acceptodds
Withdrawn from ICLR 2027

Beyond Reconstruction: Preserving and Decomposing Compositional Behavior with Multimodal Sparse Autoencoders

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Sparse autoencoders (SAEs) decompose VLM representations into interpretable features, but recovering objects, attributes, and relations does not establish that an SAE preserves how the model combines them. Reconstruction, sparsity, retrieval, and feature purity therefore provide limited evidence of compositional faithfulness. We identify two requirements for valid multimodal analysis: independently inferred image and text codes to prevent pair-identity leakage, and determinate scores for every selected feature. Without these safeguards, random dictionaries can yield near-perfect retrieval, while fixed product-TopK can select coordinates arbitrarily. We then train a multimodal SAE using Group-Hoyer sparsity, cross-modal decoding, and a variable-size positive-only paired auxiliary. Across matched CLIP, SigLIP, and SigLIP2 experiments, it improves reconstructed SugarCrepe accuracy over MGSAE by 3.7-5.1 percentage points while preserving retrieval geometry and automatically measured shared-concept coverage with 64-256 active features. Exact compositional-margin decompositions identify compact, cross-seed-recoverable circuits whose ablation and activation swapping affect held-out decisions more than energy-matched random controls. Ordered relation binding, however, remains distributed rather than localized to individual symbolic features. These results show that multimodal SAEs can preserve and causally expose compositional computations.

Primary area:
interpretability and explainable AI
Keywords:
Compositionality, Vision-Language Models, Multimodal Sparse Auto Encoders, Mechanistic Interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_bZTYOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
1(−1)
Soundness
2
Read review
Paper summary

The paper evaluates multimodal sparse autoencoders through preservation of compositional decisions rather than reconstruction alone. It identifies pair-conditioned retrieval leakage and ambiguous product-TopK selection, proposes a shared SAE with modality-local inference, and decomposes reconstructed decision margins into feature contributions. Experiments cover three VLM families, sparsity controls, and held-out feature interventions.

Critical strengths
  1. The evaluation critique is concrete and consequential: random dictionaries can achieve near-perfect retrieval when pair identity enters support construction. The distinction between reconstruction quality and decision preservation is also supported by explicit measurements rather than examples alone.
  2. The experimental controls are unusually extensive. Modality-local inference, matched-support comparisons, disjoint circuit discovery and evaluation, and energy-matched random interventions address several obvious alternative explanations. The negative result on symbolic relation binding is informative.
Critical weaknesses
  1. The advantage of the proposed training objective is less clear than the value of the evaluation framework. Appendix H.7 reports that ordinary TopK and reconstruction-only variants preserve decisions better, and H.8 does not isolate Group-Hoyer against ordinary Hoyer. The claimed benefit is therefore a fidelity--shared-feature trade-off, not better preservation alone. A common comparison including behavioral fidelity and shared-concept coverage is needed to establish that trade-off. The theoretical results are useful diagnostics, but mostly elementary identities and bounds.
  2. The causal evidence concerns SAE-reconstructed decisions, as Section 4.7 correctly states. It does not directly establish the backbone's internal compositional mechanism. Cross-seed evidence is also limited to rank recovery: Section 5.5 reports 31% top-32 recovery with decoder cosine 0.22, while causal transfer remains untested. This supports a useful intervention basis more strongly than a stable mechanistic decomposition.
Questions for authors
  1. Under a common support budget and semantic evaluator, does the full objective improve the joint trade-off between decision preservation and shared-concept coverage over ordinary TopK and the simpler cross-decoding variants?
  2. Do circuits matched across seeds retain their held-out intervention effects? If this has not been tested, can the reproducibility claim be restricted explicitly to rank recovery?

Post-rebuttal update: I read the revised PDF and the authors' replies. Please see my comment below for details on the new experiments. Fixing the base-image split leak is a good step, and the cross-seed transfer results on SugarCrepe help. Still, the new ablations show that Group-Hoyer adds very little over plain L1 or Group-L2,1, and the coverage gain over JumpReLU disappears on SigLIP. The paper also needed a major rewrite. I am lowering my score from 3 to 2.

2 comments
AuthorsOpenReview ↗

Response to Reviewer bZTY

Thank you for the careful review. Both questions led to new experiments, and we revised the manuscript to reflect what those experiments establish.

