acceptodds
Withdrawn from ICLR 2027

Separating interpretability and reconstruction objectives in Sparse Autoencoders

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Sparse Autoencoders (SAEs) are effective tools for mechanistic interpretability by decomposing neural representations into sparse and interpretable features. However, a single sparse representation in traditional SAEs is responsible for both reconstruction fidelity and interpretability, which are not guaranteed to align. To separate these objectives, we propose DualSAE, a modular framework that augments standard SAEs with a plug-in residual branch. A standard SAE serves as the Sparse Branch, learning a sparse code that contains the dominant semantics of the input. The added Residual Branch models the non-semantic information omitted by the sparse branch. We ensure that the contributions of the two branches towards the overall reconstruction have minimal overlap through a decorrelation loss. We also adopt a two-stage training strategy that first stabilizes sparse representation learning before coupling with the residual branch. Across diverse encoder–SAE configurations on two vision datasets and two language corpora, DualSAE achieves mean relative improvements of approximately 29% in reconstruction fidelity and 34% in interpretability (through the monosemanticity score). Moreover, with more interpretable representations, DualSAE yields consistent performance gains in the downstream applications of topic modeling, hypothesis generation, and LLM steering. Qualitative analysis further suggests that the Sparse Branch indeed preserves dominant semantic content, whereas the Residual Branch captures non-semantic information to improve reconstruction fidelity.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, sparse autoencoder

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_TR3cOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes DualSAE, a modular framework that augments standard SAEs with a plug-in residual branch. Traditional SAEs use a single sparse representation responsible for both reconstruction fidelity and interpretability. Since reconstruction favors any useful information, including non-semantic information that may not correspond to interpretable concepts, it can force latents to mix multiple factors. To separate these objectives, a standard SAE serves as the Sparse Branch, which learns a sparse code containing the dominant semantics of the input and reconstructs it as x_M. The added Residual Branch takes the residual r = x − x_M as input and models the non-semantic information omitted by the Sparse Branch. The final reconstruction is x̂ = x_M + r̂. Across diverse encoder-SAE configurations on two vision datasets and two language corpora, DualSAE improves both reconstruction fidelity and interpretability.

Critical strengths

The method is simple and modular, since the residual branch can be plugged into many existing SAE variants without changing their sparsity settings. The experiments are broad, covering both vision and language models.

