acceptodds
Under review as a conference paper at ICLR 2027

Conceptdiff: Contrastive Concept Discovery for Dataset Comparison

Abstract

Identifying how two datasets differ in their semantics is a basic primitive for model understanding, data analysis, and scientific discovery. It is also hard to find: differences can be subtle, their form unknown in advance, and there are usually several. Current methods require foundation models to inspect the datasets, and propose natural-language hypotheses regarding the difference. This makes them expensive, and it limits them to differences the model thinks to propose. In this paper, we argue that differences should instead be learned directly from data. We formalize this as contrastive concept discovery (CCD), the task of finding a small set of dataset difference concepts that are discriminative (each separates the two datasets), monosemantic (each encodes a single factor), and diverse (together they cover distinct differences). We then introduce ConceptDiff, a lightweight, domain-agnostic method based on training a regularized concept-based classifier designed to learn diverse concepts. Across text, image, and video use-cases, we find that ConceptDiff recovers ground truth dataset differences better than LLM-based, clustering, and sparse-autoencoder baselines, and consistently finds concepts that more discriminative, diverse and monosemantic.

Primary area:
interpretability and explainable AI
Keywords:
concept-based methods, interpretability, explainability, error discovery

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_bYVqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces ConceptDiff, a method for learning a small set of concept detectors that explain how a target dataset A differs from a reference dataset B. Built on a frozen pretrained encoder, the method aggregates local token evidence using learnable thresholds and trains the detectors with a target-oriented multiple-instance learning (MIL) classification objective, together with diversity, balance, and commitment regularization to encourage specialization. The authors evaluate concept recovery, discriminability, coverage, and interpretability across image, text, and video settings, as well as error slice discovery.

Critical strengths
  1. A practically relevant problem with a comprehensive evaluation framework Explaining dataset differences through multiple interpretable concepts has practical value for distribution shift analysis, model error diagnosis, and the analysis of generated content. The paper considers discriminability, concept coverage, semantic coherence, and stability together. This evaluation framework can expose failures such as accurate classification accompanied by concept collapse, making it more informative for assessing explanations.
  2. A coherent methodological design that provides localized evidence Local token selection, target-oriented MIL aggregation, and concept specialization regularizers are well aligned with the objective of jointly explaining how A differs from B using a small number of concepts. The learned detectors can score new examples and localize the evidence supporting each concept, facilitating inspection and interpretation of the results.
  3. Substantial improvements on selected key benchmarks On the constructed image tasks, ConceptDiff achieves an ARI of 0.460, compared with 0.275 for the strongest reported baseline. On the evaluated subset of the error slice discovery benchmark, it achieves an AP of 0.807, compared with 0.742 for the strongest baseline. These results provide empirical support for improved recovery of semantic groups, while the text and video examples demonstrate potential applicability across modalities. These findings are reported in Tables 1 and 4 of the manuscript.
Critical weaknesses
  1. The novelty relative to the closest prior work is not sufficiently delineated ProtoSumm [1] already uses interpretable classifiers to learn dataset differences and addresses specialization and collapse among multiple prototypes. NLParam [2] already optimizes multiple interpretable predicates under a task objective. The main methodological contribution of ConceptDiff therefore appears to lie in its particular combination of architecture and regularization. However, the manuscript does not sufficiently establish what additional capabilities this combination provides beyond the closest prior approaches, and it neither cites nor evaluates NLParam.
  2. Substantial baseline adaptations may affect the comparative conclusions The manuscript replaces the joint sparse feature selection in the original HypotheSAEs [3] with univariate AUROC ranking. It also replaces the semantic verifier in Zhong et al. [4]—whose adapted baseline is referred to as TextDiff in the manuscript—with text embedding similarity. These changes may affect concept complementarity and semantic scoring, making it difficult to attribute the observed performance gaps entirely to the proposed method. Their effects need to be quantified. The available evidence does not establish that the adaptations necessarily weaken the baselines.
  3. The independent contributions of the core architectural components are insufficiently validated Although the paper includes ablations of the regularization terms, it does not separately evaluate the contributions of the asymmetric concept architecture, MIL aggregation, and learnable threshold pooling. Moreover, some ablated configurations in Figure 2 obtain slightly higher ARI values than the full model. The current results therefore suggest tradeoffs among objectives, without establishing that every proposed component makes an independent and necessary contribution.
  4. Reliable concept recovery under more complex difference structures remains insufficiently established The paper includes real-world case studies and explorations with different values of K, but its strongest evidence for semantic recovery still comes from constructed tasks with known concept counts. Systematic validation remains limited when the true number of concepts is unknown, factors overlap, or concept frequencies are substantially imbalanced. In particular, whether balance and commitment regularization can cause undesirable splitting or merging of concepts under these conditions has not been sufficiently examined.

