paper that points to this, in Table 17, only the largest model CLIP ViT-L 14 outperforms random. Then there is the added confounder that the downsteam model for reasoning, Qwen3-8B is the same family as Qwen3-Embedding. Experimental setup needs to be carefully evaluated so that such confounders can be removed.

  2. The topology component, a big part of the claim is not well supported. The algorithm does not really optimize for equation 4. The landmarks are restricted to the dense region so structure in sparse regions are not captured, and rare subpoluations are likley sparse. The claims of structure/loops being rare is not supported.

  3. Some of the experimental setup needs work:

a. In GoEmotions ablation, w/o topology is 9 points blow facility location, I am not sure I understand why, should that be similar?
b. The math setup disables topology (lines 53, 75-76)
c. The mislabelled experiment is missing a bunch of baselines such as SimiFeat [1], Curvature [2], CSL [3], SSFT [4]. It is likley the noisy label works because the noise matches outlier images, a better setup woould be to use various noisy setups such as symmetric noise or otherwise. Clean lab seems to be run with ResNet18 and not CLIP features.
d. Real-noise datasets at larger scale are missing such as Red Mini-ImageNet/CNWL, WebVision, Clothing1M. And missing some older coreset baselines [5, 6, 7]
e. "Beats full data" is measured against plain cross-entropy training on noisy labels. On clean labels the coreset loses by 2.8 points to full data.

References:
[1] Zhu, Zhaowei, Zihao Dong, and Yang Liu. "Detecting corrupted labels without training a model to predict." ICML 2022.
[2] Garg et al. Memorization through the lens of curvature of loss function around samples. ICML 2024
[3] Ravikumar, et al. "Towards memorization estimation: Fast, formal and free." ICML 2025
[4] Maini, Pratyush, et al. "Characterizing datapoints via second-split forgetting." NeurIPS 2022
[5] Zheng, H., Liu, R., Lai, F., & Prakash, A. Coverage-centric coreset selection for high pruning rates. ICLR 2023
[6] I. Garg and K. Roy, "Samples with Low Loss Curvature Improve Data Efficiency," CVPR 2023
[7] Zheng, H., Liu, R., Lai, F., & Prakash, A. (2022). Coverage-centric coreset selection for high pruning rates. ICLR 2023
[8] Nagaraj, Manish, et al. "TRIM: Token-wise Attention-Derived Saliency for Data-Efficient Instruction Tuning." ICML 2026\

Questions for authors
  1. Do the results transfer to a larger real-noise dataset such as Red Mini-ImageNet/CNWL, WebVision, Clothing1M.?
  2. Does the base model have to be supervised can DinoV3 or unsupervised model be used?
  3. See weakness a, c
3 comments

Thank you for the reviewer RzA3 comments. We sincerely appreciate the valuable suggestions. We address each point below.

Q1: How can the gain from the selector be separated from the contribution of the encoder?

Response: In the main CIFAR-10N comparison, C-GeoVal and ZCore use the same CLIP embeddings and the same downstream training protocol. At the 80% budget, they reach 91.36% and 83.25%, respectively. This comparison measures the difference in selection under a given representation. Table 17 of the manuscript already tests the self-supervised DINOv2 ViT-B/14. Under this representation, C-GeoVal reaches 82.39% at the 80% budget, which is below the 83.17% of Random at the same budget in the manuscript. We will therefore state explicitly that the selector can use self-supervised representations, while the measured gain depends on the encoder. In the math experiment, C-GeoVal and DDCF share the same Qwen3-Embedding features.

Q2: What is the relationship between Algorithm 1 and Eq. (4)?

Response: Eq. (4) defines the set utility used for valuation, and Eq. (6) defines the expected marginal value of a sample. Algorithm 1 first greedily optimizes the density-weighted coverage and then performs local swaps guided by the reference persistence diagram. The approximation guarantee in Section 3.4 applies to the coverage stage, over the retained candidate set and under the fixed sparse-graph support. The landmarks are drawn from the higher-density region, so our conclusion about topological repair is limited to correcting the H1H_1 discrepancies reflected by these landmarks. For whether low-density variants are retained, please see the binned results in our response to R1. For the plain facility-location control on GoEmotions, and for why the math experiment can only evaluate the density and coverage stages, please see our response to R2.

Q3: How should the noise-robustness claim and the baseline comparison be scoped?