W1 / Q1. How does the full objective change the fidelity–coverage trade-off?

We compared methods at a common SigLIP2 support target of 256 features (seeds 17–19), using the same semantic evaluator and decision-preservation protocol. The variable-support baselines were matched on mean activity (Table 2, Figure 2, §5.3).

ModelStrict shared units ↑SugarCrepe DP ↑
Ordinary TopK39.00.909
AuxK + cross-decoding (cumulative ladder step)56.30.909
Shared latent masking (cumulative ladder step)72.70.886
Ours (selected configuration)77.70.892
BatchTopK / JumpReLU / Matryoshka43.0–52.30.897–0.908
MGSAE (K=256K=256)70.70.862

The selected configuration recovers the most strict shared units in this comparison. Ordinary TopK and the AuxK + cross-decoding variant preserve more SugarCrepe decisions. In the objective ladder, AuxK leaves coverage at 39.0 units, and adding cross-decoding raises it to 56.3 with decision preservation (DP) unchanged relative to ordinary TopK at the reported precision. Adding shared latent masking to the preceding ladder variant increases coverage from 59.7 to 72.7 units and reduces SugarCrepe DP from 0.906 to 0.886 (Figure 2, Appendix N, Table 52). The three added reconstruction-only baselines preserve 0.5–2.5 percentage points more decisions than our selected model across SugarCrepe and SugarCrepe++. We revised the paper to describe this fidelity–coverage trade-off explicitly.

Group-Hoyer ablation. We replaced Group-Hoyer with plain Hoyer, ℓ1\ell_1, and Group-L2,1L_{2,1}, each calibrated to the same initial loss contribution (§4.4, Appendix M).

PenaltyFVU ↓SugarCrepe DP ↑Strict shared units ↑
Group-Hoyer0.11210.89277.7
Plain Hoyer0.11300.88973.0
ℓ1\ell_10.11670.88579.3
Group-L2,1L_{2,1}0.11650.88581.7

At this operating point, Group-Hoyer modestly improves reconstruction and decision preservation over plain Hoyer. The largest increases in coverage along the objective ladder occur when cross-decoding and shared latent masking are added. We also clarified in §3.1 that the elementary identities and bounds serve as diagnostics that motivate our measurements.

W2 / Q2. Do matched circuits retain their held-out intervention effects?

Yes, for activation exchange on SugarCrepe. We score each edit on the frozen model's final pooled caption embedding by restoring the reconstruction residual, xedit=x^edit+(x−x^)x_{\mathrm{edit}}=\hat{x}_{\mathrm{edit}}+(x-\hat{x}). The resulting change to the dense embedding is exactly the change in the SAE reconstruction induced by editing the selected activations (§4.7).

During the rerun, we found that the original example-level split allowed base images to appear in both discovery and evaluation. The updated causal results in Table 3 and §5.5 use an image-disjoint split. We mapped the seed-17 top-32 circuit into seeds 18–21 by one-to-one Hungarian matching, separately using decoder cosine similarity and activation correlation, without reranking in the target seeds.

SugarCrepe activation exchangeExcess accuracy drop in percentage points, mean [seed range]Seeds with 95% interval above zero
Each seed's own top-32 circuit4.80 [3.12, 9.85]5/5
Mapped from seed 17, decoder matching3.41 [1.17, 6.98]4/4
Mapped from seed 17, activation matching3.49 [2.04, 5.99]4/4

The excess drop is measured relative to eight activity-matched random feature sets. All eight transfer sign-flip tests remain significant after Holm correction. Under decoder matching, an average of 32.0% of mapped features on SugarCrepe and 25.8% on ARO also appear in the target seed's own top 32.

For ablation of each seed's own circuit, the paired 95% interval was above zero in three of five seeds. ARO exchange results depended on how the circuit was selected. We therefore scope the cross-seed claim to activation exchange on SugarCrepe at the final pooled caption embedding. The encoder's internal compositional mechanism remains an open question.

We would be grateful if you could look at the revised PDF, particularly Table 2, Figure 2, and Table 3, and let us know whether it addresses your concerns.

