- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
The paper fits a full-rank ICA (whitening followed by FastICA) on frozen embeddings and maps the components back to the native space, where the directions therefore become non orthogonal. The authors make four claims: (i) the relative geometry of semantic factors directions is shared across encoders; (ii) ICA recovers more annotated factors than PCA and SAEs; (iii) the native-space geometry of the ICA directions is semantically structured, so related concepts end up closer together; (iv) the coordinates support useful interventions (retrieval, collateral damage, Waterbirds worst-group accuracy).
Critical strengths
S1 : The paper is clearly written. The method is simple and training-free, and it gives an exactly invertible map, which makes interventions easy to define.
S2 : The negentropy analysis (Fig. 3, 9) is a sensible connection to what ICA can identify: selective, presence-like factors are recovered better than continuous ones.
S3 : The qualitative visualizations (Fig. 4, 11) are nice, although they read as illustration more than evidence.
Critical weaknesses
W1 : Novelty. ICA has already been extensively studied and shown to yield interpretable axes. The oblique/biorthogonal framing is standard linear algebra (A = P⁻¹Rᵀ, dual basis B = A⁻ᵀ), and the paper itself says obliqueness is expected after inverse whitening. The remaining contribution is the claim that the cosine structure of the latent factors in embedding space is semantically meaningful. That claim carries the paper, and it is the least supported.
W2 : Illusory structure ? The paper starts from the observation that the geometric structure of ground truth latent factors is shared across models. Since this structure is based on the correlation between mean-diff vectors, any functional correlation (label co-occurence) will naturally lead to a geometric correlation. The paper needs to show that this effect goes beyond simple functional correlation (several controls are possible here, e.g., compare the cosine matrix given by (where is the dictionary) to the correlation matrix (where is the label of sample ), or take from model and from a randomly initialized model). Without proper control, the whole motivation underlying this work is null.
W3 : The central geometric claim (Table 2, Table 5) is under supported. From App. B, , so the cosines between the ICA directions depend on the covariance spectrum and the rotation. In an anisotropic embedding space, I would expect related concepts to be closer under many rotations. You need a control to show that the apparent structure you find goes beyond the covariance structure and reflects the success of ICA. Take the rotation given by ICA and an arbitrary rotation. Keep and the rest of the pipeline, and report the same statistics. Otherwise, it is impossible to tell whether the paper reports structure discovered by ICA or mere anysotropy. Section 3.3 promises controls in section 4, but the only comparison is against SAE. Direct comparison is hard to interpret due to several confounding factors.
- SAE atoms are not trained to be geometrically organised. If anything, the top-k sparsity constraint introduces well known inductive biases, biasing atoms towards mutual repulsion, since atoms compete for the same activation. The within vs across ratio in cosim is therefore not comparable for ICA and SAE. For example, if you look at the similarity between multiclass and multilabel probes for the same classes, and SAE atoms, you will find that one case is a much better fit to the SAE’s dictionary, due to their respective inductive biases that change the geometric organisation of the atoms recovered.
- Dictionary size is a massive confound, since finer grained features tend to be less correlated with the class average.
- Only ratios of mean cosines are reported. Please give the raw within and across values with confidence intervals, and state whether absolute or signed cosines are used (Table 5 uses |cos|, Table 2 is unclear). The ratio might be extremely sensitive due to the small denominator.
- Controlled references exist (mean-diff, LogReg) but are not used in this analysis.
Even if you are able to provide satisfactory controls for ICA, the existence of the same structure in SAE, plus possibly in other baselines, more comparable to ICA (dense or complete, like K-Means, NMF, SVD, etc … which might not all be exactly orthogonal in embedding space) would be a result on the hierarchical organisation of latent factors across methods than a claim about ICA itself.
W4 : Baselines are insufficient throughout. The paper claims that ICA is better than standard methods, most developed after ICA (e.g. sparse PCA, NMF, K-Means, many SAE architectures). Despite this very strong claim, the paper remains extremely lacking in its comparison to these methods …
- On the geometric organisation of the latent factors mentioned above, it compares only to SAEs.
- On subclass separability (figure 5, table 4), it does not compare to anything. Interpretability (Figure 4) is anecdotal and well known.