Response: The "exceeds full data" conclusion in the manuscript refers to standard cross-entropy training on noisy labels. On clean labels, the 80% C-GeoVal subset in the manuscript reaches 92.56%, compared with 95.36% for full-data training. The manuscript also reports results of noise-robust training on the full data, such as ELR+. The new factorial experiment tests the density preference using both the noisy labels and the clean labels of the same image pool, and the scope of the study remains limited to the existing data and noise conditions. Large-scale real-world noisy datasets, independent random label noise, and other detection baselines such as SimiFeat are controls that could further test the range of applicability. When these methods are compared, it should be stated whether each of them relies on frozen representations, observed labels, a clean validation set, or upstream training. We will also clarify in the related work the relationship to topological pruning methods such as TopoPrune, and describe our design specifically as local swaps, guided by a reference persistence diagram, applied to a coverage subset at a given budget.


We sincerely hope that this detailed response addresses the reviewer's remaining concerns, and we respectfully ask the reviewer to consider raising the score in light of the strengthened evaluation.

Thank you again for your review. We would like to clarify the scope of our results.

On CIFAR-10N, C-GeoVal and ZCore share the same CLIP embeddings and training protocol, reaching 91.36% and 83.25% at the 80% budget. This comparison isolates their difference under that representation, but the gain depends strongly on the encoder: C-GeoVal reaches 82.39% with DINOv2, below Random at 83.17%.

Algorithm 1 optimizes coverage and then applies topology-guided swaps; the approximation guarantee applies only to coverage. Density-restricted landmarks do not establish preservation of sparse rare subpopulations, and topology swaps are disabled in the math experiments.

Finally, the “beats full data” claim compares against ordinary training on noisy labels. With clean labels, C-GeoVal reaches 92.56% versus 95.36% for full-data training. Larger real-noise datasets and the additional requested baselines remain untested. We hope these clarifications help you reconsider the paper within the scope supported by its evidence.

Thank you for your response, however it does not address the fact that the gains are might be attributed to the base encoder. No new evidence was provided to show that the proposed method was the reason for the delta.

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

This paper proposes C-GeoVal, a data valuation and subset-selection framework operating on frozen embeddings without training the target model. It defines sample value through expected marginal contributions to a utility combining density-weighted coverage and topological fidelity. The practical selection algorithm performs density filtering, greedy coverage selection, and persistence-guided sample replacement. Experiments cover noisy image classification on CIFAR-10N, text classification on GoEmotions, and mathematical instruction tuning of Qwen3-8B-Base. The LLM experiments use density and coverage selection without topological repair.

Critical strengths
  1. The paper reports promising results for data selection without target-model training across image classification, text classification, and mathematical instruction tuning. The CIFAR-10N comparisons with ZCore and Prototypicality use shared CLIP embeddings and a common downstream training protocol, helping separate selection quality from representation quality.

  2. The evaluation includes useful analyses of the proposed components and their limitations. Table 7 compares topology-guided swaps with quantity-matched replacement controls. The appendices examine selection parameters, encoder dependence, noise levels, and tail-class performance. The authors explicitly acknowledge that improved overall Macro-F1 can coexist with weaker tail-class performance, and clearly distinguish the full method from the density-and-coverage variant used for LLM fine-tuning.

Critical weaknesses

The main remaining concern is the interpretation and scope of the density-based selection preference relative to ZCore [1]. ZCore’s coverage criterion can reward examples occupying distinctive, low-density regions, whereas C-GeoVal filters the lowest-density candidates and assigns greater coverage weight to dense regions. The noise-level sweep and tail-class analysis provide useful evidence for this design, and the paper appropriately acknowledges the associated trade-offs. In particular, the clean-label results show that its benefits are not confined to the primary noisy-label setting.

Nevertheless, the current results do not directly establish which low-density examples are being deprioritized: mislabeled examples, redundant examples, or correctly labeled examples representing useful rare variation. Lower retained noise and higher overall accuracy support the method’s effectiveness on the evaluated datasets, but do not fully explain why its preference is beneficial relative to ZCore’s coverage principle. Similarly, the tail-class results identify a relevant trade-off, but class frequency is not equivalent to embedding density and therefore does not resolve this question by itself.

Although the reported gains are promising, I am not yet sufficiently convinced by the justification for the method’s central density preference. Distinguishing noise suppression from the loss of useful rare examples is important to assessing its contribution relative to existing training-free selectors. I therefore currently lean toward weak rejection. A convincing explanation, supported where possible by an analysis of the existing selections, could address this concern and lead me to reconsider my recommendation.