Reviewer_bZTYOpenReview ↗

Review

Thank you for the detailed response and the new experiments. The rerun using the image-disjoint split answers my question about cross-seed transfer on SugarCrepe. Activation exchange transfers to all four target seeds under both matching rules. However, the ablation evidence is relatively weak, and the ARO results depend heavily on how the circuit is selected and matched.

My main concern remains the complete training objective. Table 2 shows a clear trade-off on SigLIP2: the proposed model finds more strict shared units, while the reconstruction-only baselines better preserve the model's original decisions. Table 44 shows that the fidelity gain brought by Group-Hoyer is very limited; the vast majority of the coverage gain in Table 52 actually comes from cross-decoding and latent masking. Since this experiment uses different hyperparameter weights from the main comparison, I am still unclear about what this complete recipe adds in terms of substantial extra benefits over the simpler "cross-decoding plus masking" variant.

Overall, the new ablations and baselines indicate that the proposed regularizer and training recipe do not offer clear advantages over simpler regularizers or standard SAE baselines. Moreover, to support these claims, the paper has essentially undergone a near-complete rewrite. Combined with the missing literature on causal interventions in VLM SAEs pointed out by Reviewer fMsY, I believe this paper still requires a major overhaul and re-evaluation. Therefore, I am lowering my score from 3 to 2 .

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

This paper studies sparse autoencoders (SAEs) that represent image and caption embeddings from vision-language encoders, such as CLIP and SigLIP, as sparse, interpretable features shared across both. It argues that reconstruction quality, retrieval scores, and recognizable features do not establish whether an SAE preserves distinctions between different combinations of concepts, such as "a person riding a horse" and "a horse riding a person." It also shows that using the matching caption to select an image’s features can produce nearly perfect retrieval scores even with random dictionaries.

The proposed method encodes images and captions independently using a shared dictionary and encourages paired inputs to activate common features. On benchmarks with correct and subtly altered captions, it preserves the original encoder’s choices better than the closest comparison method, MGSAE, although MGSAE reconstructs embeddings more accurately and a simple SAE trained only for reconstruction can preserve these choices as well or better.

Critical strengths

Experiments with random dictionaries show that nearly perfect image and text matching scores can result from using known pairings during feature selection. This exposes a concrete evaluation flaw that could otherwise be mistaken for evidence that an SAE has learned meaningful shared features.

Critical weaknesses
  1. Experiments are underdeveloped. Critically, the method is compared only with the simplest TopK baseline and an unpublished MGSAE method (Kaushik et al., 2026). Most of the results are given in the Appendix, which seems unacceptable; the main text reports only two tables.
  2. The paper’s coverage of related work is limited, demonstrating a lack of sufficient scholarship effort. It omits foundational studies on applying SAEs to CLIP (see [a,b] and papers citing them). Furthermore, [c] introduced finding small sets of SAE features that drive a model’s choice between contrastive inputs and testing them by ablation, which is essentially what this submission does in a simpler, single-layer setting without acknowledging it.
    [a] "Discover-then-Name: Task-Agnostic Concept Bottlenecks via Automated Concept Discovery" ECCV 2024
    [b] "Interpreting CLIP with Hierarchical Sparse Autoencoders" ICML 2025
    [c] "Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models" ICLR 2025

Also:

  • The abstract is unreadable; consider the following unclear phrases appearing one by one: "pair-identity leakage", "determinate scores", "safeguards", "cross-modal decoding", "variable-size positive-only paired auxiliary", "SugarCrepe", "MGSAE", "preserving retrieval geometry", "Exact compositional-margin decompositions identify compact, cross-seed-recoverable circuits whose ablation and activation swapping affect held-out decisions more than energy-matched random controls", "Ordered relation binding".
  • "0.9997 ± 0.0001", "0.9996 ± 0.0000" is hard to parse in-text, but also Table 1 is redundant if they are given in-text (choose either or)
  • CLIP and SigLIP are vision-language encoders and not vision-language models (e.g. LLaVA, GPT, Qwen-VL).
  • The main text relies too heavily on the appendix; the excessive references to it undermine the quality of the writing and overall presentation.
Questions for authors
  1. Can you compare your method with BatchTopK, JumpReLU, and Matryoshka SAEs?
  2. What is "R@1"?
  3. It would be beneficial to show and discuss some informative qualitative examples in the paper.
  4. Can you rephrase these three sentences stating your contributions in plain language so that they are human-comprehensible? It's too much unclear jargon.