Critical weaknesses
  1. The paper's main idea is to keep interpretable information in the Sparse Branch and to move non-semantic information that only helps reconstruction into the Residual Branch. But in Figure 10, after adding r̂, the retrieved images change from “any dog” to “dogs with coat color and texture very similar to the query image”. The caption says “finer visual attributes of the query, including coat color and texture”, and the label in the figure even directly says “Meaningful features to complement x^M”. But coat color and texture are not noise. They are typical interpretable concepts. Standard SAEs often learn features like “brown fur” and “striped texture”. So at least part of what the Residual Branch holds is semantic, but the authors call it “non-semantic” in the same paragraph, which is very contradictory. And the authors never define what “non-semantic” means. Attributes like coat color and texture are naturally shared across classes. A “brown fur” feature can appear in dogs, bears, and horses at the same time, so retrieving with it will of course return images from different classes. So “retrieving different classes” cannot distinguish between two cases: “it is really noise” and “it is a cross-class attribute feature”. The authors also count the latter as non-semantic. Based on this, if part of what goes into the Residual Branch is actually made up of interpretable features, then the good results do not come from a better decomposition, but from attribute features being moved out.
  2. The comparison with the baselines does not seem fair. The baselines use only 64 sparse latents (TopK, BatchTopK) to reconstruct each input, while DualSAE uses 64 sparse latents + r̂, which contains information from the real input. Obviously, DualSAE has much richer information. The reconstruction metrics of SAEs are meant to measure “how much of the model’s behavior the interpretable features can explain”. But part of DualSAE’s x̂ comes from a black box, so the metric now measures “how much the interpretable features and the black box can explain together”.
  3. Purity is computed based on ImageNet class labels. This gives an advantage to class-level features such as “dog”, but is less fair to cross-class attribute features such as “brown fur”. DualSAE happens to keep only the dominant class-level semantics, so it benefits from this metric.
  4. In Table 4 of the appendix, the row for JumpReLU SAE says “BatchTopK activation”. Maybe there are other writing errors, please check carefully.
Questions for authors
  1. For Weakness 2, maybe the authors could throw r̂ away after training DualSAE and only use x_M. Compute the R² between x and x_M, as well as ΔCE and KL. Also, for a standard SAE, authors could first train it and freeze it, then attach the same Residual Branch and train only the Residual Branch. The only difference between this and DualSAE is whether the Sparse Branch is trained jointly with the Residual Branch. This would show how much of the gain comes from the Residual Branch itself and how much comes from joint training.
  2. Could the authors evaluate with attribute-level labels or use auto-interp for testing?
  3. Could the authors report the L0 of JumpReLU and Gated in the baselines?
  4. Could the authors provide a quantitative evaluation of steering?
Reviewer_Q5J5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on sparse autoencoders (SAE) and identifies the conflict between its reconstruction fidelity and interpretability objectives. To resolve this, they propose to explicitly disentangle these objectives with a dual SAE method that adds a residual branch to incorporate uninterpretable components separately from the regular SAE that incorporates interpretable components. They introduce a decorrelation loss that minimizes the correlation between the residual component and the SAE decoder's interpretable components. For training, they use a two-stage approach: first training the sparse branch alone followed by jointly training both branches. Finally, they provide extensive experiments comparing both vision and language SAEs to demonstrate the improvements of the dual SAE.

Critical strengths
  • The approach of dual SAE is simple and intuitive. The idea of separating the reconstruction and interpretability objectives makes sense.

  • The paper provides extensive experiments across vision and language domains to show the improvements of dual SAE over standard SAEs across multiple SAE types as well.

Critical weaknesses
  • The novelty of separating interpretable and uninterpretable parts of the SAE is unclear. A relevant paper is [W1] which also separates these parts, albeit with a rank-r bottleneck instead of the residual. It would be good to clarify the similarities and differences in the high-level ideas, as well as add some quantitative comparisons.

  • It is unclear how much of the interpretable features contribute to reconstruction and how much of the residual. A simple experiment could be to track the reconstruction loss between xx and xMx^M (from sparse branch) and the reconstruction loss between xx and xM+rx^M+r. Then we could quantify how much of the activations can be made sparsely interpretable and how much of it is not interpretable.

    • Another ablation that could be useful is to remove the sparse/interpretable branch and see how much reconstruction loss the residual alone can achieve.
  • It is also unclear how much more expensive dual SAE is compared to standard SAE for training. It would be good to report training time and GPU memory usage. We need to check if the improvement is traded off by the more expensive training.

    • A related issue to training cost is that there are no hyperparameter sensitivity analyses for the newly introduced hyperparameters like the loss weights (γ\gamma, λ\lambda) or the width of the residual branch. So it is difficult to say how important hyperparameter tuning could be and that would further add to the training cost.

References

[W1] Deng et al., "Decompose Sparsely Where You Should, Absorb Densely Where You Should Not", ICML 2026 Mechanistic Interpretability Workshop

Questions for authors

Please address the weaknesses section.

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

The paper proposes DualSAE, a novel SAE architecture that addresses the problem with dual objective of reconstruction and interpretability. DualSAE adds a residual branch that improves reconstruction quality while also improving the interpretability of the main sparse branch.

Critical strengths

(S1) DualSAE is a novel architecture that addresses the dual objective of reconstruction and interpretability in SAEs and improves over the baseline SAEs. The quantitative results in Table 6 and the radar plots in Figures 6 and 13 support this.