Reference:

[1] Griffin, Marks, and Corso. Zero-Shot Coreset Selection via Iterative Subspace Sampling. WACV 2026. Paper

Questions for authors
  1. Could you analyze the existing CIFAR-10N selections by embedding-density range, separating mislabeled from correctly labeled examples, and compare retention under ZCore and C-GeoVal? Where feasible, identifying correctly labeled examples from underrepresented subpopulations would help explain the trade-off between noise suppression and distinctive coverage.

  2. How does topology-guided repair change this retention pattern? Comparing the same samples before and after repair would help clarify whether its benefit comes from preserving rare structure, improving representation within already-covered regions, or another effect. Please also clarify whether replacement candidates can include samples removed by the density prefilter.

  3. Based on the existing noise-level and tail-class results, what practical guidance can the authors give about when the density preference is likely to be beneficial and when it may be less suitable? A discussion grounded in the observed results would be useful; a universal rule for choosing between selectors is not necessary.

4 comments

Thank you for the reviewer XGJB comments. We sincerely appreciate the valuable suggestions. We address each point below.

Q1: How large is the gain of H1H_1 swapping compared to quantity-matched swap controls?

Response: Our additional experiments on GoEmotions start from the exact same Phase 2 subset and compare different swap rules under identical budgets and candidate pool constraints. We include "Random swap-out + densest-neighbor swap-in" to test the effect of the swap-out rule while keeping the density swap-in rule fixed. Facility location and Random are provided as external references.

Table R3: Macro-F1 and Weighted-F1 (%) of quantity-matched swap on GoEmotions

Selection or Swap RuleMacro-F1 · 20%Macro-F1 · 60%Macro-F1 · 80%Weighted-F1 · 20%Weighted-F1 · 60%Weighted-F1 · 80%
No Swap38.02 ± 0.7545.38 ± 0.0947.02 ± 0.4655.59 ± 0.5458.11 ± 0.2458.57 ± 0.46
H1H_1 Swap38.91 ± 0.4546.79 ± 0.1447.09 ± 0.7054.78 ± 0.1757.13 ± 0.1658.07 ± 0.18
Random Swap38.25 ± 1.0846.01 ± 0.1147.01 ± 0.6955.62 ± 0.4558.24 ± 0.1758.92 ± 0.12
Density Swap38.44 ± 0.2345.32 ± 0.4446.39 ± 1.0255.74 ± 0.6658.05 ± 0.0958.58 ± 0.79
Random out + Densest-neighbor in38.08 ± 0.5645.28 ± 1.0446.36 ± 0.4455.56 ± 0.4558.92 ± 0.2358.61 ± 0.42
Facility location37.69 ± 1.0345.09 ± 0.5046.96 ± 1.2355.12 ± 0.5858.12 ± 0.2159.07 ± 0.40
Random38.34 ± 1.0445.45 ± 0.8946.38 ± 0.8555.69 ± 0.4458.46 ± 0.5958.83 ± 1.05

Comparing against the quantity-matched swap controls with the highest mean Macro-F1 at each budget, H1H_1 swapping achieves mean advantages of 0.47, 0.78, and 0.08 percentage points at 20%, 60%, and 80%, respectively. At the 60% budget, H1H_1 reaches 46.79, outperforming No Swap (45.38) by 1.41 points and Facility location (45.09) by 1.70 points. At the 80% budget, the mean Macro-F1 of H1H_1 is close to that of Random Swap. Evaluated by Weighted-F1, H1H_1 is lower than every quantity-matched swap controls across all three budgets. The comparisons reported here are based on mean values; cross-seed stability requires evaluation via per-seed differences.

Q2: How should we interpret the tail-class results in relation to the two F1 metrics?

Response: Macro-F1 weights all classes equally, whereas Weighted-F1 weights classes by their support. If an application prioritizes protecting tail classes, one should also look directly at tail-class F1 and tail-class retention. In Table 19 of the original manuscript, topological repair at the 80% budget improves tail-class F1 from 21.15 to 27.33 (compared to 35.39 for Random), while the tail-class retention rates for post-repair C-GeoVal and Random are 45.8% and 80.1%, respectively. The overall Macro-F1 advantage in the original manuscript primarily stems from head and medium classes. A testable extension is to set a minimum retention budget for the tail classes defined by the observed labels and simultaneously evaluate tail-class F1, overall F1, and annotation disagreement. Table R3 provides the overall F1 for the new swap controls, whereas tail-class results still refer to the corresponding experiments in the original manuscript.