We formalize modality-local inference and selector-score identifiability, demonstrate identity leakage through pair-constructed supports, and characterize the indeterminate regime of fixed product-TopK selection.

We combine independent inference, modality-specific centering, a shared dictionary, structured sparsity, cross-decoding, and a positive-only paired auxiliary, improving compositional preservation over matched MGSAE across three VLM backbones.

Margin decomposition, crossseed matching, ablation, and activation exchange reveal reproducible causal circuits, while showing that ordered relation binding remains distributed.

  1. Is the sentence "The manuscript text was written by the authors." correct?
3 comments
AuthorsOpenReview ↗

Response to Reviewer fMsY

Thank you for the detailed review. We agree with the main concerns and have revised the paper substantially. Revisions are highlighted in blue in the revised PDF.

W1 / Q1. Comparison with BatchTopK, JumpReLU, and Matryoshka; reliance on the appendix.

We added all three baselines at a common SigLIP2 support target of 256 features (seeds 17–19). Each of the three added baselines uses modality-specific centering and a shared encoder and dictionary, and their variable support is matched on mean activity. The comparison is now in the main text (Table 2, Figure 2, §5.3). In the table, FVU is the fraction of variance unexplained, DP is decision preservation, the fraction of tie-free examples on which the reconstruction and the frozen encoder make the same caption choice, and SC++ is SugarCrepe++.

MethodFVU ↓SugarCrepe DP ↑SC++ DP ↑Strict shared units ↑
Ours (selected configuration)0.1120.8920.86477.7
MGSAE (K=256K=256)0.0790.8620.83470.7
Ordinary TopK0.0820.9090.88639.0
BatchTopK0.0790.9080.88943.0
JumpReLU0.0820.8970.88252.3
Matryoshka-BatchTopK0.0700.9040.88843.7

BatchTopK, JumpReLU, and Matryoshka-BatchTopK reconstruct better and preserve 0.5–2.5 percentage points more of the frozen encoder's decisions across SugarCrepe and SugarCrepe++. At the same support target, our selected configuration recovers 77.7 strict shared image–text units versus 43.0–52.3 for those three baselines, or about 1.5–1.8× as many. We now present this as a fidelity–coverage trade-off. We retain MGSAE as the closest multimodal comparison.

The main text now also contains the objective ablation (Figure 2), the Group-Hoyer ablation (§4.4), a qualitative example (Figure 3), and the causal results (Table 3). We reduced appendix references throughout.

W2. Related work. We added Discover-then-Name and Interpreting CLIP with Hierarchical SAEs to §2. We also acknowledge Sparse Feature Circuits explicitly: our circuit analysis follows that work. The differences are that we study a single pooled embedding of a dual encoder, where the caption-side margin decomposition is exact rather than approximate, and that we add held-out tests against activity-matched controls and cross-seed transfer. We thank the reviewer for these references.

Abstract and presentation. We simplified the abstract and rewrote the contribution list in plain language. We removed the redundant retrieval table and rounded the in-text numbers (§5.2). We now describe CLIP and the SigLIP family as frozen image–text dual encoders (§§1 and 3.1).

Q2. What is R@1? Recall@1: the fraction of queries whose true pair is ranked first. It is now defined at first use (§5.2).

Q3. Qualitative examples. Figure 3 shows a held-out SugarCrepe example chosen by a stated rule: "Three teddy bears laying in bed under the covers" vs. "A single teddy bear laying in bed under the covers." Exchanging the activations of the 32 selected features between the two captions moves the model's margin, with the reconstruction residual restored, from +0.044 to −0.004, flipping its decision. The dominant feature carries the text label "three". Its image-side label is unclear, so the example illustrates an intervention effect rather than a clean shared concept.

Q4. Contributions in plain language. In plain language, the revised contributions are as follows:

  1. Reconstruction, retrieval, and feature labels can look excellent while an SAE changes the model's compositional decisions. A valid evaluation needs two checks: encoding image and text independently, and checking whether a tie at the cutoff makes the selected features ambiguous. When the known matching caption helps select an image's features, even random dictionaries can achieve near-perfect retrieval.
  2. At the selected SigLIP2 operating point, our SAE recovers more shared image–text concepts than the three added reconstruction-only baselines. In a separate exact-support sweep on SigLIP2, it improves reconstructed SugarCrepe accuracy over MGSAE. We characterize the trade-off rather than claim a universal winner.
  3. Exchanging small, automatically selected feature sets between a correct caption and its foil changes the frozen model's decisions on held-out images, and on SugarCrepe this effect transfers across independently trained SAEs.
  4. The codes carry information about relation categories, but pair-order decoding stays near chance and very few individual features distinguish reversed roles under our tests.

