acceptodds
Under review as a conference paper at ICLR 2027

Towards Open-Ended Visual Scientific Discovery with Sparse Autoencoders

Abstract

Foundation models in several scientific domains, including visual domains, learn representations that capture complex semantics from their respective fields. Despite this, most existing applications focus on pre-specified concepts or targets, excelling at confirmation but not open-ended discovery of unknown patterns. We study whether sparse autoencoders (SAEs) with heuristic feature ranking can provide a practical methodology for surfacing candidate visual features from vision foundation models for scientific discovery without concept-specific supervision. We evaluate this approach in three stages: (1) general-purpose concept rediscovery, (2) domain-specific concept rediscovery, and (3) question-driven feature ranking. On ADE20K, SAEs outperform baseline methods (k-means, PCA, SemiNMF) across most downstream metrics for concept rediscovery. On FishVista, the same approach recovers fine-grained fish anatomical features without concept supervision. Finally, on Heliconius butterflies, a heuristic based on the Mann-Whitney U statistic consistently surfaces features corresponding to real diagnostic keys between mimic subspecies. Together, these results show that SAEs with heuristic ranking can efficiently learn and surface semantic structure in vision foundation model representations. We view this as a methodological prerequisite for open-ended visual scientific discovery: the present experiments validate feature surfacing and ranking through rediscovery, while prospective discovery of previously unknown scientific phenomena remains future work.

Primary area:
interpretability and explainable AI
Keywords:
sparse autoencoders, mechanistic interpretability, vision foundation models, concept rediscovery, AI for scientific 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_ZqbxOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies sparse autoencoders as tools for extracting browsable visual concepts from foundation-model patch representations. It evaluates concept rediscovery on ADE20K and FishVista, compares decomposition methods and SAE variants, and ranks features separating Heliconius mimic pairs using a Mann-Whitney-based selectivity score. Expert interpretation links highly ranked butterfly features to known diagnostic traits. The paper explicitly frames these results as preparation for future discovery rather than a new biological discovery.

Critical strengths
  1. The evaluation measures a property that reconstruction alone cannot guarantee. A representation can reconstruct inputs accurately while spreading a semantic concept across many coordinates. Since the intended use is expert inspection of individual features, axis-aligned probes and concept coverage are appropriate measures of utility. The improvements over PCA and SemiNMF on ADE20K, together with FishVista anatomy recovery, support the value of sparse feature extraction for this purpose.
  2. The biological case study demonstrates a concrete human-in-the-loop use. The ranking heuristic reduces a large dictionary to a manageable set of candidates, and the displayed activations can be checked against known diagnostic traits. The authors appropriately distinguish this rediscovery result from establishing novel biological knowledge. This bounded empirical contribution is sufficient to be useful without requiring a new SAE architecture.
Critical weaknesses
  1. Label-free feature learning is not the same as label-free workflow selection. SAE parameters are learned without concept annotations, but layer and SAE choices use labeled probe performance in the rediscovery experiments. In an unannotated discovery setting, that selection signal is unavailable. Furthermore, selecting the best axis from 16,384 latents gives a different search opportunity from a smaller decomposition, so the results combine semantic organization with dictionary-size effects. The central practical question is whether an annotation-independent selection rule retains the reported utility. A fixed transferred configuration or a declared reconstruction/sparsity-based rule would address this directly.
  2. Species separability does not uniquely identify biological novelty or useful trait diversity. The selectivity score rewards any feature correlated with the species label, including acquisition cues, and does not distinguish a new trait from several redundant encodings of one known trait. Appendix I illustrates this distinction: several top features correspond to the same diagnostic pattern. Expert validation is therefore essential, but feature-level hit rate alone does not measure the number of distinct useful hypotheses produced. Object-only ranking controls and counting distinct validated traits would align the evaluation more closely with the intended discovery task.
Questions for authors
  1. Clarification: Which layer and SAE selection decisions require concept labels, and which rule would you use unchanged on a dataset without such labels? Can the existing experiments show how a fixed or annotation-independent choice compares with probe-selected configurations?
  2. Evidence affecting the recommendation: At a comparable dictionary or candidate-inspection budget, does the advantage over the strongest decomposition baseline persist? This would clarify whether improved browsability comes from the representation itself or primarily from a larger candidate pool.
  3. Evidence affecting the recommendation: Do the top butterfly features remain associated with the diagnostic wing regions when backgrounds and collection labels are excluded? How many distinct diagnostic traits are represented among the top five features for each pair? These checks would sharpen the biological interpretation without requiring a prospective discovery study.
1 comment

We thank the reviewers for taking the time to give their thoughtful and thorough comments and questions. Our responses to the listed weaknesses and questions are provided below:

The reviewer notes that concept labels were used to select the best SAE for comparison with baseline methods, and that a dimensionality mismatch favors SAEs over baseline methods.

