acceptodds
Under review as a conference paper at ICLR 2027

Mind the Tail: Dataset Selection Through Sparse Concepts

26 views

Abstract

Data selection reallocates training support across visual concepts, including objects, attributes, color, texture, and backgrounds. Different allocation strategies may produce varied downstream model behavior, especially for tail concepts supported by only a few samples. Discarding them can hurt test-time robustness, while overemphasizing samples lying on category boundaries may introduce unstable training signals. This work explicitly investigates this relationship across fine-grained, spurious-correlation, and general image classification, as well as multimodal pretraining. Specifically, we use sparse autoencoders (SAEs) to extract concept activations from individual images. During selection, we accumulate the dataset-level concept distribution for each added sample. Diminishing returns are applied to this distribution to down-weight well-represented concepts and move the focus to the distribution tail. We then incorporate local support to find samples that fill the distributional gap. Empirical results show consistent improvements over strong data selection baselines across all benchmarks. We also reveal the connection between concepts our method enhances and the induced behavioral changes, including robustness against out-of-distribution and spurious contexts. This work establishes concept support, beyond individual sample importance, as a useful principle for data-efficient and robust learning.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Data selection

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_tz3AOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a data selection method that uses sparse autoencoders (SAEs) to decompose pretrained DINOv3 features into sparse, more interpretable concept activations. It then applies submodular optimization with diminishing returns on accumulated concept support, which reallocates the training budget toward under-represented concepts rather than preserving the source distribution, aiming to improve downstream generalization and robustness. A local-support term additionally credits samples whose concepts recur among their neighbors, to avoid overemphasizing isolated rare samples.

Critical strengths
  1. Strong performance on fine-grained and spurious-correlation benchmarks at low selection rates (e.g., +3.2 accuracy on NABirds at 10%, and +7.6/+7.5 WGA on Waterbirds at 10%/30% over the strongest baseline).

  2. Comprehensive experiments and ablations (representation choice, local-support weight, neighborhood size, utility function, SAE pretraining data, downstream architecture). These make the contribution of individual components reasonably traceable, which increases my confidence that the gains do not hinge on a single design choice.

Critical weaknesses

I do not see any major weaknesses that would independently justify rejection.

However, the method implicitly assumes that SAE sparsification makes individual dimensions more semantically meaningful than the original DINOv3 features. The paper offers some support (Fig. 4 shows multi-modal per-dimension activation distributions; Table 16 shows that K-Center also benefits from SAE features), but this evidence is descriptive. A multi-modal activation distribution does not by itself establish that the dimensions correspond to more meaningful concepts, nor does it explain why the concept-coverage objective benefits from them. A more mechanistic or theoretical explanation would strengthen the paper.

Questions for authors

1. Fairness of feature extractors.

The paper does not state which feature extractor or scoring model each baseline uses, only that the original protocols are followed. Could the authors list this explicitly for every baseline?

I note that K-Center (whose Table 1 numbers match K-Center (DINO) in Table 16), RAM-APL, and SDM already rely on foundation-model features. However, training-dynamics-based baselines (Forgetting, EL2N, Moderate, CCS, D2-Pruning, InfoMax) derive their scores from early training of the downstream model, which is much weaker than DINOv3. For these comparisons, it is difficult to rule out that part of the improvement comes from the stronger representation rather than the selection objective. Could the authors report, for at least a few of these baselines (e.g., CCS and D2-Pruning), a variant whose scores are computed from DINOv3 features (e.g., via a linear probe on DINOv3)?

2. Within-class vs. global submodular optimization.

After inspecting the released code, I noticed that all labeled classification datasets use independent within-class submodular optimization, while CC12M uses global selection. This is not described in Section 3, which presents a single global objective.

My concern is that concept support is then not accumulated across classes. For example, if bird species A and B both frequently appear on branches, global selection would account for the accumulated support of this concept across both classes, whereas within-class selection treats them independently, so the support from species A does not affect selection within species B. This makes the optimized quantity class-conditional concept coverage, while the paper's framing and analyses (Fig. 1, Table 4) concern dataset-level concept support.

Could the authors: (a) explain the motivation for this design and state it explicitly in the paper; (b) clarify how per-class quotas are set, and whether the baselines use the same per-class allocation, since per-class allocation can itself act as class balancing on imbalanced datasets such as NABirds; (c) report results with global selection on at least one classification dataset (e.g., NABirds or Waterbirds)? This seems computationally feasible given that global selection on CC12M takes about 1.3 hours (Appendix A.3).

In particular, on Waterbirds, within-class selection may itself drive the increase in bias-conflicting examples (Table 5): once the water-background concept saturates within the waterbird class, the objective naturally turns to land-background waterbirds. It would be informative to know whether the robustness gains persist under global selection.

5 comments
AuthorsOpenReview ↗

Rebuttal part 1 of 3

Dear reviewer,

We sincerely thank the reviewer for the constructive feedback and for recognizing our strong low-budget results and comprehensive experiments and ablations. Our responses are listed below.