Continued in another comment.

Q5. AI-use statement. Yes. The authors wrote the manuscript and used generative AI to polish grammar, phrasing, and concision, as indicated in our OpenReview disclosure. The revised AI Use Statement on p. 11 makes this explicit. We would be grateful if you could look at the revised PDF, particularly Table 2, Figure 3, §2, and the abstract, and let us know whether it addresses your concerns.

I will maintain my clear reject rating. This practically became a new paper that needs a comprehensive re-review. It remains poorly written and its claims are not necessarily novel.

  • After reading the other reviews and the authors’ responses, it is clear that a lot of work has gone into the paper's substantial revision. The whole paper is blue, as it should be.

  • The new baselines reconstruct the embeddings more accurately and are less likely to alter the original model’s caption choices, whereas the proposed method identifies more concepts shared between images and text (Table 2).

  • The new ablations also show that most of this coverage gain comes from cross-decoding and latent masking, with Group-Hoyer providing only modest improvements in reconstruction and decision preservation.

  • Regarding "how this differs from prior work": Hierarchical CLIP SAEs [b, Sec. 5.2–5.3] already tested how feature edits change classification and retrieval, while discover-then-name [a, Sec. 5.3] explicitly evaluates concept interventions. PatchSAE [d, Sec. 4.1] already tests how retaining or removing selected SAE features affects CLIP’s predictions, including comparisons with random features. VISTA [e, Sec. 5] evaluates targeted concept removal and replacement and shows that interventions reverse specific visual judgments. DIAL [f] likewise evaluates changes in VLM predictions after removing SAE-identified directions. These papers go beyond reconstruction, retrieval and feature labels. Your particular test of whether reconstruction preserves choices between minimally different captions may be distinct, but causal testing of visual or multimodal SAE features is already established.

[d] "Sparse Autoencoders Reveal Selective Remapping of Visual Concepts During Adaptation" ICLR 2025

[e] "Interpretability Transfer from Language to Vision via Sparse Autoencoders" ICML 2026

[f] "Label-Free Mitigation of Spurious Correlations in VLMs using Sparse Autoencoders" ICLR 2026

Reviewer_n1CrOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper is interested in extracting concepts that are shared between encoders of two different modalities, while enforcing that these concepts are compositional. It proposes a sparse autoencoder (SAE) architecture with a separate SAE-encoder for each modality, and regularized by a proposed Group-Hoyer penalty. The SAEs are trained to reconstruct embeddings from CLIP, SigLIP, SigLIP2, on MSCOCO, and evaluate compositionality with SugarCrepe accuracy.

Critical strengths

The topic of multimodal and shared concept extraction is timely and enforcing compositionality is still an open-problem. The discussion and experiments about separation of the SAE encoders is valuable.

Critical weaknesses
  • The paper is very difficult to read and follow. Multiple notations are used without being introduced in the paper (e.g. FvF_v and FtF_t in Equation 2 and definition 1, HvH_v and HtH_t in Equation 2, AsaeA_{sae} and AdenseA_{dense} in Proposition 1, qq in definition 2,... ). The main figure from the paper does not help explaining the contributions. One of the technical novelty appears to be the "Group-Hoyer" regularization, but its description (in Section 4.4, 289-295) is very light. It does not explain where it comes from (no reference given). The only description is generic: "this promotes a small effective support without enforcing an exact number of nonzero activations" (l. 294), which is the case for any sparsity penalty. I cannot verify the soundness of this part. Learning separate encoders for shared concept dictionary is not novel by itself, it is done in "Universal Sparse Autoencoders" (USAE) [A]
  • All experiments related to the contribution about "revealing" causal circuit are in the appendix. This is never explicitly demonstrated in the main paper. A claimed contribution should be supported by evidence in the main text, not in the appendix.