Q3: How about encoder dependence, runtime breakdown, and the math experiments?

Response: We will highlight encoder sensitivity in the main text (see our response to R3 for detailed results). Regarding runtime on CIFAR-10N, 29 seconds is the selection time given pre-computed embeddings, while there is an additional one-time CLIP embedding extraction cost of 195 seconds, totaling approximately 224 seconds across both stages. Random requires no such preprocessing, and K-Center in the original manuscript uses upstream from training features. For the math experiments, the 1K, 10K, and 20K settings are subset sizes selected from the same 189K candidate pool.


We sincerely hope that this detailed response addresses the reviewer's remaining concerns, and we respectfully ask the reviewer to consider raising the score in light of the strengthened evaluation.

Thank you for Table R2, which substantially addresses my density-bin concern. One clarification: for GoEmotions Macro-F1 at the 60% budget, Table 5 reports w/o topology → full as 35.36 → 46.14 (+10.78), while Table R3 reports No Swap → H1 Swap as 45.38 → 46.79 (+1.41). How do the settings or ablation definitions differ, and how should the topology contribution be interpreted?

Thank you for the thoughtful acknowledgment and follow-up question. We provide the following clarification.

Q1: How do the settings or ablation definitions differ?

Response: Table 5 evaluates the text configuration end to end, whereas Table R3 isolates the swap rule, and for that it reuses the quantity-matched protocol of Table 7, which is defined on the label-free pipeline. Table 5 uses observed-label density stratification (Appendix B): a per-class density score, a per-class 12% prefilter, facility location on 40% of the budget with a score-based fill for the rest, then Phase 3. Table R3 uses the label-free pipeline of CIFAR-10N: global kNN density, a global 12% prefilter and density-weighted facility location over the whole budget, then the swaps. The "w/o topology" row of Table 5 and the "No swap" row of Table R3 are therefore different subsets

Q2: How should the topology contribution be interpreted?

Response: Topological repair is a bounded, local adjustment applied after the subset has been chosen. It is not the component that determines performance.


Given the limited time, we sincerely hope that this detailed response addresses the reviewer's remaining concerns, and we respectfully ask the reviewer to consider raising the score in light of the strengthened evaluation.

Thank you for your follow-up. Our previous reply did not sufficiently explain the topology results.

Table 5 uses the manuscript’s observed-label-stratified text pipeline, whereas Table R3 uses a different label-free pipeline. Their no-topology subsets are therefore different. Within Table R3, all swap rules start from the same Phase 2 subset: at 60%, H1H_1 swaps improve Macro-F1 from 45.38 to 46.79, with the best matched swap control at 46.01. At 80%, the advantage over the best matched control is 0.08 points. H1H_1 swaps have lower Weighted-F1 than all matched controls across the three budgets.

Thus, the matched evidence supports a configuration- and metric-dependent benefit. Given the density analysis and this clarification, we would be grateful if you would reconsider your recommendation.

Reviewer_X6MtOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  • This draft proposes C-GeoVal (Coverage-Geometric Valuation), a method for selecting training data coresets before any target model training.
  • The key idea is to use frozen foundation model embeddings and define sample value through two criteria: density-weighted coverage (selecting representative samples from dense regions) and topological fidelity (preserving the geometric structure of the data manifold using persistent homology).
  • The method works in three phases: first filtering out low-density outliers, then greedily maximizing a facility-location coverage objective, and finally applying a topology-guided repair step.
  • The draft evaluates C-GeoVal on noisy image classification (CIFAR-10N), text classification (GoEmotions), and mathematical instruction tuning for an 8B LLM (Qwen3-8B).
Critical strengths
  1. The observation that existing data valuation methods require training a model is a real and important limitation. Gradient-based scores computed on corrupted labels can inherit the noise they are supposed to detect. The training-free framing is clearly motivated and practically significant.

  2. Using H1 persistence diagrams as a fidelity criterion for subset selection is new in the data valuation and coreset literature. The topological guardrail appears to be a good idea, too, and the ablation shows it adds value beyond quantity-matched swap controls.