We have updated Table 1 to instead use a label-free selection criterion for which SAE to use for comparison, reporting metrics for the Pareto-optimal SAE with L0 closest to 256 (Similar L0 to PCA/SemiNMF). We aim to compare SAEs to baseline methods at similar levels of sparsity, as we argue that the scalable, overcomplete nature of SAEs is what makes them ideal for trait discovery. The results of this experiment remain the same, with SAEs improving on downstream metrics.

The reviewer notes that background features may have high selectivity scores, and that the same trait may have several highly-ranked, redundant features.

We believe this is an important limitation of this result, and as such, have updated the discussion section to touch on this point. Additionally, we have added details about the number of distinct traits identified in the top-5 selectivity features for each subspecies mimic pair (Section 4.3). Despite this, we believe that concept-label-free hypothesis generation (even when a trait is surfaced multiple times) enables visual feature discovery at scale, allowing domain scientists to quickly identify meaningful and possibly novel traits. Additionally, while background features may drive selectivity, we identify several non-background features based on selectivity ranking.

Q1 Concept labels are used to select which layer of activations to use for evaluation across all methods, comparing SAE variants and model backbones, qualitative examples of features from ADE20K, and measuring the recoverability of specific traits on FishVista. Fine-grained concept labels are not used on Heliconius images.

Q2 When compared to baseline methods, SAEs improve on the percentage of highly-ranked features that correspond to known diagnostic traits at comparable dictionary size (kk-means) and candidate inspection budget (kk-means/PCA), as detailed in Section 4.3. Furthermore, SAEs improve substantially on quantitative metrics on controlled rediscovery tasks, indicating that SAEs capture features that more holistically capture visual traits (Table 1).

Q3 While we have tested removing backgrounds from images with this approach (with similar results), this is not included as part of this work. Instead, this work is focused on providing a methodology for identifying potentially meaningful visual traits without relying on fine-grained concept labels. We hope to expand on this work in the future by utilizing background removal, and have made note of this in the provided work (Appendix A).

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

Vision Foundation Models (VFMs) learn rich representations, but they are treated as black boxes for pre-specified targets, which supports confirmation of known concepts rather than open-ended discovery. This paper proposes SAEs trained on VFM activations, primarily focused on DINOv3, combining them with heuristic ranking as a method to surface candidate visual features without concept-specific supervision: (1) general-purpose concept rediscovery on ADE20K, where SAEs outperform k-means, PCA, and SemiNMF on most downstream concept-alignment metrics (probe R, mAP, Purity@k, Coverage@τ), and where Matryoshka and TopK variants further improve results; (2) domain-specific concept rediscovery on FishVista, where SAE features align with fine-grained fish anatomical parts without part-level supervision; and (3) question-driven feature ranking on Heliconius butterflies, where a selectivity heuristic based on the Mann–Whitney U statistic ranks features that. The authors position this work as a methodological prerequisite for open-ended scientific discovery. They state explicitly that it validates feature surfacing and ranking through rediscovery, and that discovery of previously unknown phenomena is left to future work.

Critical strengths
  1. SAEs are trained only on ImageNet1K and evaluated on ADE20K labels, so no labels enter training. Results show that reconstruction metrics (NMSE, L0) do not track alignment; k-means and PCA are competitive on reconstruction, yet SAEs reach mAP 0.224 vs. 0.093 for the best baseline, and Coverage@τ 0.264 versus 0.079. This is the most solid result in the paper and supports the claim that SAE features are more semantically aligned than the baselines.

  2. Heuristic Ranking. The selectivity score needs only species-level labels and, as per the authors, surfaces known diagnostic traits of all three mimic pairs. I consider this a modest strength, since the evidence is limited to 3 pairs and 1 annotator.