[A]: Thasarathan, Harrish, et al. "Universal sparse autoencoders: Interpretable cross-model concept alignment." ICML 2025.

Questions for authors

The paper would require significant rewriting and restructuring to change my assessment (e.g. introducing experiments for a claimed contribution). I don't believe this is something doable during the rebuttal, and even if it is done, the updated paper would require a whole new review.

2 comments
AuthorsOpenReview ↗

Response to Reviewer n1Cr

Thank you for the careful review. We understand your concern about the effort required to assess a substantial revision. We revised the notation and method exposition, clarified the relationship to prior work, added a calibrated Group-Hoyer ablation, and placed the held-out intervention results in the main paper. We point to the relevant sections below so that each concern can be checked directly. Revisions are highlighted in blue in the revised PDF.

Notation. We added a notation paragraph at the start of §3.1 and now explicitly define the symbols you flagged: FmF_m and HmH_m (Eq. 2), AdenseA_{\mathrm{dense}} and ASAEA_{\mathrm{SAE}} (before Proposition 1), and the selector score qq (now introduced in §3.2 before Definition 2). The evaluation metrics (FVU, GPI, R@1) are defined in §§5.1–5.2.

Shared vs. separate encoders. We realize our original text made this unclear. Our SAE does not learn separate encoders: image and text share one encoder (WencW_{\mathrm{enc}}, Eq. 15) and one dictionary. The subscript mm only selects a modality-specific centering vector. "Modality-local" refers to inference: each modality's code is computed without access to its paired input. The text after Eq. 2 now states this explicitly. We also discuss Universal SAEs in §2: USAE aligns concepts across different models, whereas we align concepts across modalities within one model and evaluate whether the model's decisions are preserved.

Group-Hoyer (§4.4). We now cite Hoyer (2004) and related grouped Hoyer-type penalties used for structured pruning (DeepHoyer, ICLR 2020), and explain the design:

  1. The Hoyer measure uses the ratio of ℓ1\ell_1 to ℓ2\ell_2 to measure how concentrated the activations are. The ratio is scale-invariant in its unstabilized form, whereas an ℓ1\ell_1 penalty also shrinks activation magnitudes.
  2. Grouping the image and text activations of the same coordinate makes a coordinate used by both modalities count as one unit, encouraging shared use of dictionary atoms. MGSAE instead uses Group-L2,1L_{2,1}, an unnormalized sum of group norms.

We also added the missing ablation, replacing Group-Hoyer with plain Hoyer, ℓ1\ell_1, and Group-L2,1L_{2,1}, each calibrated to the same initial loss contribution (SigLIP2, K=256K=256; Appendix M):

PenaltyFVU ↓SugarCrepe DP ↑Strict shared units ↑
Group-Hoyer0.11210.89277.7
Plain Hoyer0.11300.88973.0
ℓ1\ell_10.11670.88579.3
Group-L2,1L_{2,1}0.11650.88581.7

At this operating point, Group-Hoyer modestly improves reconstruction and decision preservation over plain Hoyer. In the objective ladder, the largest increases in shared-concept coverage occur when cross-decoding and shared latent masking are added (Figure 2, Appendix N).

Causal evidence in the main text. §5.5 now contains the main causal results (Table 3), and Figure 3 shows a worked held-out example. On an image-disjoint held-out split, we exchange the activations of each seed's 32 selected features between the correct caption and its foil, and restore the reconstruction residual so that the edit acts on the frozen encoder's final pooled caption embedding. On SugarCrepe, the exchange reduced accuracy more than activity-matched random feature sets in all five seeds. The mean excess accuracy drop was 4.80 percentage points, with a seed range of 3.12–9.85 points. Circuits mapped from one seed into four independently trained SAEs retained the effect under both matching methods. For ablation, the paired 95% interval for the excess accuracy drop was above zero in 3 of 5 seeds, and ARO results are mixed; the paper states both.

Figure 1. We relabeled the panels and corrected and scoped the values in panel (c). The caption now identifies the operating points and clarifies that the comparison is unmatched in support.

Compositionality. To clarify one point in the summary: our SAE does not enforce compositionality. We evaluate whether it preserves the frozen model's compositional decisions, and the introduction now states this explicitly.