W. Multimodal activation histograms and K-Center gains do not establish that SAE latents are more semantically meaningful than DINOv3 features. Please clarify this assumption and provide a mechanistic or theoretical explanation of why the concept-coverage objective benefits from SAE activations.

Thank you for raising this point. We first clarify that our method does not assume SAE latents are more semantically meaningful than DINOv3 features. Monosemanticity is a goal of SAE feature decomposition; our method uses the resulting activation patterns to measure support for recurring visual properties, without requiring each latent to correspond to a single human-interpretable concept.

The main advantage of sparse representations lies in how they interact with our concept-coverage objective. In dense representations, individual dimensions can receive activations from many images with different visual characteristics, making their accumulated support less discriminative. In contrast, SAE sparsity allows each image to contribute to a limited subset of dimensions, providing more selective measurements of recurring and under-supported visual patterns. This enables our objective to distinguish concepts that have already received substantial training support from those that remain underrepresented.

More specifically, for a candidate image ii of class yy, the marginal gain is

Δy(i∣Sy)=∑c=1Clog⁡!(1+x~i[c]1+My,c(Sy)),My,c(Sy)=∑j∈Syx~j[c].\Delta_y(i\mid S_y)=\sum_{c=1}^{C}\log\\!\left(1+\frac{\tilde{x}_i[c]}{1+M_{y,c}(S_y)}\right),\qquad M_{y,c}(S_y)=\sum_{j\in S_y}\tilde{x}_j[c].

Here, x~i\tilde{x}_i includes local support. The same activation receives a larger reward when its latent has less accumulated support. Sparse activations make this allocation more selective across visual patterns, while diminishing returns dynamically adjust their importance throughout selection.

Figure 3 and Table 14 test this connection, with SAE local support fixed; using SAE latents for concept coverage raises NABirds accuracy at 10% selection budget from 35.4% to 42.9% compared with DINO features. Table 21 also shows that SAE features improve K-Center in five of six settings, while our objective gives further gains. These experiments establish the representation's usefulness for this selection objective. We use the multi-peaked activation histograms to describe the distribution of activations.

We have added a discussion of how SAE activations support concept allocation in Appendix D.5 of the revised manuscript.

AuthorsOpenReview ↗

Rebuttal part 2 of 3

Q1. Please specify the feature extractor or scoring model used by each baseline. Could part of the gain come from stronger representations rather than the selection objective? Please report DINOv3-based scoring variants for a few baselines, such as CCS and D², for example using linear probes.

Thank you for suggesting this comparison. We trained linear classifiers on frozen DINOv3 features and used their prediction errors (EL2N scores) to run CCS and D², with matched class quotas and downstream training settings. The original and stronger-feature results are listed together below; each cell gives the 10% / 30% / 50% results, averaged over three runs.

Method and feature settingNABirds accuracy (%)Waterbirds WGA (%)
CCS, ResNet-50 scores39.7 / 65.5 / 73.652.8 / 67.1 / 69.9
CCS, DINOv3 linear-probe EL2N scores35.4 / 64.2 / 73.748.8 / 75.1 / 79.1
D², ResNet-50 scores and features25.8 / 66.6 / 74.747.0 / 56.4 / 60.8
D², DINOv3 linear-probe EL2N scores and features40.6 / 65.8 / 73.362.0 / 69.9 / 72.6
K-Center, DINOv3 patch average31.6 / 64.9 / 74.157.5 / 64.2 / 79.4
K-Center, DINOv3 CLS26.3 / 63.2 / 74.656.1 / 73.7 / 75.4
Ours42.9 / 67.4 / 75.268.6 / 77.9 / 79.3

The DINOv3-based variants show mixed gains: at 10% on NABirds, D² improves from 25.8% to 40.6%, while CCS changes from 39.7% to 35.4%. Ours reaches 42.9% and outperforms both variants at every budget on both datasets, even with their stronger features and additional classifier-training cost.

The table below lists the main-result configurations and additional baseline variants.

MethodScoring model / criterionFeatures for distance or graph
RandomUniform sampling—
ForgettingResNet-50 forgetting events—
EL2NResNet-50 probability-error norms—
K-CenterGeometric coverageDINOv3 patch average (main); DINOv3 CLS or SAE activations (additional)
ModerateClass-median distanceResNet-50
CCSResNet-50 forgetting scores (main); DINOv3 linear-classifier EL2N scores (additional)ResNet-50 (main)
D²ResNet-50 forgetting scores (main); DINOv3 linear-classifier EL2N scores (additional)ResNet-50 (main); DINOv3 (additional)
Supervised InfoMaxClass-normalized forgetting scoresL2-normalized ResNet-50
RAM-APLCLIP ViT-L/14 and DINOv2-baseCLIP / DINOv2
SDMSAE concept-distribution matchingImage-level SAE activations
OursAccumulated SAE concept supportDINOv3-B → SAE

For multimodal pretraining, CLIPScore and its score-based baseline variants use CLIP ViT-B/16 image–text similarity.