  3. Evaluating on noisy image classification, crowdsourced text classification, and LLM instruction tuning for mathematical reasoning is wide. The math result is particularly striking: a 20K subset (10.6% of the pool) outperforms full-corpus supervised fine-tuning on MATH-500.

  4. Selection takes 29 seconds for CIFAR-10N after embedding extraction, with only 0.94 GB GPU memory, compared to 2.6 hours and 3.9 GB for TRAK. This is a meaningful practical advantage.

  5. The daft tests hyperparameter sensitivity, encoder dependence, different noise levels, and swap controls. The draft includes some results where the method performs poorly.

  6. The draft provides theoretical grounding through a marginal value decomposition, a submodularity proof, and a sketched greedy approximation guarantee, along with a connection to the broader semivalue framework; however, I have not independently verified the correctness of these proofs.

Critical weaknesses
  1. On image datasets, the topology-guided repair adds only +0.48 accuracy points at every budget, a modest but consistent gain. On text, removing it costs 8-11 Macro-F1 points, a massive drop. This nontrivial discrepancy is not well explained. The per-class analysis shows that the guardrail helps the head and medium classes but leaves the tail F1 well below that of random sampling.

  2. At the 80% budget, C-GeoVal achieves Macro-F1 of 48.71 (above full-data training at 46.65) but Weighted-F1 of only 57.53 (below random at 58.84 and full-data at 58.79). C-GeoVal improves average per-class F1 but not class-frequency-weighted F1. Does this mean its gains come mainly from head and medium classes, not from the rare tail classes?

Minor

  1. Experimental results show that performance drops substantially with weaker encoders: from 91.36% (CLIP ViT-L/14) to 82.60% (CLIP ViT-B/32) and 82.39% (DINOv2 ViT-B/14) at the 80% budget on CIFAR-10N. A 9-point gap due to encoder choice is large, and many practitioners may not have access to or budget for CLIP ViT-L/14. The method’s strong results are therefore partly attributable to using a very powerful encoder, and this dependence should be discussed more prominently in the main paper.

  2. The draft notes that embedding extraction (195 seconds for CIFAR-10N) is excluded from the 29-second selection time and shared across methods. But for methods such as Random or K-Center that do not need embeddings, this is not a shared cost. A better cost comparison would include embedding extraction time for methods that require it.

Questions for authors
  1. Can you explain why removing topology costs 8–11 Macro-F1 points on text but only 0.48 accuracy points on images? Is this a property of the datasets, the encoders, or the topology computation itself?

  2. The tail retention at 80% on GoEmotions is 45.8% for C-GeoVal versus 80.1% for random. This means C-GeoVal is systematically excluding rare-class examples. For applications where rare classes are important (e.g., medical diagnosis), this could be harmful. Can you propose a modification that better balances coverage with rare-class retention?

  3. The Weighted-F1 result shows C-GeoVal trails random and full-data training at the 80% budget. Should practitioners prefer Macro-F1 or Weighted-F1 as the evaluation criterion, and how should they decide whether to use C-GeoVal for text classification?

  4. The math experiment disables topology. At what pool size does topology become computationally feasible? Have you measured what topology would add at the 1K or 10K budget, where the embedding space is smaller, and persistence is cheaper to compute?

3 comments

Thank you for the reviewer X6Mt comments. We sincerely appreciate the valuable suggestions. We address each point below.

Q1: How much do hard prefilter and density weighting each contribute?

Response: On CIFAR-10N, we disable topological repair and cross the hard prefilter (on / off) with the coverage weights (uniform / density) under the same embeddings and budgets, each subset is trained once with the human noisy labels and once with the clean reference labels. All comparisons in Table R1 are conducted within each respective label condition and budget.

Table R1: Test accuracy (%) of the factorial ablation on CIFAR-10N

Selection ConfigurationNoisy Labels · 60%Noisy Labels · 80%Clean Labels · 60%Clean Labels · 80%
No prefilter + Uniform weights81.30 ± 0.1582.01 ± 0.3992.27 ± 0.4494.13 ± 0.36
No prefilter + Density weights82.65 ± 0.2782.55 ± 0.4693.11 ± 0.5894.93 ± 0.13
Prefilter + Uniform weights81.57 ± 0.2783.22 ± 0.3092.31 ± 0.1894.28 ± 0.11
Prefilter + Density weights82.80 ± 0.0983.51 ± 0.3493.65 ± 0.2894.99 ± 0.15
Random82.51 ± 0.3782.04 ± 0.5592.07 ± 0.3693.85 ± 0.20