Continued in another comment

Thank you again for your review. You raised two concerns about assessing the paper: the main text needed restructuring, and the intervention claim needed direct evidence there. We addressed both while retaining the same SAE architecture and independent image and text inference procedure.

  • Restructured main text. Sections 3.1–3.2 define the notation you flagged, and §4.4 explains and tests Group-Hoyer. Table 2 and Figure 2 present the baseline comparisons and objective ablations. Figure 3 and Table 3 show a worked example and the aggregate held-out intervention results.
  • New experiments prompted by the reviews. We added matched-support comparisons with BatchTopK, JumpReLU, and Matryoshka-BatchTopK, a calibrated Group-Hoyer ablation, and an objective ladder that shows how training components affect decision preservation and shared-concept coverage. We also reran the interventions on an image-disjoint split, scored the edits with the reconstruction residual restored, and tested whether the exchange effect transfers across independently trained SAEs.

For a focused reread, §§3.1–3.2 and §4.4 address the notation and method concerns, and §5.5 and Table 3 present the evidence for the intervention claim. We would welcome your feedback on whether these revisions address your concerns.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response (1/2): Overview of the paper and its positioning

Overview of the paper and its positioning

We thank all three reviewers for their detailed and constructive feedback. We have uploaded a revised PDF; revisions are highlighted in blue. This comment summarizes the paper's core argument as revised; the accompanying comment (2/2) lists the changes and maps each reviewer concern to its location in the paper. We also respond to each reviewer individually.

The paper in brief

  • Problem. SAEs on dual-encoder vision-language models (CLIP, SigLIP, SigLIP2) are usually judged by reconstruction, retrieval, and feature labels. These do not show whether the SAE keeps the model's compositional decisions, for example preferring "a person riding a horse" over "a horse riding a person" for the same image.
  • Validity (Q1, §5.2). Common multimodal SAE evaluations can mislead. If the paired caption helps select an image's features, even random dictionaries reach near-perfect retrieval, and fixed product-TopK fills its support with tied zeros whenever fewer than KK products are positive. All our evaluations encode image and text independently and select features without ties.
  • Method (§4). A shared encoder and dictionary with independent image/text encoding, trained with cross-modal decoding, a grouped Hoyer sparsity penalty, and a paired auxiliary used only during training.
  • Trade-off (Q2, §5.3). At matched support on SigLIP2, our objective recovers about 1.5–1.8× as many shared image–text concepts as reconstruction-only SAEs (BatchTopK, JumpReLU, Matryoshka), which reconstruct better and preserve 0.5–2.5 points more decisions. Against MGSAE, it improves reconstructed SugarCrepe accuracy by 3.4–5.1 points at exactly 64–256 active features on SigLIP2. We present this as a fidelity–coverage trade-off, not a universal win.
  • Causal (Q4, §5.5). An exact per-feature decomposition of the caption-preference margin selects 32-feature circuits. On held-out images, exchanging them changes the frozen model's dense decisions more than activity-matched random controls, and on SugarCrepe the effect transfers to independently trained SAEs.
  • Negative result (Q3, §5.4). Relation categories are decodable (0.80–0.82 probe accuracy), but argument order is not (0.505 on Winoground pairs, chance 0.5): relation binding is not localized in individual features.

How this differs from prior work

Prior multimodal SAEs (e.g., MGSAE, SPARC, USAE, Discover-then-Name, hierarchical CLIP SAEs) evaluate whether features are recoverable, aligned, and interpretable: reconstruction, retrieval, and semantic labels. To our knowledge, they do not test whether the SAE preserves the model's compositional decisions, or whether its features causally drive those decisions. Our paper asks this directly, at three levels:

  1. Concept presence: are object, attribute, and relation concepts recoverable?
  2. Compositional faithfulness: does reconstruction keep the model's choice between minimally different captions?
  3. Causal control: do interventions on discovered features change that choice?

Sparse Feature Circuits (Marks et al., 2025) asks a related causal question for language models. We adapt it to dual-encoder VLMs, where the caption-side margin decomposition is exact, and add held-out, cross-seed tests. We also show that the evaluation itself needs safeguards: without independent encoding and tie-free selection, multimodal SAE results can look excellent for artifactual reasons.