References [1] ProtoSumm Babbar, V., Guo, Z., & Rudin, C. (2025). “What is Different Between These Datasets?” A Framework for Explaining Data Distribution Shifts. Journal of Machine Learning Research, 26(180), 1–64. [2] NLParam Zhong, R., Wang, H., Klein, D., & Steinhardt, J. (2024). Explaining Datasets in Words: Statistical Models with Natural Language Parameters. Advances in Neural Information Processing Systems, 37 (NeurIPS 2024). [3] HypotheSAEs Movva, R., Peng, K., Garg, N., Kleinberg, J., & Pierson, E. (2025). Sparse Autoencoders for Hypothesis Generation. In Proceedings of the 42nd International Conference on Machine Learning, Proceedings of Machine Learning Research, 267, 44997–45023. [4] TextDiff: Original work underlying the adapted baseline Zhong, R., Snell, C., Klein, D., & Steinhardt, J. (2022). Describing Differences between Text Distributions with Natural Language. In Proceedings of the 39th International Conference on Machine Learning, Proceedings of Machine Learning Research, 162, 27099–27116.

Questions for authors
  1. Novelty: What specific limitation of ProtoSumm and NLParam does ConceptDiff’s main methodological innovation address?
  2. Baseline adaptations: Does ConceptDiff retain its main performance advantages when the baselines preserve their original core mechanisms and are evaluated under matched tuning and evaluation budgets?
  3. Component contributions: How much does concept recovery performance change when the core architectural components are replaced one at a time, while keeping the encoder, regularization, and training budget fixed?
  4. Complex difference structures: Can the method reliably recover the main differences when the true number of concepts is not provided, concepts overlap, and their frequencies are substantially imbalanced?
4 comments

Thank you for the insightful review and for recognizing the practical value of our problem formulation, our multi-metric evaluation framework, and our coherent design for localizing token evidence. We appreciate your constructive suggestions regarding comparisons to closest prior work (ProtoSumm and NLParam), baseline adaptations, architectural component ablations, and robustness under imbalanced or overlapping concepts. We have answered all the questions and supplemented this rebuttal with new experiments and ablations below to address your concerns.

W1 & Q1: Delineating Novelty Relative to ProtoSumm and NLParam

Comparison with ProtoSumm. While we include ProtoSumm as a baseline across Tables 1, 3, and 4, there are two main differences in how ConceptDiff works. First, ProtoSumm is symmetric and tries to learn prototypes for both dataset A and dataset B. When dataset B is just normal background data (such as normal driving or error-free predictions), modeling B is unnecessary and hurts performance on finding what makes A different. Second, ProtoSumm forces each concept to match a single training patch. This works for simple local textures, but struggles when a concept is more abstract or spans across multiple tokens. In contrast, ConceptDiff only models what triggers dataset A (using an asymmetric MIL objective) and uses learnable thresholds to combine multiple relevant tokens per input (Eqs. 1 and 2). As shown in our results, this improves ground-truth group recovery (ARI) from 0.139 for ProtoSumm to 0.460 for ConceptDiff in Table 1, improves error slice discovery AP from 0.742 to 0.807 in Table 4, and avoids the concept collapse ProtoSumm has on crash videos (Act = 0.22 vs. our 1.00 in Table 3).

Comparison with NLParam. Thank you for pointing out NLParam [1]. It is an interesting work that extends the paradigm of TextDiff [2] (which we already evaluate in Tables 2 and 5) that fits a statistical model over natural-language predicates using an LLM. However, since NLParam is very similar in setup to TextDiff, it shares the same trade-offs for our task. It relies on querying an LLM proposer and verifier across samples during optimization, which makes it expensive to scale to large image or video datasets and cannot directly localize differences to specific visual patches or video tokens. In contrast, ConceptDiff learns continuous concept detectors directly from tokens in a few minutes without an LLM in the loop, making it complementary to text-predicate methods like NLParam. We will cite and discuss NLParam alongside TextDiff in the revised paper.

We hope this clarifies how ConceptDiff differs from both ProtoSumm and NLParam, and we will include this discussion in the revised paper.

W2 & Q2: Original vs. Adapted Mechanisms in HypotheSAEs and TextDiff

We would like to clarify that the adaptations made in the paper were primarily so we could easily apply the baselines across different modalities, and were never meant to affect the comparative conclusions. In fact, some of our changes were meant to be an enhancement to the baseline, such as upgrading TextDiff to use a stronger, more modern language model and embedding scorer.

That said, to answer the reviewer's query, we ran the experiments with the original baseline mechanisms restored (using the original L1 sparse feature selection for HypotheSAEs and the released semantic verifier for TextDiff). As shown in the tables below for Images (Table 1 benchmark) and Text (EditLens Gemini-generated vs. human writing, Table 2), using the original mechanisms leads to mixed trade-offs across metrics, but ConceptDiff still maintains a clear advantage overall.

Table 1: Image Benchmark (12 tasks, K = 3)

MethodARI ↑cEER ↓Act ↑Cov ↑ΔCoh ↑Stab ↑
HypotheSAEs (pre.) (Paper)0.11720.15511.00000.53350.1176$^{\mathcal{R}}$0.1573
HypotheSAEs (cls.) (Paper)0.13100.09920.86110.75110.06510.2130
HypotheSAEs (pre.) (Original)0.14090.33801.00000.31110.0877$^{\mathcal{R}}$0.2931
HypotheSAEs (cls.) (Original)0.09410.39100.91670.18590.03970.2757
ConceptDiff (ours)0.45960.06171.00000.82720.1319$^{\mathcal{R}}$0.5221
R^{\mathcal{R}} indicates monosemanticity calculated based on the concept region.

