acceptodds
Under review as a conference paper at ICLR 2027

The Independence Prior of SAEs Fragments Visual Concepts

Abstract

Sparse Autoencoders (SAEs) decompose model activations into sparse combinations of interpretable dictionary atoms. Although SAEs are grounded in the Linear Representation Hypothesis (LRH), their objective smuggles in an additional prior: concepts across patches are treated as independent, an assumption clearly violated by natural images and by the activations they induce. We therefore specialize LRH to vision through the Markov-Field Linear Representation Hypothesis (MFLRH), which adds the missing spatial dependencies to the LRH assumptions. We thus propose Spatial-SAE as an amortized MAP estimator under the MFLRH. Spatial-SAE consistently outperforms standard SAEs in concept recovery and interpretability. Across four variants, it achieves a 96% average win rate on synthetic concept recovery and improves interpretability on DINOv2 activations, at a reconstruction cost concentrated in high spatial frequencies.

Primary area:
interpretability and explainable AI
Keywords:
Sparse Autoencoders, Vision Interpretability, Spatial Inductive Biases, Concept Recovery, Graph Regularization

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_W6uKOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses spatial fragmentation of visual concepts learned by SAEs. The authors propose the Markov-Field Linear Representation Hypothesis, which models spatial dependencies by decomposing sparse codes into smooth and innovation fields. Based on this hypothesis, they introduce SPATIAL-SAE, which incorporates spatial regularization into SAE training. Experiments on synthetic and real-world vision representations demonstrate improvements in concept recovery on synthetic data and in spatial coherence and interpretability on vision representations.

Critical strengths
  • The paper is grounded in a solid theoretical framework and supported by systematic synthetic experiments and real-world evaluations. Overall, the work appears rigorous and reflects substantial effort in both methodology and experimental validation
  • Distinguishing between smooth and innovation fields to account for the spatial characteristics of visual representations offers an interesting and novel perspective on SAE-based concept discovery
Critical weaknesses

1. The benefits and limitations of spatial regularization remain unclear

  • The paper demonstrates improvements in segmentation alignment and interpretability evaluations. However, it remains unclear whether these improvements primarily reflect better spatial localization of existing concepts or improvements in the learned representations themselves (e.g., novel concepts, finer-grained concepts, or fewer duplicated concepts.) . If the main benefit is improved activation maps, simpler post-hoc approaches, such as smoothing or segmentation-based refinement, might achieve a similar goal without modifying SAE training. The value of incorporating spatial structure directly into SAE learning could be better established.
  • SAEs can capture diverse visual concepts, including textures, patterns, and abstract semantic properties that do not necessarily exhibit spatial smoothness. The proposed spatial prior may favor spatially coherent concepts at the expense of others. An analysis of which concepts are better captured, newly discovered, or potentially lost compared with conventional SAEs would clarify the scope and limitations of this inductive bias

2. The semantic significance of the smooth and innovation decomposition remains unclear.

  • The ablation studies and spectral analyses support the effectiveness of the two-stream formulation and demonstrate differences in spatial frequency characteristics. However, they provide limited insight into the semantic information represented by each field. A more direct analysis of the concepts captured by the smooth and innovation fields would strengthen the empirical support for MF-LRH.
Questions for authors
  1. However, what types of visual concepts are captured by the smooth and innovation fields, respectively? Could the authors provide concrete examples showing whether they capture complementary semantic information in real-world representations?

  2. How does SPATIAL-SAE change the types of concepts learned compared with conventional SAEs? Are there concepts that become newly identifiable or better represented, as well as concepts that become less interpretable due to the spatial prior?

  3. Could the authors summarize the absolute improvements in MCC, support F1, and polysemanticity for each sparsifier in Table 1? Currently there is only win rate are reported, and this additional info would help assess the magnitude of the gains

  4. The synthetic benchmark varies the geometry and support of smooth and innovation fields. How representative are these assumptions of real vision-model representations, and would the proposed method remain effective when the underlying concepts exhibit different or weaker spatial dependencies?

7 comments
AuthorsOpenReview ↗

Rebuttal 1/2

We thank the reviewer for the thoughtful questions about what the prior changes. We answer in order of importance.

W1.a: [...] it remains unclear whether these improvements primarily reflect better spatial localization of existing concepts or improvements in the learned representations themselves (e.g., novel concepts, finer-grained concepts, or fewer duplicated concepts). If the main benefit is improved activation maps, simpler post-hoc approaches, such as smoothing or segmentation-based refinement, might achieve a similar goal without modifying SAE training.