(S2) The paper evaluates DualSAE on a broad set of evaluations from prior SAE work:

  • the MonoSemanticity Score [1] with class purity (Sec. 7.4)
  • topic modeling (Sec. 7.8)
  • steering (Sec. 7.9 and Appendix C)

In each, DualSAE shows advantages over the standard SAE.

(S3) Beyond reconstruction, the paper reports interpretability metrics (MonoSemanticity Score and class purity), which also improve over the SAE. This indicates that the residual branch improves reconstruction and also benefits the interpretability of the main sparse branch.

Critical weaknesses

(W1) The motivation for a residual branch is not convincing, as motivation from section 3 Motivation shows that residual error rr contains mostly non-semantic information; a cheap solution to this problem (which was done in prior SAE works) is just to add this residual back into the SAE output. The authors should explain why a learned residual branch is necessary rather than simply adding back the error term. DualSAE does improve interpretability, which suggests that the SAE branch becomes more semantic while the residual branch captures non-semantic information. However, neither the motivation nor the training objective enforces this separation.

(W2) The projection layer is underspecified and described only in the appendix Table 4 as a one-layer projector with hidden width, (L. 724), which suggests a two-layer network. Its architecture should be defined precisely, and the paper includes no ablation on its width.

(W3) Section 7.2 raises a key question: do the improvements come from the architecture or from the additional parameters? Table 1 addresses this, but the gains over some baselines, such as BatchTopK and TopK, are marginal (below 0.010.01 on most metrics) with no std. These differences should be tested for statistical significance. Moreover, all comparisons in the paper, not only those in Section 7.2, should be made at matched parameter counts. Otherwise, the results are biased in favor of DualSAE.

(W4) The paper evaluates only a single setting, with no results across different expansion factors and sparsity levels.

Minor Weaknesses

  • (W5) The structure of the paper should be improved. Sections 4–6 should be merged into a single section with subsections. Section 7 needs an experimental setup subsection that specifies the exact encoder model versions and the evaluation metrics, with references. Section 7 should also be organized into thematic subsections rather than a sequence of paragraphs.
  • (W6) The related work omits MP-SAE [3] and lacks related work on SAE downstream applications.
  • (W7) The introduction of class purity (L. 120) is unclear and hard to understand.
  • (W8) The MonoSemanticity Score is first introduced (L. 202) without a reference to [1].

References

[1] Pach, Mateusz, et al. "Sparse autoencoders learn monosemantic features in vision-language models." Advances in Neural Information Processing Systems 38 (2026): 95706-95742.

[2] Karvonen, Adam, et al. "SAEBench: A comprehensive benchmark for sparse autoencoders in language model interpretability." arXiv preprint arXiv:2503.09532 (2025). (ICML 2024)

[3] Costa, Valérie, et al. "From flat to hierarchical: Extracting sparse representations with matching pursuit." Advances in Neural Information Processing Systems 38 (2026): 34381-34424.

Questions for authors

(Q1) Can the authors give a precise description of the projector layer, including the number of parameters it adds on top of the SAE?

(Q2) Following up on (W1): if rr encodes mostly non-semantic information, why not simply add it back to the output instead of training a residual branch?

(Q3) Following up on (W3): can the authors report standard deviations for Table 1 and test whether DualSAE’s improvements are statistically significant at matched parameter counts?

(Q4) The MonoSemanticity Score and class purity results suggest that DualSAE improves the interpretability of the main SAE branch. However, the codebase implementation of the MonoSemanticity Score (loss.py) appears to differ from the original. For example, it uses the same encoder E(⋅)E(\cdot) both for the SAE activations and for computing image similarity (In [1], Appendix E.2, Until now, our analysis has focused [..] the encoders always differ). Can the authors confirm whether their implementation matches the original? If it differs, can they justify the change or report results with the original implementation?

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

Related papers

Discussion (0)

Sign in to comment.