Table 2: Text Benchmark (EditLens: Gemini-Generated vs. Human Writing, K = 4)

MethodcEER ↓Act ↑Cov ↑ΔCoh ↑Stab ↑
HypotheSAEs (pre.) (Paper)0.09621.00000.49120.27680.0886
HypotheSAEs (cls.) (Paper)0.00380.41670.99860.12330.0040
TextDiff (Paper)0.25321.00000.27910.25270.1502
HypotheSAEs (pre.) (Original)0.13290.66670.20930.2478-0.0230
HypotheSAEs (cls.) (Original)0.00030.33330.65110.09280.3333
TextDiff (Original)0.30430.75000.79740.0393-0.0159
ConceptDiff (ours)0.00211.00001.00000.22650.9267

We hope these results confirm that ConceptDiff's improvements hold regardless of whether the adapted or original baseline mechanisms are used, and we will include this comparison in the revised paper.

W3 & Q3: Ablations of Core Architectural Components (Asymmetric Head, MIL Pooling, Threshold Routing)

Following the reviewer's suggestion, we ran an architectural ablation study across all 12 tasks on the image benchmark (K = 3, 3 seeds) where we replace each of the core architectural components one at a time while keeping the frozen DINOv2-L encoder, regularization weights, and training budget fixed.

First, we tested a symmetric target and reference head that replaces our asymmetric target-only head and learned detectors for both dataset A and reference dataset B. Second, we tested a linear classifier over concepts that replaces our MIL smooth-max aggregation with a standard linear classifier over the concept scores. Third, we tested mean pooling over tokens that replaces our thresholded token routing with average pooling across all tokens. And as shown in the table below, replacing any of the three components hurts group recovery (ARI), separation (cEER), coverage (Cov), or stability (Stab).

Architecture VariantARI ↑cEER ↓Act ↑Cov ↑ΔCoh ↑Stab ↑
Symmetric target/reference head0.41000.11321.00000.56720.12370.4317
Linear classifier over concepts (w/o MIL)0.42410.08161.00000.67640.10560.4358
Mean pooling over tokens (w/o threshold routing)0.38230.06651.00000.78890.10500.4423
ConceptDiff (Full Model)0.45960.06171.00000.82720.13190.5221

Additionally, we believe there is a slight misinterpretation regarding results shown in Figure 2 of the paper. While removing both the commitment and diversity regularizers do increase ARI slightly from 0.460 to 0.472, the other evaluation aspects degrade, as separation worsens (cEER increases from 0.062 to 0.071), target coverage drops (from 0.823 to 0.808), and stability across seeds drops significantly (from 0.522 to 0.471). We want to clarify that, contrastive concept discovery is a multi-objective task where discovered differences must be discriminative, diverse, and monosemantic at the same time. Thus, looking at all metrics together is essential when evaluating a method in this task, and our full model provides the strongest overall balance across every criterion rather than over-optimizing ARI alone.

These results show that each architectural component and regularizer in ConceptDiff makes a necessary contribution to the final model. We hope this clarifies the reviewer's concern, and we will include this architectural ablation table and discussion in the final revised paper.

W4 & Q4: Discovering Dataset Differences In Real-world Scenario.

We agree with the reviewer that discovering differences in real-world scenarios is challenging, especially when the true number of concepts is unknown, factors overlap, and class frequencies can be imbalanced. This is precisely why we included our real-world case studies on EditLens text and dashcam car crash videos, where no ground-truth concept counts or balanced distributions exist.

Firstly, on how ConceptDiff handles unknown KK, we view KK as a human-defined budget that controls the resolution of the discovered differences. As we increase KK, the model naturally decomposes broad dataset differences into finer, more specific sub-concepts rather than collapsing or failing to converge, which is clearly visible in our text experiments where different KK values yield coherent sub-categories.

Additionally, as we noted in our response to Reviewer ANji, we show that stability across random seeds and cross-validation splits provides a principled way to guide the choice of KK in practice. When KK exceeds the true structural complexity of the underlying data distribution, the discovered concepts begin to split arbitrarily, leading to a sharp drop in stability across seeds. Practitioners can thus monitor stability scores across increasing values of KK to identify the point before instability sets in, serving as a reliable indicator for the ideal concept budget.

Thus, we believe ConceptDiff provides a robust and reliable way to uncover structure in complex real-world environments. We thank the reviewer for bringing up this important point, and we will expand our discussion in the revised paper.

References

[1] NLParam Zhong, R., Wang, H., Klein, D., & Steinhardt, J. (2024). Explaining Datasets in Words: Statistical Models with Natural Language Parameters. Advances in Neural Information Processing Systems, 37 (NeurIPS 2024).