Q3: Could the authors summarize the absolute improvements in MCC, support F1, and polysemanticity for each sparsifier in Table 1? Currently there is only win rate are reported, and this additional info would help assess the magnitude of the gains.

TLDR: The learned representations themselves improve, and post-hoc smoothing cannot reproduce the gain. On real data, smoothing a baseline SAE reaches Spatial-SAE's semantic smoothness (region coherence), yet its concept quality (Normalized F1) is lower than Spatial-SAE's for all four sparsifiers. On synthetic data, where the true concepts are known, this gain is evident: Spatial-SAE directly improves the quality of the learned concepts as measured by their alignment with true concepts (MCC), which no post-processing of activation maps can change.

Smoothing control on real data (the same table as in our reply to NVWv W1). We smooth each baseline's codes with a Gaussian kernel of one patch (the best of six smoothing kernels for concept quality), re-sparsify every patch to its original ℓ0\ell_0, and decode with the baseline's own dictionary. Cells are ΔSpatial−smoothed\Delta_{\text{Spatial} - \text{smoothed}}, in bold where Spatial-SAE is better:

SparsifierConcept quality Δ\Delta (Normalized F1)Smoothness Δ\Delta (region coherence)Faithfulness Δ\Delta (R2R^2)
TopK+5.0+3.4-1.7
BatchTopK+3.4+2.4-1.6
ReLU+4.5-0.9-0.5
JumpReLU+4.9-0.3+2.3

Synthetic data, with the absolute gains of Table 1 (Q3). Spatial-SAE improves all three identification metrics for every sparsifier: clearly for TopK, BatchTopK, and JumpReLU, and slightly for ReLU. For those three, relative to the baseline, alignment with the true concepts (MCC) rises by 10 to 21%, support recovery (F1) rises by 25 to 83%, and polysemanticity falls by 21 to 51%. Cells are the mean ΔSpatial−baseline\Delta_{\text{Spatial} - \text{baseline}} over the 300 matched pairs of Table 1, with the change relative to the baseline in parentheses, in bold where Spatial-SAE is better:

SparsifierMCC Δ\Delta ↑\uparrowPoly. Δ\Delta ↓\downarrowF1 Δ\Delta ↑\uparrowR2R^2 Δ\Delta ↑\uparrow
TopK+0.13 (+17%)-0.20 (-51%)+0.05 (+25%)-0.01 (-2%)
BatchTopK+0.14 (+21%)-0.19 (-36%)+0.09 (+83%)-0.01 (-3%)
ReLU+0.02 (+2%)-0.02 (-5%)+0.01 (+4%)-0.02 (-12%)
JumpReLU+0.07 (+10%)-0.10 (-21%)+0.05 (+30%)+0.09 (+26%)

W1.b / Q2: [...] The proposed spatial prior may favor spatially coherent concepts at the expense of others. An analysis of which concepts are better captured, newly discovered, or potentially lost compared with conventional SAEs would clarify the scope and limitations of this inductive bias.

TLDR: Concepts that span many patches gain the most. For TopK, gains do not depend on object size; for the other sparsifiers, small, thin objects can lose.

We compared each class's best-feature F1 between baseline and Spatial-SAE (the 84 ADE20K classes with at least 10 held-out images):

  • TopK: 71% of classes improve (mean +3.1pp), independently of class size. Gains include small objects as well as regions; the clearest losses are small rigid objects.
  • BatchTopK, ReLU, and JumpReLU: 49 to 55% of classes improve, and the change grows with class area. Large regions gain most, and small objects lose.

With TopK, small classes gain on average (+2.7pp), so the loss is not inherent to the prior. Under MF-LRH, small objects should be captured by innovation features, but in our trained models these act mainly as a local residual at object boundaries (our reply to W2.a), so for the other three sparsifiers no feature matches small objects as well as the baseline's best one. We will discuss this in the limitations.

AuthorsOpenReview ↗

Rebuttal 2/2

W2.a / Q1: [...] a more direct analysis of the concepts captured by the smooth and innovation fields would strengthen the empirical support for MF-LRH. [...] What types of visual concepts are captured by the smooth and innovation fields, respectively? Could the authors provide concrete examples showing whether they capture complementary semantic information?

TLDR: The smooth field carries the concepts; the innovation field carries the residual: fine local detail and noise. Reconstructing from the smooth field alone keeps 94 to 99% of the concept information (linear-probe segmentation F1) of the full reconstruction. The innovation field alone keeps little, and adding it helps mainly at object boundaries, so the two fields are complementary.