Critical weaknesses
  1. The only evidence for the ranking heuristic is the Heliconius butterflies. The selectivity score is never applied to the DINOv3 features, PCA components, or k-means clusters, so the paper cannot show that SAE is responsible for the result. There’s also no chance comparison, and the match criterion is applied flexibly. Diffuse features are excluded (Fig. 17), and one top feature has no documented trait (Fig. 19). The evidence supports “the method can surface some known traits,” not that it’s an effective discovery tool.

  2. PCA is capped at 1,024 components, while SAEs and k-means use 16,384, and SAE variants in Table 2 span very different sparsity levels. The components (ReLU, TopK, and Matryoshka SAEs, and a Mann–Whitney-based score) are all existing methods.

Questions for authors
  1. Can you apply the same selectivity score to (A) raw DINOv3 patch activations, (B) PCA activations, and (C) k-means clusters on the Heliconius data, and report how often the top-5 features match known diagnostic traits? Without this, it is unclear whether the SAE is responsible for the result or whether any feature basis would surface these traits.

  2. How many of the 16,384 latents exceed a selectivity of 0.9 for each mimic pair? What fraction of a random sample of those high-selectivity features would an expert judge to match a known trait? This is needed to interpret “the top-5 are interpretable”.

  3. Was the expert blind to feature rank and to which SAE a feature came from? Were the annotation criteria fixed in advance? How were features with diffuse activation maps (Fig. 17) handled, and would counting them change the 60–100% figure?

1 comment

We thank the reviewers for taking the time to give their thoughtful and thorough comments and questions. Our responses to the listed weaknesses and questions are provided below:

The reviewer notes that selectivity ranking and expert interpretation are performed only for SAE features, not for PCA components, k-means clusters, or raw DINOv3 dimensions.

To address this, we apply the same selectivity ranking and expert interpretation to PCA components, kk-means clusters, and raw DINOv3 dimensions. We find that 0-20% of the 5 highest-selectivity PCA components correspond to known diagnostic traits, 40-100% of the 5 highest-selectivity kk-means clusters correspond to known diagnostic traits, and none of the 5 highest-selectivity DINOv3 dimensions correspond to known diagnostic traits (Section 4.3). We also compare top-10 selectivity scores between SAEs, PCA components, kk-means clusters, and raw DINOv3 dimensions, and find that SAEs achieve the highest average top-10 scores (Section 4.3, Figure G6) in all but one subspecies pair for one value of kk. These results, in conjunction with those in Table 1, are consistent with our original results, supporting that SAEs outperform baselines for surfacing relevant, meaningful, cohesive visual traits.

With this, we agree that this work focuses on surfacing known traits. However, we argue that surfacing known traits without using fine-grained concept labels is a critical and previously unstudied step in the discovery process. As such, this work presents a principled and tested method for generating visual hypotheses.

The reviewer notes a dimensionality mismatch between SAEs and other baselines.

We have updated Table 1 to instead use a label-free selection criterion for which SAE to use for comparison, reporting metrics for the Pareto-optimal SAE with L0 closest to 256 (Similar L0 to PCA/SemiNMF). We aim to compare SAEs to baseline methods at similar levels of sparsity, as we argue that the scalable, overcomplete nature of SAEs is what makes them ideal for trait discovery. The results of this experiment remain the same, with SAEs improving on downstream metrics.

Additionally, the reviewer notes that Table 2 has SAEs of varying sparsity levels.

The SAEs in Table 2 are chosen according to the maximum Probe RR for each SAE type to give a best-case comparison.

Finally, the reviewer notes that the components used in the proposed method are all pre-existing.

While this is true, this work aims to combine existing methods into a novel framework for proposing visual trait hypotheses, an unexplored and untested application.

Q2

Each mimic pair had varying numbers of features with selectivity > 0.9:

  • 6 for cyrbia/cythera
  • 7 for lativitta/malleti
  • 3 for notabilis/plesseni.

While additional expert scoring beyond the addition of baseline method interpretations is outside the scope of the rebuttal period, 3, 5, and 4 of the top-5 selectivity features correspond to known diagnostic traits for cythera/cyrbia, lativitta/malleti, and notabilis/plesseni, respectively. All but 2 of the presented features with selectivity > 0.9 correspond to known diagnostic traits.

Q3

The expert was blind to which SAE a feature came from, as well as the feature rank. The annotation criteria were fixed, with the expert being given the 6 images and activation maps with the highest and lowest maximum feature values for each subspecies and per-image maximum feature values. The expert was then asked if the feature maps correspond to a known diagnostic trait. We aim to perform a larger and more rigorous evaluation of expert interpretation in the future, and this has been noted in the updated version of the paper.