- Semantic coverage (Table 1) is compared against PCA and SAE only.
- Collateral damage (figure 6) is compared against supervised baselines but no other method (apart from PCA).
- Retrieval intervention (figure 7) and spurious correlation removal (table 3) are not compared against any baseline or competing method. About comparison to SAEs : for such geometric claims at the level of the method, you need to report several architectures, not several expansion factors. The latter is part of the ablations unless you have a specific point about dimensionality. The method is what is being tested here.
W5 : Subclass separability (Fig. 5, Table 4) is anecdotal. The metric is not defined in the main text. From appendix D.4 it seems to be a binary classifier trained on the n most extreme samples of two ICs (e.g. 100 fedora and 100 beret), with manual verification of the samples. This is a much easier task than detecting a subclass within its parent class or within the whole population. It covers only four subclasses with no explanation of how they were chosen, and 0.98 -- 1.00 on easy pairs (fedora vs. beret, light aircraft vs. fighter jet) would probably be reached by almost any embedding direction. No baseline is given. No comparison to other methods is proposed.
W6 : Retrieval intervention (Fig. 1, 7, 13, 14) is anecdotal. Same as above, this would usually be a major weakness. It is a compelling narrative illustration but does not serve any scientific purpose if not backed by quantitative results
minor points :
Spurious mitigation (table 3) is undersupported ? In all 4 out of 4 columns that report stds, the effect fall within the confidence interval. You claim that for the ERM columns std is negligible. Can you still report the values ? Assuming that they are indeed negligible and the effect is statistically significative for ERM, do you have any explanation as to why it would be in ERM but not in CA and DFR ? Moreover, you do not explain the selection algorithm for the number of components to remove. The number of removed ICs and of extreme samples was selected per setting as the best-performing configuration (3 to 30 removed ICs). If this selection used test worst-group accuracy, the conclusion is null.
And again, you do not compare to any baseline or other method.
Table 1 is hard to interpret.
- Coverage is defined only in words. As I understand it, for each annotated factor the best-aligned direction is selected, and coverage is the number of unique divided by the number of factors. Please define it formally.
- It is unclear whether matching is done in whitened or native space, and whether the measurement direction (rows of W) or the intervention direction (columns of A) is used. App. B shows these have different geometry.
- Matching each attribute to the best of D directions using labels inflates AUC, and a random rotation would show how much of ICA’s apparent results are just best-of-D selection.
- It is unclear whether AUC is computed on data held out from the ICA fit, and no variance across seeds is given although FastICA is stochastic.
- The “highest separability” claim holds for AUC, but SAEs match or beat ICA on coverage in some settings.
Questions for authors
see weaknesses.
5 comments
Response 2
W3. Is hierarchical geometry attributable to ICA rather than anisotropy?
We thank the reviewer for suggesting this important control. We implemented the proposed random-rotation baseline (RandRot), keeping the whitening and inverse-whitening transformations fixed while replacing the ICA rotation with a random orthogonal rotation. We repeat the experiment three times and report the best result, providing an optimistic comparison for this control. In Table 2, ICA achieves within/across-parent similarity ratios of 2.82/5.72/4.05 across CLIP/DINO/I-JEPA, compared with 1.64/1.21/1.26 for RandRot. We additionally report the raw mean absolute cosine similarities (Table 9), showing that ICA consistently exhibits both higher within-parent and lower across-parent alignment than RandRot. These results directly demonstrate that embedding covariance and anisotropy alone cannot explain the hierarchical organization recovered by ICA.
We further expanded the evaluation to include Sparse PCA, NMF, and K-Means, alongside SAE baselines. While K-Means achieves a particularly high hierarchical ratio on DINO, its performance is substantially weaker on CLIP and I-JEPA. In contrast, ICA consistently reveals strong hierarchical organization across all three encoders. We also added supervised MeanDiff and LogReg references (Table 10). ICA exhibits stronger hierarchical organization than LogReg on CLIP and DINO, while remaining below MeanDiff. Importantly, these supervised methods provide useful references rather than direct competitors to unsupervised discovery methods such as ICA.
We recognize that different decomposition methods have distinct inductive biases affecting their geometric organization. Therefore, our central evidence comes from the covariance-controlled ICA–RandRot comparison, while the broader baselines establish the consistency and relative strength of the recovered semantic structure.