Without hard prefilter, density weighting increases the mean test accuracy of noisy-label training by 1.35 and 0.54 percentage points at the 60% and 80% budgets, respectively, and improves clean-label training by 0.84 and 0.80 percentage points. With density weighting enabled, adding hard prefilter at the 80% budget yields mean differences of +0.96 and +0.06 points for noisy-label and clean-label training, respectively. The gains under clean labels indicate that the benefit of density weighting comes at least in part from which images are selected. Its relationship with mislabeled sample filtering should be interpreted in conjunction with the retention results below. When both hard prefilter and density weighting are enabled, the configuration outperforms the re-run Random baseline by 0.29, 1.47, 1.58, and 1.14 points across the four conditions, respectively.

Q2: What samples do C-GeoVal and ZCore retain in low-density regions?

Response: We rank and bin all 50,000 images using the same set of CLIP local densities. After subset selection, we use the clean reference labels to identify mislabeled samples. Each retention rate in Table R2 is computed using the number of correctly labeled or mislabeled samples within that specific bin as the denominator.

Table R2: Per-bin retention rates (%, correctly labeled / mislabeled) on CIFAR-10N at the 80% budget

Density Percentile (Low to High)Bin SizeBin Mislabeling Rate (%)RandomFacility locationZCoreC-GeoVal (Pre-repair)C-GeoVal (Post-repair)
0–12% (Pre-filtered)6,00035.280.2 / 80.998.9 / 99.379.9 / 80.70.0 / 0.00.0 / 0.0
12–23%5,50030.680.0 / 79.498.2 / 98.579.1 / 79.430.6 / 27.737.3 / 34.3
23–34%5,50026.278.7 / 79.597.0 / 96.179.6 / 80.598.4 / 97.999.7 / 99.5
34–45%5,50021.979.8 / 80.395.2 / 95.879.4 / 80.399.5 / 99.793.9 / 94.1
45–56%5,50017.680.3 / 81.391.3 / 91.080.3 / 80.899.8 / 99.797.7 / 97.6
56–67%5,50013.580.2 / 79.783.7 / 86.279.6 / 80.799.9 / 100.099.5 / 99.7
67–78%5,5009.480.1 / 80.371.1 / 72.279.9 / 80.699.9 / 99.999.8 / 99.8
78–89%5,5005.480.5 / 78.849.3 / 52.080.6 / 79.8100.0 / 100.099.9 / 100.0
89–100%5,5001.779.9 / 80.732.3 / 33.480.7 / 81.299.8 / 100.0100.0 / 100.0

In the lowest-density 12% bin, the mislabeling rate is 35.2%, roughly twice the full-pool mislabeling rate of 18.1%. Estimated by the in-bin ratio, this region still contains approximately 3,900 correctly labeled images. ZCore retains about 80% of both correctly labeled and mislabeled samples in this bin. In contrast, C-GeoVal's hard prefilter redirects the budget toward the remaining regions. This comparison quantifies the trade-off between the two strategies: when the low-density region is enriched with mislabeled data, reducing selection from this region suppresses the retention of these mislabeled samples. However, if a task relies on valid variants in the low-density region, the same rule also reduces the chance of including those variants in the subset. While the binning results identify correctly labeled low-density samples, determining whether they form useful rare subgroups requires subgroup definitions and downstream evaluation.

Q3: Does topological repair swap pre-filtered samples back in?

Response: At the reporting precision of Table R2, the retention rates for both types of samples in the lowest-density bin remain 0.0% before and after topological repair. In the adjacent 12–23% bin, the retention rates of correctly labeled and mislabeled samples increase from 30.6% to 37.3% and from 27.7% to 34.3%, respectively, showing a net increase for both groups in that bin.


We sincerely hope that this detailed response addresses the reviewer's remaining concerns, and we respectfully ask the reviewer to consider raising the score in light of the strengthened evaluation.

Reviewer_X6MtOpenReview ↗

Retaining the rating, but not championing

I thank the authors for providing their rebuttal (which I find selective). However, after carefully reviewing the discussion, I agree with the other reviewers regarding the reliability/robustness of the performance gains offered by the approach. Under these circumstances, I would not champion this paper for acceptance, but I would suggest that the authors revise it based on these discussions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.