[2] TextDiff: Original work underlying the adapted baseline Zhong, R., Snell, C., Klein, D., & Steinhardt, J. (2022). Describing Differences between Text Distributions with Natural Language. In Proceedings of the 39th International Conference on Machine Learning, Proceedings of Machine Learning Research, 162, 27099–27116.

Thank you for the additional experiments and clarifications. Restoring the original baseline mechanisms and providing controlled architectural ablations substantially alleviate my concerns about baseline adaptations and component contributions. I consider the architectural ablation concern largely addressed.

Presenting examples in descending activation order may provide cues that inflate simulation correlations, and no randomized-order results are provided to exclude this possibility. This leaves an important concern about the evidence supporting monosemanticity.

The discussion of NLParam also remains incomplete: its task-driven optimization of multiple interpretable predicates is directly relevant to the novelty claim. The architectural differences are now clearer, but the contribution should be more carefully delineated against this prior work.

The rebuttal strengthens the empirical case, and I have updated my assessment positively in that respect. Nevertheless, the remaining concerns about novelty prevent me from moving to acceptance. I therefore retain my weak-reject rating.

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

The paper presents a new task for determining a small set of concept difference between two datasets. This small set are expected to include samples that are discriminative, monosemantic, and diverse. The paper then proposes ConceptDiff, a method for contrastive concept discovery. ConceptDiff learns domain-agnostic regularized concept classifer with diverse concepts. The paper demonstrates the effectiveness of the proposed method in discovering the dataset differences, outperforming the existing baselines.

Critical strengths

The proposed task for contrastive concept discovery seems new, and the method is technically interesting and sensible.

Critical weaknesses

ConceptDiff architecture uses an encoder to produce tokens, which are then projected to feature vectors for scoring each concept. This architecture is rather common, and the technical contribution seems quite limited.

The performance of ConceptDiff seems mixed.

  • For image datasets, ConceptDiff achieves good results, but it is only experimented with three annotated groups. The number of concepts are limited to only 3 as well. It is not clear whether the experiment can be scaled with more groups and concepts.

  • For EditLens, the improvement from ConceptDiff seems not significant enough, as the baseline kmeans works quite well (Table 2). The same case happens for car crash videos (Table 3).

  • It remains unclear why stability varies significantly across Table 1, 2, 3.

The experiments are done on a relative small scale (with a few thousands of images as indicated in Fig. 8 in appendix). For computer vision datasets, the scale could be much larger, from tens of thousands toward millions and even more. How well does CCD and ConceptDiff scale to discover differences in large datasets?

Questions for authors
  1. Please clarify the technical novelty of the paper.

  2. Please clarify the significance of the performance gap in Table 1, 2, 3 compared to the baselines.

  3. Please comment on the scale of CCD and ConceptDiff. Why only a few concepts can be discovered? How well can the difference discovery work for large image or video datasets?

2 comments

We are thankful to the reviewer for the feedback. We appreciate the reviewer's recognition of our Contrastive Concept Discovery (CCD) task as new and our method as technically interesting and sensible. However, we believe there are some key misunderstandings regarding the technical contributions and experimental results of the proposed method. We have answered all the questions and provided new large-scale experiments below to clarify these queries for the reviewer. In the text below, Q1 & W1, refer to Question #1 and Weakness #1 respectively.

W1 & Q1. Is the technical contribution limited to an encoder and token projection?

We would like to clarify a potential misunderstanding that using a frozen encoder and projecting tokens (Stages 1 and 2) is our main contribution. The use of encoder to extract features is to simplify the input interface across modalities, and our baselines (like k-means, HypotheSAEs and ProtoSumm) use this exact same frozen token setup. The core technical contribution of ConceptDiff lies in how we route, pool, and regularize those tokens to identify differences (Stages 3 and 4, Section 3.2.2):

  1. Asymmetric Multiple-Instance Learning (MIL) formulation (Stage 4, Eq. 2): Standard concept-based classifiers (such as ProtoSumm) are symmetric and learn features for both datasets. However, when comparing a target dataset A (e.g., crash videos) against a reference dataset B (normal driving), finding concepts for B is uninformative. We instead formulate dataset comparison as an asymmetric Multiple-Instance Learning (MIL) problem using a smooth-max pooling layer (Eq. 2), where an input is predicted as A if and only if at least one concept fires.
  2. Adaptive token selection to filter out background context (Stage 3): In images, text, or videos, most tokens represent shared background context rather than what separates A from B, so global pooling can wash out subtle differences. To solve this, each concept learns its own threshold to softly claim and pool only the tokens that clear that threshold (Eq. 1). This allows ConceptDiff to isolate localized differences (such as specific image regions or text spans) without any spatial or span supervision.
  3. Joint anti-collapse regularization (Section 3.2.2): Because a MIL classifier can trivially satisfy the A-vs-B objective using a single head, we introduce three complementary regularizers over the target assignment probabilities (Eq. 3). Diversity to decorrelate concept scores across A, balance to ensure all K heads are utilized across batches, and commitment to force each target example to be explained by a single concept rather than an uninterpretable mixture. Together, these losses prevent collapse and raise group recovery (ARI) from -0.0001 to 0.460 (as shown in Figure 2 of paper).