In particular, we applied the segmentation probe from our reply to NVWv Q1 to each stream's reconstruction on its own:

  • Concepts sit in the smooth stream: its reconstruction recovers 94 to 99% of the full reconstruction's segmentation F1, while the innovation stream alone recovers little (F1 8-16 vs 41-45).
  • The innovation stream adds local detail: adding it to the smooth stream raises F1 more on boundary patches than inside objects, for every sparsifier and seed (+0.4 to +3.1pp vs 0.0 to +0.8pp), and it specializes in high frequencies (Figure 6c, dotted), the band that contains boundary detail (our reply to NVWv W2).

Furthermore, this split is what signal processing predicts. The innovation field specializes in high spatial frequencies (Figure 6c), the band with the lowest signal-to-noise ratio: natural signals concentrate their energy at low frequencies while noise spreads across all of them [1, 2], also on graphs [3].

We will revise Section 1 and Appendix B.3 to describe the innovation field this way, and add a gallery of randomly sampled smooth and innovation stream features with their descriptions to the appendix.

[1] Donoho, David L., and Iain M. Johnstone. "Ideal spatial adaptation by wavelet shrinkage." Biometrika, 1994.

[2] Wiener, Norbert. "Extrapolation, interpolation, and smoothing of stationary time series." MIT Press, 1949.

[3] Shuman, David I., et al. "The emerging field of signal processing on graphs." IEEE Signal Processing Magazine, 2013.


Q4: [...] How representative are these assumptions of real vision-model representations, and would the proposed method remain effective when the underlying concepts exhibit different or weaker spatial dependencies?

TLDR: The synthetic data reproduces the spatial dependence of real activations, and the gain persists when that dependence is weaker. Our generator matches the spatial dependence measured in DINOv2 (the decay of patch similarity with distance), and Spatial-SAE's gain in concept recovery (support F1) over the baseline stays positive for every sparsifier as the dependence weakens.

In particular, Figure 2c shows the offset-similarity kernels of real DINOv2 activations, and Figure 2d shows that our generator reproduces their decay with distance, while an independent (LRH) model does not. To test weaker spatial dependence, we lower τ\tau, the diffusion strength that sets how far each smooth concept spreads over the grid, from 64 to 4. With a smaller τ\tau each concept covers fewer patches, so neighboring patches share less of their code. Cells are ΔSpatial−baseline\Delta_{\text{Spatial} - \text{baseline}} in F1 (pp), in bold where Spatial-SAE is better:

SparsifierStrong dependence (τ=64\tau = 64)Weak dependence (τ=4\tau = 4)
TopK+4.5+4.3
BatchTopK+9.2+8.1
ReLU+0.4+0.6
JumpReLU+6.5+2.7

Thank you for the detailed rebuttal and additional experiments. I appreciate the authors' efforts,

However, my concern about the limitations of the spatial prior remains. The current evaluations mainly focus on segmentation-based concept alignment, which naturally favors spatially localized concepts. While the additional class-wise analysis clarifies the trade-offs across object sizes, it does not address whether spatial regularization comes at the cost of learning other types of visual concepts, such as textures or abstract attributes (e.g., emotions, actions), that may not exhibit the same spatial structure. Comparing image-level concept identification performance against standard SAEs on classification datasets could help clarify this potential trade-off.

Reviewer_yKo6OpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper identifies that standard tokenwise SAEs in vision models implicitly impose a patch-independent prior on their latent codes and argues that this is inappropriate for natural images and the spatial dependencies present in vision-model activations. To address this, they propose a decomposed representation consisting of a smooth field that captures spatially persistent information and an innovation field that captures local deviations from this information. In their approach, the smooth field is regularized to promote similar representations across neighboring tokens, while the innovation field is regularized to promote contrast between neighboring patches. This decomposition allows for coarser concepts that span multiple patches to be learned in the smooth field, while local details and deviations can be represented by the innovation field. They evaluate this method both on synthetic data and natural images, finding improved concept recovery and semantic alignment, respectively, at the cost of reduced reconstruction fidelity on real-image features.

Critical strengths

S1. Relieving the tokenwise SAE objective of its implicit patch-independence prior to allow for modelling concepts that span neighboring patches is a great insight, and modelling this behavior with a smooth and an innovation field is elegant.

S2. The proposed MF-LRH improves concept recovery on synthetic data and produces more concept-aligned features on real DINOv2 activations, suggesting that the approach is useful beyond synthetic examples.

S3. The paper is very well written, conveying a clear narrative, the minimal necessary mathematical abstraction and intuitive explanations of every step. Reading the paper was a real pleasure.