In the revised manuscript, we have added the baseline configurations in Appendix A.2, Table 7, and the stronger-feature comparisons in Section 4.2, Table 3(b), and Appendix D.7, Table 23. Due to time constraints, these additional comparisons currently cover NABirds and Waterbirds at 10%, 30%, and 50% selection; we will add the complete results to the main tables in the future.

AuthorsOpenReview ↗

Rebuttal part 3 of 3

Q2. The code uses within-class selection for labeled datasets and global selection for CC12M, while Section 3 presents a global objective. Please explain and document this design and reconcile class-conditional coverage with dataset-level support analyses. How are per-class sample counts set, and do baselines match them to control for class balancing? Please report global-selection results on a classification dataset and assess whether the Waterbirds robustness gains persist under global selection. Could within-class selection itself increase bias-conflicting examples and contribute to those gains?

Thank you for asking about this design. We use within-class support accumulation because shared contexts need training support in different classes. For example, seeing a hand with species A does not give species B training examples in the same context. With global support accumulation, hand examples from A lower the reward for adding hand support in B, even if B has no such examples yet. Within-class accumulation preserves the reward for support across classes.

For labeled classification, neighbors and support accumulation are class-specific, with proportional quotas By=max⁡(1,⌊rny⌋)B_y=\max(1,\lfloor r n_y\rfloor), where nyn_y is the class size and rr is the selection rate. Classification baselines use the same class quotas; global controls match the total budget ∑yBy\sum_y B_y. Figure 1 and Table 5 report raw SAE support aggregated across all classes after selection, summarizing the dataset-level allocation produced by the class-conditional objective.

We compare within-class and global selection on NABirds and Waterbirds. Within-class selection selects separately within each class. Global selection with fixed class counts selects across classes while keeping the same per-class sample counts and neighbor graphs. Global selection uses a global graph without fixing class counts. Random (global) samples uniformly from the full pool. All settings use the same total budget and downstream training. Each entry below reports the 10% / 30% / 50% results.

Selection modeNABirds accuracy (%)Waterbirds WGA (%)
Within-class selection42.9 / 67.4 / 75.268.6 / 77.9 / 79.3
Global selection with fixed class counts32.7 / 63.4 / 72.548.6 / 53.3 / 65.7
Global selection30.1 / 59.7 / 71.459.8 / 62.8 / 72.0
Random (global)35.2 / 63.9 / 73.551.6 / 63.4 / 74.9

Within-class selection performs better at every budget when the class counts and graphs are fixed, isolating the contribution of separate support accumulation.

The hand concept occurs in 108 classes in the full NABirds training pool. The table below compares hand-context class coverage under within-class selection and global selection with fixed class counts.

Selection budgetWithin-class selectionGlobal selection with fixed class counts
10%50 / 108 classes37 / 108 classes
30%74 / 108 classes64 / 108 classes
50%88 / 108 classes80 / 108 classes

Within-class selection retains hand-containing images in more classes at every budget, even with the same number of samples for each class. On Waterbirds, Table 6 further shows that matching all four label–background group counts raises Random’s WGA from 52.3% to 63.7%, while Ours reaches 68.6%. Thus, a more balanced background composition explains part of the gain, but which images provide support within each group also matters.

For large-scale image–text pretraining on CC12M, we use global support accumulation to promote broad concept coverage across diverse image–text pairs. This matches the pretraining goal of learning visual–language associations across the full data pool. Global accumulation reduces the marginal reward for repeatedly supported patterns and directs the selection budget toward under-supported concepts across the pool.

In the revised manuscript, we have clarified the implementation in Section 3.2 and baseline allocations in Appendix A.2, and added the controlled comparisons and hand-context analysis in Appendix D.3, Tables 18–19.

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

The paper treats data selection as moving training support across visual concepts. A TopK sparse autoencoder (SAE) runs on frozen DINOv3 features, and each image becomes a sparse concept vector. Selection process tries to ensure that concepts that are already well covered give less reward, so the selection moves toward rare concepts. Moreover it favors rare concepts that are backed by similar images over rare concepts that appear alone.

Critical strengths
  1. The starting observation is useful. Fig. 1 and Table 4 show that K-Center, Forgetting, and D2 all produce almost the same concept-support profile as random (prevalence ratio 0.97 vs 1.00). This says something real about the whole "representative subset" line of work.

  2. The method is simple and the optimization is sound. The objective is submodular and lazy greedy gives a (1−1/e) guarantee.

Critical weaknesses
  1. The main claim is too broad: The paper says data selection "need not preserve the source distribution." This is only true when the test set weights concepts differently from the training set. When train and test match, keeping the source distribution is the right target. ImageNet-1K result shows this. In Table 19 (appendix only), CCS matches or beats the method on the main metric at two of three rates (50.1/64.8/68.4 vs 49.2/65.2/67.7).

  2. It is not chencked whether the "concept" extracted by SAE is aligned withthe task. The SAE is trained without labels, so nothing guarantees its directions match the evidence that limits accuracy. And only few hand-picked examples are shown in the main paper.