Diffuse feature maps are considered uninterpretable, as they do not clearly highlight any specific visual trait. Because of this, diffuse features are not counted in the 60-100% number given in Section 4.3.

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

This paper studies whether sparse autoencoders (SAEs) trained on DINOv3 patch activations, plus a ranking heuristic, can surface candidate visual features for scientific discovery without concept labels. The approach is tested in three stages: ADE20K concept rediscovery against k-means, PCA, and SemiNMF; FishVista body-part rediscovery; and ranking features for Heliconius mimic pairs with a Mann-Whitney selectivity score. SAEs do best on most alignment metrics, and 60 to 100% of the top-5 butterfly features match known diagnostic traits.

Critical strengths
  1. The framing is honest. The paper validates SAEs as an instrument through controlled rediscovery and says plainly that prospective discovery is future work, so most claims stay within the evidence.
  2. The evaluation covers three settings that get more domain-specific, with ablations over backbone, ViT size, layer, and SAE variant and released code. That makes it a practical reference for applying SAEs to scientific images.
Critical weaknesses
  1. The most important opportunity, I think, is to add controls to the butterfly experiment. The SAE variants and the selectivity score (|2·AUC - 1|) are existing tools, so the contribution rests on the evaluation, and this is the stage closest to discovery. Right now one co-author labels the top-5 features for 3 pairs (L425 to 427), pooled over several Pareto-optimal SAEs. The paper does not report how many images each subspecies has. Max-pooling over all patches also lets background, ruler, or label patches drive selectivity, and backgrounds switch between green and grey across specimens (Fig. 6, Figs. I7 to I9). Four additions would make this convincing: a label-permutation null, the same ranking applied to PCA components and raw DINOv3 dimensions, a species-supervised baseline such as ProtoPNet [1], and blinded labels from two or more experts.

  2. The ADE20K comparison needs matched baselines. Picking the best latent per class favors large dictionaries, yet the SAE has 16,384 latents against 256 PCA/SemiNMF components (Table 1). The text says reconstruction is comparable (L323, Table 1 caption), but at matched sparsity Fig. G5 shows the SAE well ahead (roughly 0.37 vs. 0.61 NMSE at L0 ≈ 64), so part of the downstream gain may come from better reconstruction. SAEs are also selected by ADE20K probe loss (L268 to 269), which is supervision a real discovery setting will not have. An overcomplete SemiNMF at matched L0, Archetypal SAE [2], and a label-free selection rule would address this.

[1] Chen, C., et al. "This Looks Like That: Deep Learning for Interpretable Image Recognition." NeurIPS, 2019.
[2] Fel, T., et al. "Archetypal SAE: Adaptive and Stable Dictionary Learning for Concept Extraction in Large Vision Models." ICML, 2025.

Questions for authors
  1. How many dorsal images per subspecies does each pair have, and what maximum selectivity does a label-shuffled run reach over all features and SAEs?
  2. Of the top 50 features per pair, how many fire mainly off the wing?
  3. How is a k-means latent defined for the 1-D probe? Its Probe R of 0.001 (Table 1) suggests the choice matters.

I'm willing to raise my score if the rebuttal adds the null control and the matched baselines.

3 comments

We thank the reviewers for taking the time to give their thoughtful and thorough comments and questions. Our responses to the listed weaknesses and questions are provided below:

The reviewer notes that background features may drive selectivity, and that selectivity ranking and interpretation are not done for PCA components or raw DINOv3 dimensions.

We agree with the reviewer that background features such as rulers, labels, or other imaging artifacts can drive high selectivity scores. Cropping/segmenting wings would address this, and we have updated the paper to list this as a possible improvement to this work.

With this, we apply the same selectivity ranking and expert interpretation to PCA components, kk-means clusters, and raw DINOv3 dimensions. We find that 0-20% of the 5 highest-selectivity PCA components correspond to known diagnostic traits, 40-100% of the 5 highest-selectivity kk-means clusters correspond to known diagnostic traits, and none of the 5 highest-selectivity DINOv3 dimensions correspond to known diagnostic traits (Section 4.3). We also compare top-10 selectivity scores between SAEs, PCA components, kk-means clusters, and raw DINOv3 dimensions, and find that SAEs achieve the highest average top-10 scores (Section 4.3, Figure G6) in all but one subspecies pair for one value of kk. These results, in conjunction with those in Table 1, are consistent with our original results, supporting that SAEs outperform baselines for surfacing relevant, meaningful, cohesive visual traits.