Critical weaknesses

W1. Appendix C.1 states that the decomposed models receive separate smooth and innovation streams, whereas the baseline receives only their mixture. To me this reads as if Table 1 presents results using oracle decomposition access, which would be an unfair comparison. Please clarify the exact inputs and report recovery when Spatial-SAE must infer both fields from H alone, as in Section 4.2.

W2. The paper shows that the reconstruction gap is larger at high spatial frequencies, but it remains unclear what information is lost and whether this affects downstream tasks. The overall reconstruction drop is also substantial (roughly 11–18 R² percentage points at the selected sparsity budget). For example, does Spatial-SAE preserve coarse semantic structure while losing finer details such as object boundaries or small structures? One possible experiment would be to compare segmentation performance using original activations, baseline-SAE reconstructions, and Spatial-SAE reconstructions. This could be evaluated both by transferring a segmentation head trained on original activations to the reconstructed features, and by training separate heads directly on each reconstruction. Evaluating the resulting segmentation masks at different spatial scales, together with boundary-sensitive metrics such as Boundary IoU, could help determine whether the reconstruction loss primarily affects fine-grained spatial information or also removes useful semantic information.

W3. The concept F1 evaluation only considers images containing the target class, meaning that false activations on images without that class are ignored. This could overestimate how specific individual features are to a concept. Evaluating class-best features on the full dataset, including class-absent images, with feature selection performed on a separate validation set, would provide a stronger measure of concept specificity.

W4. The ablations show that the two-stream decomposition and spatial regularization improve performance, but they do not isolate the contribution of the repulsive regularization on the innovation field. Would a two-stream model with smooth-field regularization but no spatial regularization on the innovation field perform similarly? This comparison would help establish whether the repulsive prior is actually necessary or whether a separate sparse residual stream is sufficient.

Questions for authors

Q1. Figure 2: It is not clear what the different colors encode in (b). The legend explains the hatching but not the color mapping.

Q2. Algorithm 1: E_S and E_R are not explained when the algorithm is introduced. They are presumably the smooth and innovation encoders, as defined later in Appendix D.1. Could they be defined before their first use?

Q3. 5.1 Main experimental setting. What is DGP? Presumably data-generating process? Please define the abbreviation at first use.

Q4. Appendix D describes training JumpReLU with an L1 penalty, whereas the original JumpReLU method uses direct L0 regularization. Why was this changed? Do the reported gains persist when both the baseline and Spatial-SAE are trained using the original L0 objective at matched sparsity?

4 comments
AuthorsOpenReview ↗

Rebuttal 1/2

We thank the reviewer for the detailed, constructive review and the kind words about the writing. We answer in order of importance.

W2: The paper shows that the reconstruction gap is larger at high spatial frequencies, but it remains unclear what information is lost and whether this affects downstream tasks. [...] does Spatial-SAE preserve coarse semantic structure while losing finer details such as object boundaries or small structures?

TLDR: Coarse semantics are kept and improve inside objects for every sparsifier. Boundaries and small classes are preserved for TopK and BatchTopK; ReLU and JumpReLU lose some accuracy there, most on ReLU's small classes.

We followed the suggested design: a segmentation probe trained on raw activations and transferred unchanged to each reconstruction (cells are baseline -> Spatial-SAE, in bold where Spatial-SAE is better):

SparsifierF1 (all)mIoUBoundary F1Interior F1Small-class F1
TopK43.9 -> 45.024.4 -> 25.743.3 -> 43.639.3 -> 43.435.0 -> 35.0
BatchTopK43.8 -> 44.424.6 -> 24.243.2 -> 43.039.1 -> 43.335.0 -> 34.3
ReLU42.2 -> 41.122.7 -> 20.341.8 -> 39.937.2 -> 39.233.4 -> 29.7
JumpReLU42.0 -> 42.022.7 -> 22.441.7 -> 40.537.0 -> 41.033.1 -> 31.3

The same probe on raw activations: 45.7 / 26.1 / 45.0 / 41.6 / 37.4. Boundary patch: its label set differs from at least one 4-neighbor (54% of patches). Small classes: below the median training-set patch frequency. Patch-level on the 16×1616 \times 16 grid, not pixel-level Boundary IoU.

A probe retrained on each reconstruction gives the same picture (Spatial-SAE minus baseline F1: +1.3 / +0.9 / -0.3 / +0.5 for TopK / BatchTopK / ReLU / JumpReLU). Across spatial scales, the band Spatial-SAE reconstructs least (structure finer than about 5-6 patches) holds little concept information: removing it from the raw activations lowers boundary F1 by 3% and small-class F1 by 5% (relative; our reply to NVWv W2), about what JumpReLU loses here, while ReLU loses more. We will add these analyses to the appendix and the ReLU/JumpReLU cost to the limitations.