  3. Selection never uses labels, and every quantity is a marginal count. Nothing measures how often a concept appears with a class. So the method cannot tell a useful rare concept from a rare concept that is a shortcut. It's suggested to build a version where the shortcut is rare but perfectly predictive, and measure worst-group accuracy.

Questions for authors

How is β\beta chosen without labels? What happens with one fixed β\beta for all datasets?

Does the method still help when the spurious attribute is rare but perfectly predictive?

Every table says "averaged over three runs" but reports no variance.

After rebuttal.

Thanks for author's rebuttal, my concerns on when training and test distributions match have been solved. But my concerns on how SAE extracted rare concepts help with downstream task still remains, simply comparing the proposed method with random sampling does not address this issue. It needs some explanations that are more mechanic.

5 comments
AuthorsOpenReview ↗

Rebuttal part 1 of 2

Dear reviewer,

We sincerely thank the reviewer for the constructive feedback and for recognizing the value of our concept-support observation, the simplicity of our method, and the soundness of the optimization. Our responses are listed below.

W1. The claim that data selection need not preserve the source distribution is too broad. When training and test distributions match, selection should preserve that distribution. The ImageNet-1K results support this concern, CCS matches or outperforms the proposed method at two of three selection rates.

Thank you for raising this question. To better understand the relationship between train–test distribution differences and the effectiveness of concept allocation, we conducted an additional distribution analysis across the adopted datasets. MMD² [1] measures differences between feature distributions, while C2ST [2] measures how accurately a classifier can distinguish training from test images.

DatasetMMD²C2ST
NABirds25.2 ± 15.71.3 ± 1.5
Stanford Cars23.1 ± 27.20.1 ± 0.8
ImageNet-1K64.8 ± 8.73.1 ± 2.1

Although NABirds and Stanford Cars show smaller measured train–test discrepancies than ImageNet, concept-based selection brings larger improvements instead. Thus, the benefits of adjusting concept support do not simply correlate with a larger train–test distribution gap.

On ImageNet-1K, our method surpasses Random by 0.6% / 1.0% / 0.5% at the three budgets (Table 27). As discussed in Appendix E.1, random subsets may already retain much of the evidence needed for standard classification, leaving less room for concept-support redistribution. Meanwhile, the same selected subsets yield better results on robustness benchmarks like ImageNet-A, ImageNet-R, and ImageNet-D at all three budgets. This conclusion is closely related to our case study on NABirds (Section 5.1), where the enhanced concepts may correspond to class-agnostic background features but improve model robustness.

Figure 5(b) shows larger recognition gains for concept groups receiving increased training support; the hand case illustrates that these gains also include shared contexts. These results suggest that our method helps by increasing training support for useful visual patterns, including shared contexts, even on datasets with smaller measured train–test gaps.

We agree that the Introduction's statement that data selection “need not preserve the source distribution” is too broad. In the revised manuscript, this now reads: “Under a limited training budget, data selection should allocate training support to evidence most useful for downstream learning by adjusting concept support.”

We have added the train–test gap diagnostics and their setup in Appendix E.1, Table 26 of the revised manuscript.

[1] Arthur Gretton et al. A Kernel Two-Sample Test. Journal of Machine Learning Research, 13:723–773, 2012.

[2] David Lopez-Paz and Maxime Oquab. Revisiting Classifier Two-Sample Tests. ICLR, 2017.

W2. SAEs are trained without labels, so their concepts are not guaranteed to capture task-relevant evidence. Please provide systematic evidence of task alignment beyond a few selected examples.

Thank you for raising this point. We deliberately train the SAE on patch tokens to capture recurring local visual properties, including both category-discriminative attributes and category-agnostic contexts. The relevance of these concepts to downstream learning is not limited to their direct association with category labels. Useful training support also includes the contextual variations under which a category must be recognized. For example, seeing a bird species in the presence of a hand provides training support for recognizing that species in the same context at test time.

We illustrate this through the hand concept in the 10% NABirds subset:

Hand supportFull training poolRandomOurs
Images1602060
Species covered1082050

Accuracy on strongly hand-activated test images increases from 28.6% to 51.4%. This example shows how increasing support for a category-agnostic contextual concept can be associated with improved recognition under that context.

Importantly, our evidence extends beyond selected examples. Figure 5(b) systematically examines the relationship between concept support and downstream recognition across NABirds and CC12M. We first group SAE concepts according to whether their training support is enhanced, unchanged, or reduced relative to random selection, independently of downstream accuracy. Across NABirds and CC12M, images associated with enhanced concepts show the largest recognition gains. The systematic analysis and the hand case together support that enhanced concepts are associated with larger downstream recognition gains.

We have added the comparison of hand-context coverage in Appendix D.3, Table 19 of the revised manuscript.

AuthorsOpenReview ↗

Rebuttal part 2 of 2

W3 / Q2. Can selection based on marginal concept counts, without labels or concept–class association statistics, distinguish useful rare concepts from shortcuts? Please test a rare but perfectly predictive shortcut and report worst-group accuracy.