We believe that formulating Contrastive Concept Discovery as an asymmetric problem and designing a domain-agnostic algorithm that discovers differences across vision, text, and video is a novel technical contribution. We will highlight this distinction more clearly in Section 3.2 of the camera-ready version, and we hope the reviewer will consider these contributions in their evaluation.

W2 (Points 2 & 3) & Q2. Significance of the Performance Gap and Stability Across Tables 1, 2, and 3.

We would like to clarify and explain that ConceptDiff does improve over baselines like k-means. As defined in our formulation (Section 3.1), to solve contrastive concept discovery, the discovered concepts need to be discriminative, diverse, and monosemantic. Thus, we urge the reviewer to look at these values together. Just looking at one aspect, such as cEER or stability alone, might make baselines like k-means appear to be working well, when in other aspects they are not. For example, frozen k-means (pre.) has high stability, but it misses the actual differences between datasets and gets very low coverage (0.395 vs. our 0.827 in Table 1, and 0.439 vs. our 1.000 in Table 3). Similarly, classifier k-means (cls.) gets a low cEER because its MLP is trained to classify A vs. B, but it fails on monosemanticity and stability (on EditLens, its simulation score drops to 0.456 vs. our 0.670, and stability drops to 0.257 vs. our 0.927). In contrast, ConceptDiff is the most balanced overall across all three aspects, and these gains are consistent across seeds, as seen through our stability scores. We hope this clarifies the reviewer's misunderstanding on how ConceptDiff improves over the baselines.

Why does stability vary?

We believe this is due to how cleanly the frozen backbone represents data in each modality. Since ConceptDiff relies on tokens as input, noisy tokens can indeed lead to less stable results. This is evident from our text and video examples. Text tokens from T5-large are much cleaner and thus produce very stable groups (0.927 in Table 2), while raw dashcam video tokens from V-JEPA 2.1 contain heavy background clutter, leading to lower stability (0.383 in Table 3).

Having said that, we want to clearly state that having a good backbone does not mean it will find the differences on its own. If it did, k-means (pre.) and HypotheSAEs (pre.) would easily recover the groups in Table 1 since they use the exact same frozen features, yet they only reach 0.189 and 0.117 ARI vs. our 0.460. Having a strong backbone simply helps in getting more stable differences. We will clarify this in the paper.

W2 (Point 1), W3, & Q3. Can ConceptDiff scale to more concepts (K) and larger datasets?

Yes, ConceptDiff is highly scalable in both the number of concepts K and dataset size. In Table 1, we intentionally started with K = 3 and a controlled dataset size to establish a clean ground-truth benchmark where we know the exact differences ahead of time and can measure group recovery (ARI). In general settings, K is simply a user-defined budget, and we already show in the paper that ConceptDiff works well with K = 5 on DCBench (Table 4), K = 6 on crash videos and StoryScope (Tables 3 and 6), and up to K = 12 on EditLens (Figure 6 and Appendix A.5.3), where increasing K splits broad differences into finer sub-concepts without collapsing.

For dataset size, because ConceptDiff only trains a lightweight head on top of cached tokens, training scales linearly with the number of samples. On a single NVIDIA RTX 4090 GPU, training takes just 0.42±0.150.42 \pm 0.15 seconds per epoch for 1,0001,000 images and 3.60±1.043.60 \pm 1.04 seconds per epoch for 10,00010,000 images. This makes it much faster and easier to scale to large datasets compared to LLM-in-the-loop baselines (such as VisDiff) that have to query a VLM across individual images.

Following the reviewer's suggestion, we constructed a new large-scale benchmark with 20,603 images (scaled from 1.5k in original image benchmark) and K = 9 planted concept classes spanning three scenarios: (a) Neckwear: Bow tie, bolo tie, Windsor tie, (b) Headwear: Cowboy hat, sombrero, mortarboard (graduation cap), and (c) Emergency vehicles: Ambulance, fire engine, police van.

Even at this larger scale (20,603 images) and higher concept budget (K=9K = 9), ConceptDiff keeps all 9 heads active (Act = 1.000), reaches 0.9808 coverage, and substantially outperforms kk-means in both ground-truth group recovery (0.6443 vs. 0.5005 ARI) and separation (0.0113 vs. 0.0402 cEER):

MethodKARI ↑cEER ↓Act ↑Cov ↑Stab ↑
k-means90.50050.04021.0000.78470.9753
ConceptDiff (ours)90.64430.01131.0000.98080.6769

These results show that ConceptDiff is not limited to small KK or small datasets and can scale in both of these axes without any issue.

We thank the reviewer again for their constructive feedback, which helped us clarify our core algorithmic contributions and demonstrate the scalability of ConceptDiff. We will include these findings in our final camera-ready version. Since we have addressed all of your questions and validated large-scale performance through new experiments, we hope you will consider raising your score in light of our review.