Table 2 (in the revised paper)
| Model | ICA | RandRot | PCA | SPCA | NMF | K-Means | SAE-D | SAE-2D | SAE-4D | SAE-8D | SAE-16D |
|---|---|---|---|---|---|---|---|---|---|---|---|
| CLIP | 2.821 | 1.637 | -- | 0.188 | 1.066 | 1.252 | 1.504 | 1.467 | 1.326 | 1.243 | 1.203 |
| DINO | 5.721 | 1.206 | -- | 1.308 | 1.004 | 9.363 | 1.279 | 1.439 | 1.454 | 1.330 | 1.220 |
| I-JEPA | 4.046 | 1.255 | -- | 0.271 | 1.077 | 1.039 | 1.139 | 1.081 | 1.074 | 1.044 | 1.029 |
Table 9 (in the revised paper):
| Method | CLIP Within | CLIP Across | DINO Within | DINO Across | I-JEPA Within | I-JEPA Across |
|---|---|---|---|---|---|---|
| ICA | 0.228 | 0.081 | 0.129 | 0.023 | 0.203 | 0.050 |
| RandRot | 0.165 | 0.101 | 0.073 | 0.061 | 0.093 | 0.074 |
| PCA | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 | 0.000 |
| Sparse PCA | 0.001 | 0.005 | 0.000 | 0.000 | 0.000 | 0.000 |
| NMF | 0.183 | 0.172 | 0.153 | 0.153 | 0.173 | 0.161 |
| K-Means | 0.879 | 0.702 | 0.468 | 0.050 | 0.959 | 0.923 |
| SAE-D | 0.052 | 0.035 | 0.037 | 0.029 | 0.027 | 0.024 |
| SAE-2D | 0.047 | 0.032 | 0.040 | 0.028 | 0.025 | 0.023 |
| SAE-4D | 0.042 | 0.032 | 0.039 | 0.027 | 0.024 | 0.023 |
| SAE-8D | 0.038 | 0.030 | 0.034 | 0.026 | 0.023 | 0.022 |
| SAE-16D | 0.036 | 0.030 | 0.031 | 0.025 | 0.023 | 0.022 |
Table 10 (in the revised paper):
| Model | ICA Within | ICA Across | ICA Ratio | MeanDiff Within | MeanDiff Across | MeanDiff Ratio | LogReg Within | LogReg Across | LogReg Ratio |
|---|---|---|---|---|---|---|---|---|---|
| CLIP | 0.228 | 0.081 | 2.821 | 0.623 | 0.168 | 3.701 | 0.177 | 0.069 | 2.573 |
| DINO | 0.129 | 0.023 | 5.721 | 0.247 | 0.035 | 7.152 | 0.140 | 0.037 | 3.799 |
| I-JEPA | 0.203 | 0.050 | 4.046 | 0.586 | 0.094 | 6.262 | 0.308 | 0.057 | 5.362 |
Response 3
W4 and W6. Breadth and fairness of baselines.
We substantially expanded the comparison suite. The revised factor-recovery evaluation (Table 1) now includes RandRot, Sparse PCA, NMF, K-Means, and an additional SAE architecture (SAE-Best), alongside PCA and the original SAE variants. The geometric analysis (Tables 2 and 9) also includes these additional decomposition methods. Importantly, the semantic intervention analysis (Figure 6, 18) demonstrates that ICA achieves substantially lower collateral semantic drift than competing unsupervised decompositions and supervised reference directions, approaching the zero-drift oracle. We further added LEACE and SpLiCE as concept-removal baselines (Figure 17), where ICA again achieves stronger preservation of non-target semantics, and quantitative retrieval comparisons against PCA and SAEs (Figure 16), demonstrating improved identity preservation.
Importantly, the expanded baselines reveal a distinction between recovering many apparently distinct concepts and recovering semantically predictive factors. For example, K-Means achieves the highest raw coverage in Table 1, but with relatively low AUC. Under the stricter evaluation in the new Table 4, which requires distinct directions with AUC ≥ 0.8, its coverage decreases substantially, whereas ICA achieves the highest thresholded coverage across all six dataset and encoder settings. This demonstrates that ICA combines broad semantic coverage with strong separability. These additions provide a substantially broader comparison across semantic recovery, geometric organization, selective interventions, and identity preservation.