Thank you for suggesting this test. We added a shortcut marking to 1%, 2%, or 5% of positive-class Waterbirds training images (11, 22, or 55 images). The cue is a 32 × 32 checkerboard inserted after augmentation.

The table below reports cue selection results. Each cell shows cue count / selected-set share (%) / enrichment relative to the full training pool (×). “1% cue,” for example, means cue presence in 1% of the positive training class.

Cue conditionMethod10% selection30% selection50% selection
1% cueRandom0 / 0.0% / 0.0×6 / 0.4% / 1.8×7 / 0.3% / 1.3×
1% cueOurs1 / 0.2% / 0.9×4 / 0.3% / 1.2×7 / 0.3% / 1.3×
2% cueRandom2 / 0.4% / 0.9×8 / 0.6% / 1.2×10 / 0.4% / 0.9×
2% cueOurs2 / 0.4% / 0.9×8 / 0.6% / 1.2×13 / 0.5% / 1.2×
5% cueRandom8 / 1.7% / 1.5×12 / 0.8% / 0.7×29 / 1.2% / 1.1×
5% cueOurs3 / 0.6% / 0.6×17 / 1.2% / 1.0×35 / 1.5% / 1.3×

Ours gives the cue 0.6–1.3× enrichment, showing no significantly larger amplification in these settings. We report WGA on the original test images without the synthetic cue. To measure sensitivity to the cue, we add it to each test image and report the flip rate: the percentage of images whose predicted class changes after adding the cue. Clean denotes training without the synthetic cue. Each cell below reports WGA ↑ / flip rate ↓, both in percent.

Cue conditionMethod10% selection30% selection50% selection
CleanRandom38.8 / 1.265.1 / 0.667.1 / 0.5
CleanOurs68.5 / 0.277.2 / 0.478.4 / 0.4
1% cueRandom39.4 / 1.065.1 / 0.667.3 / 0.7
1% cueOurs68.7 / 0.376.5 / 0.679.6 / 0.2
2% cueRandom38.3 / 1.165.3 / 0.566.5 / 0.4
2% cueOurs67.2 / 0.376.4 / 0.578.3 / 0.2
5% cueRandom38.5 / 1.463.6 / 0.666.8 / 1.1
5% cueOurs67.2 / 0.576.8 / 1.777.3 / 1.5

Ours outperforms Random in WGA by 10.5–29.3% under all cue conditions and budgets, and its WGA stays within 1.3% of the clean setting. Flip rates stay at or below 0.6% with 1%/2% cues. The WGA drop of random selection is even larger than our method in the 30% selection and 5% cue case. Thus, the method remains effective under a rare predictive shortcut.

In the revised manuscript, we have added the cue construction and evaluation protocol in Appendix E.2, with pool prevalence in Table 28, selected cue exposure in Table 29, and WGA and prediction flip rates in Table 30.

Q1. How is β chosen without labels, and what happens when one fixed value is used across all datasets?

Thank you for the question. Table 17 already evaluates each fixed β\beta across all five datasets. Below we compare β=0.05\beta=0.05 and β=2\beta=2 with the best result for each dataset and budget over the full sweep {0,0.05,0.1,0.25,0.5,1,2}\{0, 0.05, 0.1, 0.25, 0.5, 1, 2\}. Each entry lists the 10% / 30% / 50% results.

Dataset and metricBest over full β\beta sweepFixed β=0.05\beta=0.05Fixed β=2\beta=2
NABirds accuracy42.9 / 67.4 / 75.242.9 / 67.4 / 75.241.9 / 66.0 / 73.3
Stanford Cars accuracy25.3 / 65.8 / 78.823.9 / 65.3 / 78.825.1 / 65.4 / 77.8
Waterbirds WGA69.8 / 77.9 / 79.368.6 / 77.9 / 79.365.4 / 67.6 / 77.3
Spawrious WGA76.1 / 76.9 / 74.974.3 / 74.7 / 72.376.0 / 72.4 / 73.4
Tiny-ImageNet accuracy23.5 / 40.4 / 50.015.4 / 37.0 / 46.523.5 / 40.4 / 50.0

A fixed β=0.05\beta=0.05 is within 2.6% of the best tested setting across the fine-grained and spurious-correlation datasets, while β=2\beta=2 attains the best results on Tiny-ImageNet. As discussed in Section 4.3, these results reflect different demands for concept coverage and local support, where concentrated recognition tasks benefit from broader coverage of under-supported visual patterns, while general classification benefits from a stronger local contribution. We therefore adopt β=0.05\beta=0.05 and β=2\beta=2 as the respective group defaults. When labels or task granularity are not pre-defined, we recommend using β=2\beta=2, which yields more stable performance across data types.

The full β\beta sweep and its interpretation are reported in Appendix D.2, Table 17.

Q3. Results are averaged over three runs, but no variability is reported. Please provide these statistics.

Thank you for requesting these statistics. The reported classification results are means over three runs. We have added the corresponding standard deviations in Appendix G, Tables 32–33 of the revised manuscript.