W1: Appendix C.1 states that the decomposed models receive separate smooth and innovation streams, whereas the baseline receives only their mixture. To me this reads as if Table 1 presents results using oracle decomposition access, which would be an unfair comparison. Please clarify the exact inputs and report recovery when Spatial-SAE must infer both fields from H alone, as in Section 4.2.

TLDR: The prior improves recovery with access held fixed (Figure 5 in Section 5.1). The synthetic study isolates the effect of the decomposition and of the prior. Inferring both fields from H\mathbf{H} alone is the real-data setting, where both interventions still help.

We compare three models (Appendix C.1): (1) B, a standard single-stream SAE on H\mathbf{H}, (2) D, a two-stream SAE given the true smooth and innovation fields, without the prior, and (3) DT, the same two-stream SAE with the prior (Spatial-SAE). B -> D isolates the decomposition, and D -> DT the prior. Table 1 reports DT vs B, Figure 5 reports DT vs D, and Table 2 and Figures 9-10 report all comparisons. With access held fixed, the prior alone raises F1 across all smoothness and innovation settings, by up to 6.9pp (Figure 5), and it wins every MCC, F1, and polysemanticity comparison against D for TopK, BatchTopK, and JumpReLU (no effect for ReLU).

Inferring the fields directly from H\mathbf{H} in these comparisons would confound the effect of each intervention with estimation error, so the synthetic study gives D and DT the true fields. The test the reviewer asks for is the real-data setting (Section 5.2), where Spatial-SAE only sees H\mathbf{H} and we measure alignment with annotated classes. There both interventions still help (ADE20K, TopK, static estimator in Appendix D.4): adding the prior raises Normalized F1 by a median +6.5pp, in all 14 matched pairs (Table 5), and the decomposition adds +6.1pp over a single-stream model with the same prior, also in all 14 (Table 4). Across sparsifiers, Spatial-SAE improves Normalized F1 over the baselines by +4.3 to +7.9pp.

AuthorsOpenReview ↗

Rebuttal 2/2

W3: The concept F1 evaluation only considers images containing the target class, meaning that false activations on images without that class are ignored. This could overestimate how specific individual features are to a concept.

TLDR: Our implementation already evaluates F1 on the full held-out pool, including class-absent images, as the reviewer suggests. Every number in the paper uses it; only Equation 10 misdescribes our code.

True and false positives are counted over all held-out patches (1448 images ×\times 256 patches), so a feature firing on an image without the class counts as a false positive. Restricted to class-present images, Spatial-SAE's Normalized F1 lead is +5.5 / +2.7 / +2.3 / +3.6 (TopK / BatchTopK / ReLU / JumpReLU), against +7.0 / +4.6 / +5.0 / +4.2 on the full pool (both with baselines retrained with the paper's configuration). Spatial-SAE's class-best features therefore also fire less on images without their class. We will correct Equation 10.

W4: The ablations show that the two-stream decomposition and spatial regularization improve performance, but they do not isolate the contribution of the repulsive regularization on the innovation field. [...] This comparison would help establish whether the repulsive prior is actually necessary or whether a separate sparse residual stream is sufficient.

TLDR: The repulsive prior is necessary. Without it, the smooth codes collapse to zero (or a spatial constant), which minimizes their spatial penalty, and the model degrades to a baseline SAE that only uses the innovation stream's share of the budget.

The smooth-field prior penalizes spatial variation in the smooth codes, and zero or spatially constant codes cost nothing under it. When both fields are inferred from H\mathbf{H}, the innovation stream sees the target patch and, without the repulsive term, is not regularized, so it can reconstruct the activation on its own. The cheapest solution is then to switch the smooth stream off: what remains is a standard per-patch SAE on H\mathbf{H} with only the innovation stream's atoms, a weaker version of the baseline. We observed this collapse in preliminary experiments. The repulsive term prevents it by penalizing spatially smooth content in the innovation stream, which makes that content cheaper to carry in the smooth stream. We will state this role explicitly in Section 3 and in the ablation discussion.

Q4: Appendix D describes training JumpReLU with an L1 penalty, whereas the original JumpReLU method uses direct L0 regularization. Why was this changed?