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

The paper defines contrastive concept discovery: given a target set A and a reference set B, find K scalar concepts that separate A from B, cover different differences, and each track one factor. ConceptDiff freezes a pretrained encoder, maps each token through a small MLP, and scores it with K linear detectors. Each concept softly weights tokens with a learned threshold, pools them into one score, and a smooth maximum over those scores classifies A versus B. Diversity, balance, and commitment losses are there to stop one concept from taking the whole decision. The concepts are the product. The classifier is only a training device.

The controlled benchmark is 12 constructed image pairs with K=3 (seven CelebA-HQ attribute splits, five ImageNet-family splits). The same head is then run on EditLens text (T5-large, K=4), crash versus normal driving (V-JEPA 2.1, K=6), and error slices on a filtered DCBench CelebA subset (K=5).

Critical strengths
  • The task is stated cleanly, and the asymmetry with ordinary concept-bottleneck models is real. The method looks for concepts present in A and absent from B. The multiple-instance head matches that goal: A is predicted when some concept fires.
  • Table 1 is the strongest result in the paper. On the planted differences, ConceptDiff reaches ARI 0.460, cEER 0.062, and coverage 0.827. The best k-means variant, k-means on classifier features, is at ARI 0.275 and coverage 0.747. The best baseline cEER is HypotheSAEs (cls.) at 0.099. This is a recovery and separation gain. k-means (pre.), HypotheSAEs (pre.), VisDiff, and ProtoSumm already keep every head active (Act 1.00), so head usage is not what Table 1 adds. Figure 2 shows why the regularizers are in the method: classification alone collapses to one active head (33% of K=3) with ARI 0.0001, and the paper reports that removing balance leaves most heads inactive.
  • On the filtered DCBench subset, slice AP is 0.807, against 0.742 for k-means (pre.) and 0.718 for Domino (Table 4). ConceptDiff cEER is 0.169, in line with Domino at 0.162.
  • The same objective runs on DINO tokens, T5 tokens, and V-JEPA tokens. On EditLens the discovered concepts localize to spans (Figure 5), and raising K from 4 to 12 splits broad writing styles into finer ones (Figure 6). That qualitative behavior is useful. The quantitative edge there is stability and the simulation score, which Tables 2 and 3 put next to the separation numbers.
Critical weaknesses
  • The use case that opens and closes the paper is not solved. The introduction asks whether safety-critical clips are fast pedestrians, bad weather, or rare events. Section 4.4, with K=6 on V-JEPA 2.1 features, returns night driving and a motorcycle rider's viewpoint (Figure 7). The authors say these groups do not isolate collision mechanisms. The conclusion still says ConceptDiff addressed the driving-log problem. A weak video encoder is a plausible cause, and it also means the empirical case for the problem the introduction says the paper exists to solve is still missing. Only two of the six concepts are shown.
  • Monosemanticity is not in the training loss. The limitations section says perceptual regularizers were tried and failed, so diversity and commitment are the only pressure toward a single factor. The simulation score then asks an explainer model to name a concept from its top examples and a simulator model to predict scores from that name (Section 3.1, Appendix A.3). Coherence is a separate embedding statistic (SigLIP2, X-CLIP, or MPNet). The introduction criticizes foundation models for proposing only differences they already think of. The main interpretability number depends on the same kind of model. Neither score shows that a head is a single human factor.
  • Separation and discovery are easy to conflate. cEER is the mean only over heads with at least five assigned target examples (Appendix A.3), so a collapsed solution can still post a low cEER. On EditLens, classifier k-means has cEER 0.0002 versus ConceptDiff at 0.0021, and both keep all four heads active. ConceptDiff is more stable (0.927 vs 0.257) and has a higher simulation score (0.670 vs 0.456). On crash video, HypotheSAEs (cls.) and ProtoSumm sit at cEER 0 with Act 0.22, about one or two of six heads. ConceptDiff also has cEER 0, keeps all six heads active, and reports coherence 0.151 against at most 0.141. Frozen k-means is far more stable: 0.981 versus 0.383 on video, and 0.990 versus 0.522 on the image benchmark. The abstract's claim of concepts that are consistently more discriminative, diverse, and monosemantic is stronger than Tables 2 and 3.
  • The planted CelebA concepts the paper shows (eyeglasses, necktie, hat; Figures 1 and 3) are attributes a current vision encoder already represents. The other CelebA splits are facial attributes of the same kind. Figure 4 reports that a stronger frozen backbone improves cEER and regional coherence, with DINOv3-L best. ConceptDiff learns the split from the contrast, on top of a frozen foundation-model encoder in every modality. That is a weaker claim than "learned directly from data, without a foundation model in the loop."
  • The DCBench number is on a selected slice. Of 520 CelebA problems, 142 are kept by requiring that at least 70% of the annotated slice is misclassified. Table 4's caption lists 143 slices on those 142 problems. The filter matches the method's definition of an error slice. The paper reports that subset only. The following sentence, which says the omitted problems are those with under 50% of the slice correctly classified, describes the other tail of the same statistic.
  • K is an input. The main experiments set it to 3, 4, 5, or 6, and the paper gives no selection rule for how many differences exist. The K sweep (Appendix A.5.3) is a qualitative text study. Image cross-seed stability of 0.522 means two runs often disagree on the grouping. A baseline that takes the top linear directions between A and B is also missing. Contrastive PCA is only cited.
Questions for authors
  1. Please show all six crash concepts. Figure 7 shows night driving and a motorcycle viewpoint, and Section 4.4 says these are not collision mechanisms. If the full set is scene type and camera viewpoint, that does not answer the question in the introduction. If the encoder is replaced by one that represents motion and interaction, does any concept correspond to a cause of the crash?
  2. What is slice AP, and cEER, on all 520 DCBench CelebA problems, without the filter that keeps a problem only when at least 70% of the annotated slice is misclassified?
  3. How should a user pick K when the number of differences is unknown, and what does the method return when K is larger than the number of real differences?
  4. On a sample of concepts, if a person writes the description instead of the explainer model, does the simulation score still hold? The paper does not name the explainer or simulator.
4 comments