AuthorsOpenReview ↗

Follow-up Response Part 1 of 2

Reviewer’s follow-up comment

Thanks for author's rebuttal, my concerns on when training and test distributions match have been solved. But my concerns on how SAE extracted rare concepts help with downstream task still remains, simply comparing the proposed method with random sampling does not address this issue. It needs some explanations that are more mechanic.

Response

Thank you for the continued discussion and for confirming that our previous response resolved your concern about matching training and test distributions. We explain below how SAE-extracted rare concepts can help downstream learning. Useful training support includes both category-discriminative attributes and category-agnostic contexts.

1. Contextual support and recognition

Prior work shows that changing image backgrounds can substantially affect recognition even when the foreground object remains unchanged [1]. More directly, Nagarajan et al. study a classification task where image color is independent of the label, models still fail on an underrepresented color, and increasing its training coverage alleviates this failure [2]. Group distributionally robust optimization (group DRO) minimizes the worst-group training loss, including across groups defined by bird category and background in Waterbirds [3]. This highlights that high average accuracy can hide failures under particular contextual conditions. Such conditions can matter for robust recognition even when they do not define the category.

Our insight is that data selection should explicitly consider how training support is allocated across visual concepts, including shared contexts that do not directly distinguish categories. When an enhanced concept represents an under-supported context, selecting more images that carry it gives the model more examples on which to learn recognition under that condition. Concept-support allocation can therefore contribute to the robustness goal of group DRO by improving training coverage of such conditions.

2. A theoretical explanation

Dao et al. [4] analyze training over label-preserving transformations for linear predictors in feature space with a smooth convex loss. Their second-order approximation combines loss at the mean prediction with a prediction-variance penalty. For one example with a scalar prediction score, we restate it as

Ez[ℓ(uz,y)]≈ℓ(μ,y)+12ℓ′′(μ,y)Var⁡z(uz),μ=Ez[uz].\mathbb{E}_{z}[\ell(u_z,y)] \approx \ell(\mu,y) + \frac{1}{2}\ell''(\mu,y)\operatorname{Var}_{z}(u_z), \qquad \mu=\mathbb{E}_{z}[u_z].

Here, z indexes label-preserving transformations, u_z is the prediction score, and y is the unchanged label. The mean score is μ; ℓ is the loss, and ℓ'' is its curvature with respect to the prediction score. Var measures prediction variance across transformations. The variance term penalizes sensitivity to changes that preserve the label.

This motivates our interpretation of contextual support in image selection. When a context is underrepresented, the model receives few training examples under that condition. Increasing its support gives the model more examples on which it must recognize the object despite the contextual variation. Our selection objective directs additional support toward underrepresented concepts, while local support favors recurring evidence. The experiments below examine this connection.

3. Evidence from the main paper

Figure 5(b) reports larger recognition gains on test images associated with enhanced concepts, across NABirds and multimodal pretraining. The hand-context case in Section 5.1 further shows that these gains extend to shared contexts, the accuracy on strongly hand-activated test images increases from 28.6% to 51.4%. Section 5.2 examines the corresponding robustness through background changes and controls that match label–background group counts. The ablation in Figure 7 also shows why the samples providing this support matter. At the 10% NABirds budget, pure-rarity selection achieves 33.1%, concept coverage without local support achieves 37.9%, and the full objective achieves 42.9%.

Reviewer_FmDfOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper uses a sparse autoencoder-based method to implement data selection, and verifies the effectiveness of the method on several benchmarks.

Critical strengths

1.The layout of the article is reasonable.

2.The results are verified on multi-modal datasets.

Critical weaknesses

1.The contribution of the paper is unclear, and the author needs to point out the main contribution of the paper.

2.The comparison method is insufficient. The SAE based data selection methods mentioned in the manuscript [1-2] need to be compared.

3.To understand, the author needs to define the concept of sparsity and explain what it specifically refers to.

4.Innovation is limited, and this method simply weights K-neighbor samples based on SAE. The author needs to conduct ablation experiments on the number of neighboring samples.

[1]Diversity-driven data selection for language model tuning through sparse autoencoder.

[2]Rethinking representativeness and diversity in dynamic data selection.

Questions for authors

1.The contribution of the paper is unclear, and the author needs to point out the main contribution of the paper.

2.The comparison method is insufficient. The SAE based data selection methods mentioned in the manuscript [1-2] need to be compared.

3.To understand, the author needs to define the concept of sparsity and explain what it specifically refers to.

4.Innovation is limited, and this method simply weights K-neighbor samples based on SAE. The author needs to conduct ablation experiments on the number of neighboring samples.

[1]Diversity-driven data selection for language model tuning through sparse autoencoder.

[2]Rethinking representativeness and diversity in dynamic data selection.

3 comments
AuthorsOpenReview ↗

Rebuttal part 1 of 2

Dear reviewer,

We sincerely thank the reviewer for the constructive feedback and for recognizing the organization of our paper and its evaluation on multimodal datasets. Our responses are listed below.

W1 / Q1. The paper’s main contributions are unclear and should be stated explicitly.