We use the JumpReLU SAE of the overcomplete library, which implements the activation and learned thresholds of [1] and is trained with the library's ℓ1\ell_1 penalty. Baseline and Spatial-SAE share this objective, and a simplified version of the target-ℓ0\ell_0 controller of [2] adjusts its coefficient so that every model sits at E[ℓ0]=16\mathbb{E}[\ell_0] = 16. We will state each sparsifier's training objective in the appendix.

Q1-Q3: Figure 2: It is not clear what the different colors encode in (b). [...] Algorithm 1: ESE_S and ERE_R are not explained when the algorithm is introduced. [...] What is DGP?

Thank you; we will fix all three: add the missing color coding of Figure 2b to its legend (colors denote the layer the activations are extracted from: first, middle, or last), define ESE_S and ERE_R in Algorithm 1 as the smooth and innovation stream encoders, and define "DGP" (data-generating process) at first use.

[1] Rajamanoharan, Senthooran, et al. "Jumping ahead: Improving reconstruction fidelity with JumpReLU sparse autoencoders." arXiv:2407.14435, 2024.

[2] Chanin, David, and Adrià Garriga-Alonso. "SynthSAEBench: Evaluating sparse autoencoders on scalable realistic synthetic data." arXiv:2602.14687, 2026.

Thank you for the detailed rebuttal and additional experiments, and apologies for my late response.

I appreciate the authors' efforts in addressing my concerns. In particular:

  • W1: The clarification of the synthetic setup and the additional references to the controlled comparisons largely address my concerns about evaluation fairness.
  • W2: The new segmentation experiments provide convincing evidence about the downstream impact of the reconstruction loss and clarify which information is affected.
  • W3: The clarification that class-absent images were already included in the evaluation resolves my concern about concept specificity.
  • W4: The explanation of the innovation stream's role and the collapse observed without repulsive regularization provides a plausible justification for this design choice.
  • Q1–Q4: Thank you for the clarifications and proposed revisions.

I also found the concerns raised by the other reviewers about whether the method improves general concept discovery or primarily favors spatially coherent concepts particularly interesting. The additional experiments address the concern that the improvements merely reflect spatial smoothing, while the applicability to less spatially structured concepts remains an open question. I think this distinction and the potential trade-offs across concept types would be worth discussing explicitly in the paper.

Overall, I find my main concerns sufficiently addressed and appreciate the substantial additional experimental evidence. I maintain my original rating of 4 (Clear Acceptance).

Thank you for your response, we are glad that our rebuttal addressed your concerns!

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

In this paper, the authors make an interesting observation that many existing Sparse Autoencoders (SAEs) impose a strong prior that treats concepts across patches as independent, which does not reflect the nature of natural images. To address this limitation, they propose the Markov-Field Linear Representation Hypothesis (MF-LRH), which replaces this prior with one that better matches vision-model activations. The proposed method achieves competitive results in the experiments presented in the paper.

Critical strengths
  1. In general, the paper is well written and easy to follow. The proposed method is well motivated by a clear observation of a practical problem with current baselines. The idea of addressing this problem by replacing the mismatched factorised prior with an MRF over patch codes is conceptually simple and intuitive.

  2. In terms of evaluation and experiments, the paper evaluates various sparsifiers, uses synthetic ground truth, and tests on two real-world datasets, transferring the configuration from ADE20K to COCO-Stuff without re-tuning. The ablation studies are sufficient, covering separate decomposition, trace strength, and estimator choices, including a single-stream Laplacian-only control. Qualitative examples are also well presented.

Critical weaknesses
  1. All evaluation metrics (e.g., patch-mask F1 and region coherence) reward spatially smooth, segment-shaped maps. Since the smooth-stream encoder never sees the target patch, its code is effectively a learned neighborhood average.

  2. The claim that the lost high-frequency variance corresponds to information where “concepts do not live” seems unjustified and is not empirically tested. High-frequency components in DINOv2 features may carry useful information, such as boundaries, positional cues, and part-level details.

This raises the question of whether the gains of the proposed method are from exploiting smoothness that resembles human segmentation, rather than from producing more faithful representations of the model’s features.

Questions for authors

Apart from my previous comments, I have three additional questions:

  1. What is the downstream performance or faithfulness cost of the drop in R2R^2? For example, could you evaluate linear segmentation on ADE20K or k-NN/linear classification using reconstructed activations, comparing the baseline with Spatial-SAE?

  2. How robust are the auto-interpretability results to the choice of rater and feature selection? For example, could you use a second LLM or conduct a small human study, and evaluate randomly selected features rather than only the best-performing features for each class?

  3. Do the conclusions generalise to other backbones?

5 comments
AuthorsOpenReview ↗

Rebuttal 1/2

We thank the reviewer for the careful review. We answer in order of importance.