Table 1 (in the revised paper):
| Method | CelebA CLIP Cov. (%) | CelebA CLIP AUC | CelebA DINO Cov. (%) | CelebA DINO AUC | CelebA I-JEPA Cov. (%) | CelebA I-JEPA AUC | MSCOCO CLIP Cov. (%) | MSCOCO CLIP AUC | MSCOCO DINO Cov. (%) | MSCOCO DINO AUC | MSCOCO I-JEPA Cov. (%) | MSCOCO I-JEPA AUC |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| ICA | 70.0 | 0.798 | 70.0 | 0.773 | 62.5 | 0.804 | 73.8 | 0.931 | 83.8 | 0.902 | 73.8 | 0.878 |
| RandRot | 57.5 | 0.765 | 57.5 | 0.767 | 65.0 | 0.785 | 63.8 | 0.876 | 76.2 | 0.858 | 65.0 | 0.843 |
| PCA | 20.0 | 0.736 | 10.0 | 0.741 | 7.5 | 0.751 | 15.0 | 0.827 | 36.2 | 0.850 | 12.5 | 0.783 |
| SPCA | 22.5 | 0.736 | 10.0 | 0.742 | 7.5 | 0.751 | 13.8 | 0.824 | 37.5 | 0.852 | 12.5 | 0.783 |
| NMF | 12.5 | 0.730 | 25.0 | 0.741 | 10.0 | 0.755 | 41.2 | 0.858 | 65.0 | 0.871 | 47.5 | 0.806 |
| K-Means | 85.0 | 0.750 | 82.5 | 0.739 | 77.5 | 0.745 | 87.5 | 0.873 | 92.5 | 0.913 | 76.2 | 0.839 |
| SAE-D | 62.5 | 0.774 | 30.0 | 0.768 | 15.0 | 0.733 | 63.8 | 0.906 | 78.8 | 0.857 | 33.8 | 0.796 |
| SAE-2D | 60.0 | 0.780 | 37.5 | 0.766 | 22.5 | 0.741 | 66.2 | 0.916 | 77.5 | 0.876 | 32.5 | 0.829 |
| SAE-4D | 72.5 | 0.735 | 32.5 | 0.739 | 20.0 | 0.742 | 65.0 | 0.916 | 80.0 | 0.870 | 38.8 | 0.812 |
| SAE-8D | 67.5 | 0.754 | 40.0 | 0.748 | 25.0 | 0.765 | 72.5 | 0.917 | 77.5 | 0.874 | 41.2 | 0.833 |
| SAE-16D | 65.0 | 0.768 | 50.0 | 0.757 | 17.5 | 0.763 | 67.5 | 0.920 | 76.2 | 0.873 | 50.0 | 0.828 |
| SAE-Best-D | 64.3 | 0.766 | 45.8 | 0.758 | 37.5 | 0.785 | 72.5 | 0.935 | 76.6 | 0.872 | 71.6 | 0.853 |
| SAE-Best-4D | 71.8 | 0.789 | 61.8 | 0.762 | 40.8 | 0.795 | 73.8 | 0.934 | 83.4 | 0.887 | 72.9 | 0.860 |
Table 4 (in the revised paper):
| Method | CelebA CLIP Cov. (%) | CelebA CLIP AUC | CelebA DINO Cov. (%) | CelebA DINO AUC | CelebA I-JEPA Cov. (%) | CelebA I-JEPA AUC | MSCOCO CLIP Cov. (%) | MSCOCO CLIP AUC | MSCOCO DINO Cov. (%) | MSCOCO DINO AUC | MSCOCO I-JEPA Cov. (%) | MSCOCO I-JEPA AUC |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| ICA | 65.0 | 0.818 | 57.5 | 0.810 | 60.0 | 0.817 | 96.2 | 0.939 | 93.8 | 0.916 | 88.8 | 0.890 |
| RandRot | 45.0 | 0.780 | 50.0 | 0.790 | 55.0 | 0.793 | 86.2 | 0.891 | 78.8 | 0.868 | 71.2 | 0.842 |
| PCA | 32.5 | 0.740 | 25.0 | 0.737 | 42.5 | 0.752 | 70.0 | 0.846 | 70.0 | 0.860 | 50.0 | 0.790 |
| SPCA | 30.0 | 0.736 | 30.0 | 0.741 | 42.5 | 0.752 | 67.5 | 0.843 | 72.5 | 0.863 | 50.0 | 0.793 |
| NMF | 42.5 | 0.753 | 45.0 | 0.774 | 45.0 | 0.772 | 80.0 | 0.877 | 81.2 | 0.883 | 62.5 | 0.831 |
| K-Means | 37.5 | 0.761 | 42.5 | 0.770 | 50.0 | 0.768 | 81.2 | 0.886 | 92.5 | 0.923 | 71.2 | 0.846 |
| SAE-D | 57.5 | 0.802 | 55.0 | 0.804 | 40.0 | 0.766 | 86.2 | 0.914 | 81.2 | 0.877 | 52.5 | 0.806 |
| SAE-2D | 57.5 | 0.804 | 55.0 | 0.805 | 45.0 | 0.769 | 90.0 | 0.923 | 81.2 | 0.882 | 65.0 | 0.836 |
| SAE-4D | 52.5 | 0.782 | 55.0 | 0.810 | 45.0 | 0.775 | 87.5 | 0.922 | 80.0 | 0.885 | 60.0 | 0.820 |
| SAE-8D | 47.5 | 0.799 | 55.0 | 0.799 | 52.5 | 0.782 | 91.2 | 0.923 | 81.2 | 0.887 | 71.2 | 0.841 |
| SAE-16D | 52.5 | 0.794 | 47.5 | 0.793 | 50.0 | 0.785 | 90.0 | 0.922 | 83.8 | 0.889 | 68.8 | 0.837 |