The reviewer also notes that the current work would be strengthened by more information about butterfly data and a label-permutation null to validate the usefulness of SAE features.

We have added a table detailing the number of images per subspecies (Table E2) in the Appendix. It is also included below:

Mimic PairSubspeciesNumber of Images
1H. erato cyrbia412
H. melpomene cythera389
2H. erato lativitta276
H. melpomene malleti301
3H. erato notabilis198
H. melpomene plesseni224

Additionally, we have added a label permutation null experiment (Appendix H), which demonstrates that the discovered SAE features being predictive of species is not due to random chance.

The reviewer notes that the current work would be strengthened by including additional experts in the interpretation process.

While we would like to enlist additional expert review and species-supervised baselines, this is beyond the scope of the current work, as only one Heliconius expert is included as an author. We plan to implement a more robust interpretation process in future work and have listed this as a future direction in the limitations section, accordingly.

The reviewer notes a dimensionality mismatch between SAEs and other baselines.

We have updated Table 1 to instead use a label-free selection criterion for which SAE to use for comparison, reporting metrics for the Pareto-optimal SAE with L0 closest to 256 (Similar L0 to PCA/SemiNMF). We aim to compare SAEs to baseline methods at similar levels of sparsity, as we argue that the scalable, overcomplete nature of SAEs is what makes them ideal for trait discovery. The results of this experiment remain the same, with SAEs improving on downstream metrics.

The reviewer notes that the inclusion of additional methods, such as Archetypal SAEs or ProtoPNet, would strengthen this work.

In the future, we would like to explore using Archetypal SAEs or comparing to ProtoPNet, but believe that this is outside of the scope of this work. Rather, we intend for this work to substantiate that SAEs outperform baseline methods in learning features corresponding to biologically relevant morphology, and that heuristic ranking allows for efficient surfacing of meaningful features.

Q2

While it is difficult to provide a direct answer to the authors' full question in the discussion timeframe, among the top-5 selectivity features among all 3 subspecies pairs, only 1 may correspond to a non-wing trait (Figure J9, Rank #4). Anecdotally, at lower selectivity, background features become more prevalent.

Q3

For the 1-D probes, the nth k-means latent is given a value of 1(1−(distance to centroid))\frac{1}{(1 - (\text{distance to centroid}))} if it is closest to the nth centroid, and given a value of 0 if a patch is closest to a different centroid. This means that each patch has exactly one non-zero latent.

Thank you for the detailed response. The new baseline rankings, the label-permutation null, and Table E2 are useful additions. Two of my concerns remain. First, background patches can still drive selectivity, and I think a wing-cropped or background-removed run is needed to rule this out. Please also report the maximum selectivity under the null. With about 200 images per class it is likely well below 0.9, so it shows the features beat chance, which acquisition cues would also do. Second, Table 1 now matches L0 but not dictionary size (16,384 vs. 256). The ReLU SAE in Table 1 (L0 264) also reaches mAP 0.386 and Coverage 0.589, while the ReLU row in Table 2 (L0 379) gives 0.224 and 0.264. Could you explain this gap? I will keep my score for now and would reconsider if these points are resolved.

Thank you for the further suggestions on how to improve this work. We appreciate the reviewer's time in responding to edits.

While background patches may drive selectivity, it is unfortunately infeasible to perform a background-removed experiment in the discussion timeframe. With this, we have noted that background removal may help to strengthen the results, and we hope to perform such experiments in the future.

A table has been added with maximum selectivity values among real SAE features as well as under the label-permutation null, along with pp-values (Table H4). This table is also included below:

Subspecies PairMax. selectivityNull Max. Selectivitypp-value
cyrbia / cythera0.9000.4060.001
lativitta / malleti0.9700.1180.001
notabilis / plesseni0.9620.2300.001

Finally, while Table 1 compares baseline methods to SAEs at matched L0, Table 2 compares SAE methods at optimal RR-Probe values for each approach. In line with reviewer comments, we have updated this to use a similar approach to Table 1: SAEs are now selected based on Pareto-optimality at comparable L0. The results are similar, with Matryoshka Top-K SAEs beating vanilla and Matryoshka ReLU SAEs on all downstream metrics, and with DINOv3 outperforming PE as a backbone (Table 2).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.