We thank the reviewer for their thoughtful evaluation and for highlighting our clean task formulation, the asymmetric multiple-instance design, and our experiment results. We appreciate your constructive feedback regarding the driving concepts, evaluation metrics, DCBench filtering, and concept budget KK. We supplement this rebuttal with additional experiments to answer the questions.

Please refer to this Rebuttal Code where we have compiled detailed explanations and visualizations for better understanding. In the text below, Q1 & W1, refer to Question #1 and Weakness #1 respectively.

Q1 & W1: Additional Autonomous Driving Crash Concepts

We have uploaded examples for all six discovered concepts at the anonymized link. As shown in the legend of Figure 7 (left), the six concepts on the Car Crash Dataset are (1) Sunny, dry roads, (2) Motorcycle rider view, (3) Highway overtaking, (4) Snowy roads, (5) Lane intrusion, and (6) Night driving. While two of these capture weather and lighting conditions (Snowy roads, Night driving) and one reflects camera perspective (Motorcycle rider view), Concepts 3 and 5 group clips with dynamic traffic interactions (Highway overtaking and Lane intrusion). We believe these discovered differences address the main motivation from the Introduction of finding crash concepts that differ from normal driving.

Our comment in Section 4.4 that these groups do not isolate distinct collision mechanisms was about the failure of the base V-JEPA 2.1 model which is known to struggle with fine-grained motion and intuitive physics (https://arxiv.org/abs/2512.06232), and not a failure of the method itself. We apologize if our phrasing gave the impression that the method failed on this task, and we will make this clearer in the final draft.

Following the reviewer's suggestion, we also tested our method with two new video backbones (matched at K=6K=6) to see if other models extract finer-grained crash concepts. As shown in the table below, across all tested backbones, V-JEPA 2.1 still remains the best choice among them all. We believe with better and newer backbones these results will improve further.

Video Backbone (K=6)cEER ↓Act ↑Cov ↑ΔCoh ↑Stab ↑
EmbeddingGemma 20.0003 ± 0.00021.001.000.0617 ± 0.01910.181
VideoMAE V2-Giant0.00001.001.000.0859 ± 0.00570.548
V-JEPA 2.1 (Used in Paper)0.0000 ± 0.00011.001.000.1505 ± 0.00800.383

Q2 & W5: Unfiltered DCBench Evaluation

We apologize for the confusing wording on Line 456. That sentence should have stated that in the majority of omitted DCBench problems, less than 50% of the annotated ground-truth slice was misclassified (meaning the classifier actually got most of the slice right). Since our setup compares a model's errors (AA) against its correct predictions (BB), a slice that the model mostly gets right sits in the reference set BB rather than the error set AA. We used the 70% misclassification filter in Table 4 simply to ensure that the ground-truth slices we evaluate on are genuine error modes concentrated in AA. We will fix this in the final version.

To answer the reviewer's question on the unfiltered benchmark, we evaluated all 520 DCBench CelebA problems (K=5K=5, 3 seeds) without the 70% filter and presented the results below. Specifically, we see that even in the unfiltered setting, ConceptDiff finds better slices (0.650 Slice AP) with more discriminative features (0.208 cEER) compared to Domino (0.537 Slice AP, 0.249 cEER). While frozen kk-means gets a slightly higher unfiltered Slice AP (0.662), it clusters raw image features without separating AA from BB, failing to find discriminative features (0.269 cEER). We will include these findings in the final version.

MethodFiltered (142 problems) Slice AP ↑Filtered (142 problems) cEER ↓Unfiltered (All problems) Slice AP ↑Unfiltered (All problems) cEER ↓
Frozen k-means0.7420.3090.6620.268
Domino0.7180.1620.5370.249
ConceptDiff (ours)0.8070.1690.6500.208

Q3 & W6: Selecting K and cPCA Baseline

For using ConceptDiff in the real world, we believe KK to be a user-defined parameter whose role is to find and group data based on differences. Intuitively, if KK is very large (for instance, equal to the size of dataset AA), then each datapoint will be assigned to a new group since every datapoint has its own difference, which is why we keep KK small. That said, when the number of differences is unknown, we believe the best way to pick KK is to go with the highest cross-seed stability on the validation set. We tried this approach on our image benchmark, where we know the ideal KK is 3. Across different values of KK, we see that K=3K = 3 gives the best stability (0.522) and group recovery (0.460 ARI), while for larger KK, all heads remain active (Act = 1.00) and split the main groups into finer sub-concepts as separation continues to improve (cEER drops from 0.062 to 0.047). Here are the results:

Concept Budget (K)ARI ↑cEER ↓Act ↑Cov ↑Stab ↑
K = 20.326 ± 0.0180.0741.000.8230.339
K = 3 (Ideal K)0.460 ± 0.0200.0621.000.8270.522
K = 40.343 ± 0.0220.0551.000.8110.430
K = 60.246 ± 0.0160.0471.000.7950.358

In addition to that, we also tested Contrastive PCA (cPCA) on the 12 image benchmarks (K=3K=3) and saw that ConceptDiff works substantially better in both group recovery and separation. Since cPCA only captures global linear directions over pooled features rather than localized token-level concepts, we tried two forms of grouping, one based on the top-KK directions and the other using kk-means on the cPCA features. In comparison to both cases, ConceptDiff performed much better.

Method (K = 3, 12 Image Tasks)ARI ↑cEER ↓Act ↑Cov ↑Stab ↑
cPCA (top-K directions)0.1980.1611.000.5991.0 *
cPCA + k-means0.2670.2121.000.1960.999
ConceptDiff (ours)0.4600.0621.000.8270.522

* Directions are deterministic.

Q4: Use Of Foundation Models For Evaluation

We would like to clarify that ConceptDiff does not use a foundation model to generate or discover concepts. The foundation model (explainer and simulator) is only used as a post-hoc evaluation metric to measure monosemanticity.

In the introduction, our critique is about methods like VisDiff that rely on an LLM or VLM in the discovery loop to propose candidate differences. ConceptDiff instead discovers the groups and highlights the relevant tokens or patches directly from the data. To evaluate monosemanticity across our experiments, we directly followed the standard explainer-simulator protocol used by [1, 2] as a scalable replacement for human evaluation. In actual deployment, a user does not need the explainer model at all and can simply inspect the top-activating examples and highlighted regions directly.

For our evaluations, we used Gemini 3.1 Pro Preview for both the explainer and simulator roles. We apologize for missing the model name, and we will add this detail to the final version.

W3: Multi-Objective Metric Trade-offs

The reviewer is right that in our contrastive concept discovery setup, no single metric tells the full story in isolation because baselines often optimize one property at the expense of another. For example, frozen k-means has very high stability (0.990 on images and 0.981 on video) simply because clustering frozen features without contrastive supervision is nearly deterministic, yet it fails to capture the actual differences between datasets (ARI of 0.189 vs. our 0.460, and coverage of 0.395 vs. our 0.827). Conversely, classifier-based baselines like k-means (cls.), HypotheSAEs (cls.), and ProtoSumm can post a low cEER on active heads because their MLP is already trained to classify A vs. B, yet they either collapse onto 1 to 2 heads (Act = 0.22 on crash video) or lose stability and monosemanticity (on EditLens, k-means (cls.) has 0.257 stability and 0.456 simulation vs. our 0.927 and 0.670).

Across all benchmarks, ConceptDiff is the only method that keeps all heads active (Act = 1.00) with high coverage while simultaneously achieving strong separation, stability, and monosemanticity. That is what we intended to convey in the abstract, and we appreciate the reviewer pointing out that our wording could be read as claiming the best score on every individual metric in isolation. We will refine the sentence in the abstract to clearly state that ConceptDiff achieves the strongest overall balance across discriminability, coverage, active concept utilization, and monosemanticity without suffering from concept collapse.

W4: Use Of Frozen Token Encoders

The reviewer brings up a fair point on terminology. When we wrote "without a foundation model in the loop," we meant that we do not query a generative LLM or VLM during discovery to propose text hypotheses (as in VisDiff and TextDiff), not that we train from raw data without a feature extractor.

Modern pipelines across vision, text, and video are natively token-based. We intentionally designed ConceptDiff as a lightweight, plug-and-play module that attaches to any existing frozen encoder to understand how the backbone interprets data and use that to mark differences. Unlike LLM-in-the-loop baselines, which require thousands of expensive VLM/LLM calls and can only find differences the language model thinks to verbalize, ConceptDiff learns the differences directly from the contrastive token geometry in minutes. We appreciate the reviewer pointing out this ambiguity, and we will update the phrasing in the Abstract and Introduction to "without a generative LLM/VLM hypothesis proposer in the discovery loop" so the distinction is immediately clear.

References:

[1] Bills, et al., "Language models can explain neurons in language models", 2023.

[2] Movva, Rajiv, et al. "Sparse autoencoders for hypothesis generation." arXiv preprint arXiv:2502.04382 (2025).

Reviewer_xYSeOpenReview ↗

Official comment

After reviewing all reviewers' comment and authors' reply. I decided to keep my score.

I have read the paper and used AI to generate partial reviews. Frankly, I do not have sufficient expertise to assess this submission.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.