Response 1
We thank the reviewer for the detailed critique and concrete experimental recommendations. We agree that standard ICA and biorthogonal algebra are not, by themselves, novel, and that the main claim requires controls beyond qualitative examples. We have substantially expanded our experiments and the manuscript to better support our claims. We address each point below.
W1. Novelty and scope of the claims.
We agree that ICA and biorthogonal linear algebra are established tools, and do not claim either as a methodological novelty. Our central contribution is the finding that semantic factors in vision foundation models exhibit a coherent, non-orthogonal geometric organization, where semantically related factors are geometrically aligned yet remain individually controllable. We demonstrate that ICA provides a complete, unsupervised biorthogonal coordinate system that reveals this organization and supports selective semantic interventions.
Importantly, the semantic organization is not merely a consequence of covariance anisotropy. We added new random-rotation controls (Table 2) that preserve these properties but exhibit substantially weaker hierarchical organization than ICA across all three encoders. This is further supported by raw within- and across-parent similarities (Table 9) and comparisons with randomly initialized encoders (Figure 10).
Beyond these geometric controls, we have substantially expanded the evaluation with additional decomposition baselines (Tables 1 and 2, Figure 6), a complementary semantic recovery evaluation (Table 4), quantitative identity-preserving retrieval (Figure 16), concept-removal baselines (Figure 17), and additional disentanglement analyses (Figures 18–19). Together, these experiments strengthen both the evidence for meaningful semantic geometry and its practical value for selective interventions. We discuss the individual additions in detail below.
W2. Could cross-encoder agreement reflect only co-occurrence or architecture?
We agree that correlations between annotated attributes contributes to the observed similarity between mean-difference RSMs. Importantly, this is a motivating observation, not evidence that the underlying factors are statistically independent. Our hypothesis is that the geometry of learned representations reflects the semantic relationships present in the data, including, but not limited to, attribute co-occurrence.
To further examine the role of architecture and training, we added randomly initialized encoder controls (Figure 10). These exhibit weaker agreement with pretrained representations than the agreement observed between pretrained encoders. The moderate correlations (r=0.50 for CLIP and r=0.53 for DINOv2) indicate that data and architectural biases contribute to this structure, while representation learning further reinforces it. Thus, while attribute co-occurrence may contribute to the shared geometry, the stronger agreement among pretrained encoders suggests that this structure is not solely explained by data correlations or architectural biases.