Thank you for the question. Our central contribution is to establish concept allocation as a perspective for understanding and designing data selection. We observe that different sample selection strategies can produce similar concept-support distributions (Figure 1 and Table 5), suggesting that selecting different samples does not necessarily change the represented concept distribution. This motivates explicitly considering concept support at the dataset level and investigating its relationship to downstream model behavior. Our contributions are:

  1. We study data selection from the perspective of concept curation, focusing on how a limited training budget is allocated across visual concepts. We leverage patch sparse autoencoder features to characterize recurring visual patterns, including both category-discriminative and category-agnostic concepts.

  2. We propose a simple and efficient selection objective to track accumulated concept support and apply diminishing returns to progressively emphasize under-supported concepts. We further incorporate local support to identify samples that reliably strengthen these concepts, mitigating the risk of being biased by rarity.

  3. We systematically study concept-based selection across three data distribution scenarios: fine-grained, spurious-correlation, and general tasks. Beyond consistent improvements without training on target datasets, our findings reveal a strong connection between concept support and downstream model behavior, with enhanced concept allocation improving both recognition and robustness.

Our findings suggest that effective data selection should go beyond identifying individually valuable or diverse samples to explicitly consider how training support is distributed across visual concepts. Importantly, under-supported concepts can include category-agnostic contextual information that contributes to robustness, while aggressively prioritizing rare concepts without sufficient local support can hurt performance.

We have summarized the three contributions at the end of the Introduction in the revised manuscript.

W2 / Q2. The comparisons are insufficient. Please compare with the cited SAE-based selection methods by Yang et al. and Zhou et al.

Thank you for suggesting these comparisons. We have extended our experiments to include both cited SAE-based approaches. Specifically, Yang et al. [1] use SAE features to measure data diversity. Zhou et al. [2] combine common-factor representativeness and rare-factor diversity through dynamic selection. SDM [3] selects subsets to match a reference concept distribution. Our work does not simply change representation spaces for data selection; instead, it explicitly controls concept allocation, particularly for tail, under-represented concepts.

To enable fair comparison under a fixed data-selection budget, we adapt Yang et al.'s LLM selection method to visual SAE features and Zhou et al.'s dynamic method to a static subset. All methods use the same SAE features and downstream training settings. Each entry below reports the 10% / 30% / 50% results, averaged over three runs.

MethodNABirds accuracy (%)Waterbirds WGA (%)
Yang et al.36.4 / 65.3 / 74.046.4 / 73.3 / 73.7
Zhou et al.37.0 / 63.4 / 71.845.2 / 65.9 / 68.1
SDM37.5 / 64.5 / 73.761.0 / 70.4 / 75.2
Ours42.9 / 67.4 / 75.268.6 / 77.9 / 79.3

Our method achieves the best performance at every evaluated budget on both datasets. At the 10% selection ratio, it improves over the strongest SAE-based baseline by 5.4% on NABirds and 7.6% on Waterbirds.

We have included these comparisons in the revised manuscript: a summary in Section 4.2, Table 3(a), per-budget results in Appendix D.6, Table 22, and adaptation details in Appendix A.2. Due to time constraints, these additional comparisons currently cover NABirds and Waterbirds at 10%, 30%, and 50% selection; we will add the complete results to the main tables in the future.

[1] Xianjun Yang et al. Diversity-driven Data Selection for Language Model Tuning through Sparse Autoencoder. arXiv:2502.14050, 2025.

[2] Yuzhe Zhou et al. Rethinking Representativeness and Diversity in Dynamic Data Selection. arXiv:2603.04981, 2026.

[3] Arnav M. Das et al. Matched Data, Better Models: Target Aligned Data Filtering with Sparse Autoencoders. ICLR, 2026.

AuthorsOpenReview ↗

Rebuttal part 2 of 2

W3 / Q3. Please define sparsity and explain what it refers to in this method.

Thank you for the question. Sparsity refers to the number of nonzero SAE latents (L0L_0) relative to the dictionary size. For each image patch, our TopK SAE retains at most 32 positive activations from a dictionary of 16,384 latents (0.2%). We then apply max pooling across patches to obtain the image-level concept representation used for data selection. Each image activates around 300 latents, which is approximately 2-2.4% of the dictionary on average. Therefore, each image represents a sparse set of the learned concepts.

Full training poolMean patch L0L_0Mean max-pooled image L0L_0Dictionary size
NABirds32.0327.716,384
Waterbirds32.0389.716,384

We have added the sparsity definitions and patch-level and image-level measurements in Appendix B.2, Table 12 of the revised manuscript.

W4 / Q4. The method appears to offer limited novelty beyond weighting neighboring samples with SAE features. Please clarify its novelty and ablate the number of neighbors.

Thank you for raising this point. Our core idea is to move from sample curation to concept curation, explicitly controlling how a limited training budget is allocated across visual concepts. This goes beyond simply applying existing data selection strategies in SAE space.