We appreciate the strengths highlighted across the reviews, including the evaluation flaw exposed by the retrieval experiments, the experimental controls, and the negative result on relation binding. The revision builds on these findings with additional baselines and held-out intervention experiments, and clarifies the scope of the claims. We would be grateful if you could revisit the revised PDF, especially Table 2 for the matched-support comparisons and Table 3 for the held-out intervention results.

AuthorsOpenReview ↗

General response (2/2): Summary of changes and changelog by reviewer concern

Summary of changes and changelog by reviewer concern

Presentation

  • Rewrote the abstract, introduction, and contributions in plain language; reduced coined terminology.
  • Added a notation paragraph (§3.1) and defined every symbol and metric before use (FVU, GPI, R@1, SC/SC++, ITT).
  • Moved the main baseline, trade-off, causal, and qualitative results from the appendix into the main text (Table 2, Fig. 2, Fig. 3, Table 3).
  • Stated explicitly that image and text share one encoder and dictionary; only inference is per-modality.

New experiments

  • Three new baselines at matched support: BatchTopK, JumpReLU, Matryoshka-BatchTopK (Table 2, App. N).
  • An objective ladder from ordinary TopK to the full objective, mapping the fidelity–coverage trade-off (Fig. 2).
  • Group-Hoyer vs. plain Hoyer, ℓ1\ell_1, and Group-L2,1L_{2,1} under calibrated loss contributions (§4.4, App. M).
  • Causal interventions rerun on an image-disjoint split, scored on the frozen model's dense embedding with the reconstruction residual restored, plus executed cross-seed transfer under two matchers (§4.7, §5.5, Table 3).

Corrected or narrowed claims

  • The gain over MGSAE is now stated as 3.4–5.1 points on SigLIP2 at exact KK (the submitted abstract's range mixed operating points); Figure 1(c) values corrected.
  • Claims of better preservation replaced by the measured trade-off; causal claims scoped to the final pooled embedding; partial ablation necessity (3/5 seeds) and mixed ARO results stated explicitly.
  • Decision preservation is now computed on the tie-free subset used in Proposition 1, making the accounting identity exact.
  • Added a provenance appendix mapping every result to its checkpoint and configuration (App. L); clarified the AI-use statement.

Changelog

ConcernReviewerChangeWhere
Undefined notation (F,H,ASAE,qF, H, A_{\mathrm{SAE}}, q)n1CrNotation paragraph; all symbols defined before use§3.1–3.2
Figure 1 unclearn1CrPanels relabeled; values in (c) corrected and scopedFig. 1
Group-Hoyer unexplained, no referencen1Cr, bZTYMotivation; Hoyer (2004) and DeepHoyer cited; calibrated ablation§4.4, App. M
Separate encoders / novelty vs. USAEn1CrShared encoder and dictionary stated explicitly; USAE contrasted§2, §3.1
Causal evidence only in appendixn1Cr, fMsYCausal table and qualitative example in main text§5.4–5.5, Table 3, Fig. 3
Too few baselinesfMsYBatchTopK, JumpReLU, Matryoshka at matched supportTable 2, App. N
Missing related workfMsYSparse Feature Circuits, Discover-then-Name, hierarchical CLIP SAEs, USAE§2
Abstract and contributions hard to readfMsYRewritten in plain languageAbstract, §1
Redundant retrieval table, dense in-text numbersfMsYSubmitted retrieval table removed; numbers rounded§5.2
CLIP/SigLIP called "VLMs"fMsY"Dual-encoder vision-language models" defined§1
R@1 undefinedfMsYDefined at first use§5.2
No qualitative examplesfMsYHeld-out example with stated selection ruleFig. 3
Manuscript authorship and AI assistancefMsYClarified: authors wrote the text; AI tools were used for language polishingAI use statement
Trade-off vs. ordinary TopK unclearbZTYCommon-support comparison and objective ladder§5.3, Table 2, Fig. 2
Theory elementarybZTYPropositions framed as diagnostics§3.1
Causal effects only on reconstructionsbZTYResidual-restored dense scoring§4.7, §5.5
Cross-seed evidence only rank recoverybZTYExecuted transfer under two matchers§5.5, Table 3

We would be grateful if reviewers could look at the revised PDF, particularly §5.3 and §5.5, and let us know whether it addresses their concerns.

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

Related papers

Discussion (0)

Sign in to comment.