W1: All evaluation metrics (e.g., patch-mask F1 and region coherence) reward spatially smooth, segment-shaped maps. Since the smooth-stream encoder never sees the target patch, its code is effectively a learned neighborhood average [...] This raises the question of whether the gains of the proposed method are from exploiting smoothness that resembles human segmentation, rather than from producing more faithful representations of the model's features.

TLDR: Spatial-SAE's interpretability gains cannot be reached by smoothing an SAE. We smooth SAEs with the kernel that maximizes their concept quality (Normalized F1) across sparsifiers. This also brings them close to Spatial-SAE's faithfulness (R2R^2) and semantic smoothness (region coherence), yet Spatial-SAE's concept quality stays 3.4 to 5.0pp higher for all four sparsifiers.

In particular, we smooth each baseline's codes with a Gaussian kernel of one patch (the best of six smoothing kernels for concept quality), re-sparsify every patch to its original ℓ0\ell_0, and decode with the baseline's own dictionary. Cells are ΔSpatial−smoothed\Delta_{\text{Spatial} - \text{smoothed}} in the reported metric, in bold where Spatial-SAE is better:

SparsifierConcept quality Δ\Delta (Normalized F1)Smoothness Δ\Delta (region coherence)Faithfulness Δ\Delta ( R2R^2 )
TopK+5.0+3.4-1.7
BatchTopK+3.4+2.4-1.6
ReLU+4.5-0.9-0.5
JumpReLU+4.9-0.3+2.3

W2: The claim that the lost high-frequency variance corresponds to information where "concepts do not live" seems unjustified and is not empirically tested. High-frequency components in DINOv2 features may carry useful information, such as boundaries, positional cues, and part-level details.

TLDR: The high-frequency band carries little concept information: removing it from DINOv2 activations barely changes segmentation F1. Concept information (linear-probe segmentation F1) goes from 45.7 to 45.2: the high band holds 27% of the activations' energy, yet nearly all concept information sits in the remaining, smooth 73%.

In particular, we remove from the raw DINOv2 activations the graph-frequency band where Spatial-SAE's reconstruction gap concentrates (modes 40-255; Figure 6c). A linear segmentation probe trained on the raw activations is then applied to what remains, and to the band alone:

ActivationsF1mIoU
Raw45.726.1
High band removed45.225.6
High band alone22.67.9

Finally, the reviewer is right that the band is not empty: it carries some boundary and small-object detail (e.g., removing it lowers boundary F1 by 3% (relative)). But the high-frequency band has the lowest signal-to-noise ratio: natural signals concentrate their energy at low frequencies while noise spreads across all of them, which is why classical denoising attenuates the high band [1, 2], and the same holds for signals on graphs [3]. The high band is therefore the cheapest place to give up a little faithfulness for interpretability. "Where concepts do not live" was too strong; we will replace it with this measured statement and add the analysis to the paper.

[1] Donoho, David L., and Iain M. Johnstone. "Ideal spatial adaptation by wavelet shrinkage." Biometrika, 1994.

[2] Wiener, Norbert. "Extrapolation, interpolation, and smoothing of stationary time series." MIT Press, 1949.

[3] Shuman, David I., et al. "The emerging field of signal processing on graphs." IEEE Signal Processing Magazine, 2013.

AuthorsOpenReview ↗

Rebuttal 2/2

Q1: What is the downstream performance or faithfulness cost of the drop in R2? For example, could you evaluate linear segmentation on ADE20K or k-NN/linear classification using reconstructed activations, comparing the baseline with Spatial-SAE?

TLDR: The drop in R2R^2 does not reduce downstream performance. We ran the suggested experiment. Spatial-SAE's reconstructions reach the same downstream performance (linear segmentation F1 on ADE20K) as the baseline's, within about 1pp for every sparsifier, although their faithfulness (R2R^2) is 11 to 18pp lower.

In particular, we train a linear segmentation probe on the raw DINOv2 activations, where it reaches F1 45.7 and mIoU 26.1, and apply it unchanged to held-out reconstructions. Cells are ΔSpatial−baseline\Delta_{\text{Spatial} - \text{baseline}} in the reported metric, in bold where Spatial-SAE is better (seed standard deviation ≤1.2\leq 1.2):

SparsifierDownstream Δ\Delta (F1)Downstream Δ\Delta (mIoU)Faithfulness Δ\Delta ( R2R^2 )
TopK+1.1+1.3-18.0
BatchTopK+0.6-0.4-16.7
ReLU-1.1-2.4-14.7
JumpReLU0.0-0.3-10.8

Finally, the only clear cost is for ReLU, and it is localized on boundaries and small classes (see our reply to yKo6 W2). We will add this experiment to the paper.