First, we need to accumulate a distribution of individually measurable visual concepts during selection. SAEs offer this opportunity to characterize concept support across the dataset. As shown in Figure 4, SAE features exhibit distinct activation patterns compared with dense DINO embeddings. They provide more informative signals for identifying frequently supported and underrepresented concepts.

Second, we proactively control the concept support in the dataset. Specifically, we apply diminishing returns to the concept distribution accumulated during selection. Subsequent selection will have a stronger focus on less-represented concepts. The neighborhood component serves a complementary purpose: after identifying underrepresented concepts, it helps to determine which samples can reliably support them. This avoids excessive emphasis on isolated rare samples. Therefore, neighborhood weighting is only one component of our framework, which combines concept-level allocation with sample-level support.

The requested neighborhood-size ablation was already included in the submitted paper and is reported in Appendix D.2, Table 15 of the revised paper. It evaluates k∈{1,2,3,5,10,20}k \in \{1, 2, 3, 5, 10, 20\} on NABirds and Waterbirds at 10%, 30%, and 50% selection rates. The results support k=3k=3 as a strong common default. Section 4.3 and Appendix D also examine component removal, representations, local-support weight, and utility functions.

We have clarified the concept-curation motivation in the Introduction of the revised manuscript.

AuthorsOpenReview ↗

Follow-up on Our Responses

Dear Reviewer,

Thank you again for your time and constructive feedback. In our responses and revised manuscript, we have clarified the contributions and concept-curation motivation, added comparisons with the requested SAE-based methods, defined sparsity, and pointed to the neighborhood-size ablation.

As the discussion period is nearing its close, we would be grateful to hear whether these updates address your concerns, particularly regarding our contribution and its distinction from prior work. Please let us know if any points remain unclear; we would be happy to provide further clarification.

Comments on the paper

2 comments

Dear Reviewers,

The authors have provided their rebuttals to the reviewers' comments. Please take a moment to review the responses and continue the discussion on any remaining concerns, unresolved issues, or points that may require further clarification.

Your further input and discussion are highly appreciated.

Regards,
AC for this paper

AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers (R1: FmDf, R2: YqQT, R3: tz3A) for their constructive feedback. During the rebuttal period, we have clarified our contributions and added further experiments to address the reviewers’ concerns. We summarize the key updates below:

1. What distinguishes our contribution from existing data-selection approaches?

(R1 W1/Q1, W4/Q4) We clarified that we study data selection from the perspective of concept curation, focusing on how a limited training budget is allocated across visual concepts. Our findings motivate a simple and efficient objective combining accumulated concept support with local support, supported by systematic evaluation across different learning settings. (Clarified contributions and motivation at the end of the Introduction)

2. Do the gains persist against other SAE-based methods and baselines using stronger features? (R1 W2/Q2, R3 Q1) We extended our comparisons to the cited SAE-based approaches and DINOv3-based baseline variants. Our method outperforms these additional baselines across the evaluated settings, supporting the value of concept-support allocation beyond the feature extractor. We also documented the scoring models and feature configurations used by each baseline. (New results in Section 4.2, Table 3, and Appendices D.6–D.7, Tables 22–23; configurations in Appendix A.2, Table 7)

3. Do the benefits of concept allocation depend on larger training–test distribution differences? (R2 W1) We added distribution diagnostics showing that the benefits of adjusting concept support do not simply correlate with a larger train–test distribution gap. We refined the Introduction’s distribution-preservation statement and discussed how the ImageNet recognition and robustness results relate to useful training support. (Revised Introduction; new diagnostics in Appendix E.1, Table 26)

4. Why are SAE concepts useful for downstream learning? (R1 W3/Q3, R2 W2, R3 W) We clarified that useful training support includes both category-discriminative attributes and category-agnostic contexts, supported by the concept-group analysis and hand-context case. We also explained how sparse activations interact with the selection objective and clarified the meaning of sparsity. (Added representation–objective explanation in Appendix D.5; sparsity definitions and measurements in Appendix B.2, Table 12)

5. Does the method remain effective when a shortcut is rare but perfectly predictive? (R2 W3/Q2) We conducted additional shortcut experiments examining selected cue exposure, downstream robustness, and prediction sensitivity. The results show that our method retains its robustness advantage without substantially amplifying the rare cue in the tested settings. (New results in Appendix E.2, Tables 28–30)

6. Why use within-class selection, and are the gains explained by class balancing? (R3 Q2) We clarified that shared contexts need training support in different classes. Controlled comparisons isolate the benefit of separate support accumulation from class balancing, and the hand-context analysis shows broader coverage across classes. We also documented the class-specific implementation and the global-selection setting. (Clarifications in Section 3.2 and Appendix A.2; new comparisons in Appendix D.3, Tables 18–19)

7. How sensitive are the results to parameter choices and training runs? (R1 W4/Q4, R2 Q1/Q3) We discussed the existing neighborhood-size and fixed-β comparisons and clarified the parameter defaults. We also added means and standard deviations across three training runs for the classification results. (Existing parameter comparisons in Appendix D.2, Tables 15 and 17; added run statistics in Appendix G, Tables 32–33)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.