Q3: Do the conclusions generalise to other backbones?

TLDR: The gains carry over to a second backbone without any re-tuning, and are mostly larger there. On a supervised ViT-B/16 (AugReg), Spatial-SAE improves concept quality (Normalized F1) by 3.8 to 14.3pp and smoothness (region coherence) by 12.8 to 21.5pp, for every sparsifier on both ADE20K and COCO-Stuff.

In particular, we train the baselines and Spatial-SAE on AugReg activations with exactly the DINOv2 hyperparameters. Cells are ΔSpatial−baseline\Delta_{\text{Spatial} - \text{baseline}} as ADE20K / COCO-Stuff, in bold where Spatial-SAE is better (standard error at most 1.8, 0.2, and 1.7 in the three columns):

SparsifierConcept quality Δ\Delta (Normalized F1)Smoothness Δ\Delta (region coherence)Faithfulness Δ\Delta ( R2R^2 )
TopK+10.0 / +12.4+15.9 / +15.3-19.5 / -16.7
BatchTopK+5.1 / +6.8+13.9 / +12.8-15.7 / -13.6
ReLU+5.9 / +14.3+18.3 / +17.4-1.5 / -16.7
JumpReLU+3.8 / +12.3+19.0 / +21.5-0.6 / -12.3

Q2: How robust are the auto-interpretability results to the choice of rater and feature selection? For example, could you use a second LLM or conduct a small human study, and evaluate randomly selected features rather than only the best-performing features for each class?

TLDR: The results hold with human raters, with a second LLM, and for randomly selected features at matched firing frequency. Six blind human raters prefer Spatial-SAE's features in 91.0% of decided judgments for map quality and in 89.0% for concept consistency. A second LLM (Claude Opus 5.5 in place of Haiku 4.5) keeps Spatial-SAE ahead for every sparsifier.

In particular, six blind raters compared the class-best feature visualizations of baseline and Spatial-SAE for 50 class-sparsifier pairs, without knowing which model produced each and in randomized order. In the table below, cells are the percentage of decided judgments preferring Spatial-SAE, with 95% confidence intervals in brackets, in bold where Spatial-SAE is better:

Human ratersMap qualityConcept consistency
All judgments (6 raters ×\times 50 pairs)91.0% [86.8, 93.9]89.0% [84.2, 92.5]
Majority vote per pair97.9% [88.9, 99.6]92.0% [81.2, 96.8]

Each of the six raters prefers Spatial-SAE on both criteria: 74 to 97% of decided pairs for map quality and 71 to 96% for concept consistency. When two raters both pick a side, they pick the same one in 89% of cases for map quality and 88% for concept consistency.

For feature selection, Opus 5.5 rated 50 random non-class-best features per model, one per firing-frequency quantile (400 visualizations). Cells are ΔSpatial−baseline\Delta_{\text{Spatial} - \text{baseline}} in score (1-10 scale) at matched firing frequency, with 95% confidence intervals in brackets, in bold where Spatial-SAE is better:

Random featuresMap qualityMonosemanticity
Δ\Delta score+0.67 [+0.32, +0.99]+0.84 [+0.41, +1.25]

Randomly selected Spatial-SAE features score higher on both criteria, with confidence intervals that exclude zero.

Dear Reviewer NVWv,

We thank you again for your careful review! As the discussion period is coming to an end (less than a day left), we would greatly appreciate your input on whether or not our responses resolved your concerns, and we are happy to engage with you further if needed.

Since our initial reply to your concerns and during our discussion with Reviewer W6uK, we ran a new experiment that also addresses your W1, namely whether or not the observed gains of Spatial-SAE over the baselines depend on metrics that reward segment shaped maps. Specifically, we trained the baselines and Spatial-SAE on textures (DTD), actions (Stanford 40), and emotions (EmoSet), using the paper's hyperparameters and no retuning. DTD and EmoSet have image labels only, so there Normalized F1 is computed at image level (a feature fires on an image if it fires on any patch), which does not reward spatially smooth maps. We found that Spatial-SAE significantly outperforms the baselines on textures and actions for every sparsifier (Normalized F1 gains of up to 7.7pp and 8.0 to 9.6pp), and on emotions it gains up to 3.3pp and wins or ties for three of four sparsifiers. The full results are in our follow up reply to Reviewer W6uK.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.