acceptodds
Under review as a conference paper at ICLR 2027

De-Absorption Reparameterization: Post-hoc Repair of Feature Absorption in Sparse Autoencoders.

Abstract

Sparse autoencoders (SAEs) are widely used to decompose neural-network activations into sparse, human-interpretable features. However, when the underlying features are hierarchically related, the same sparsity that promotes interpretability can distort their representation. In particular, a broad parent feature may stop activating when one of its more specific child features fires. This phenomenon, known as feature absorption, makes individual SAE features less reliable for interpretation and intervention. Existing approaches mainly aim to reduce absorption by modifying SAE training. In contrast, we ask whether feature absorption can be corrected after the SAE has already been trained. To address this question, we introduce De-Absorption Reparameterization (DAR), a post-hoc method that corrects feature absorption for a supplied relation between a parent feature and one or more child features in a trained SAE. DAR redistributes absorbed parent-related information from the child decoder back to the parent feature, while applying a compensating decoder update that preserves the SAE reconstruction exactly. Since reconstruction is preserved for any transfer coefficient, reconstruction alone cannot identify the correct repair. DAR therefore estimates the coefficients from parent–child activation patterns using sign-constrained least squares, ensuring that the parent activation only increases while all latent activations remain nonnegative. We evaluate DAR on JumpReLU and BatchTopK SAEs for Gemma-2-2B and CLIP ViT-L. Given the supplied parent–child relations, DAR restores parent activations on absorbed inputs across both language and vision. On the main language benchmark, DAR restores parent activation on up to 44.6% of previously absorbed examples. The repaired activations are more usable for their intended interpretation: intended-latent AUROC on absorbed inputs increases from 0.535 to 0.941. Casually, ablating the repaired parent produces a 15× larger effect on the relevant model logit.

Primary area:
interpretability and explainable AI
Keywords:
Interpretability, Post-hoc Repair, Reparameterization, Representation Learning, Dictionary Learning, Mechanistic Interpretability, Sparse Autoencoders

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_dG4yOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose a new method, DAR, that addresses the problem of feature absorption in already trained SAEs (i.e., it is the first such post-hoc method). Given a specified feature hierarchy (one parent and several children), the method calibrates the internal activations and the decoder matrix so that the output remains the same, but the parent feature fires more strongly. The experiments were conducted on 3 different models across text and visual modalities, on 3 different datasets with a given concept hierarchy. All experiments confirm that the level of feature absorption drops noticeably, which is further supported by a small causal experiment.

Critical strengths
  1. The authors propose a very simple, clear, cheap, and effective (absorption almost disappears across all metrics) method for addressing feature absorption in already trained SAEs.
  2. The authors evaluated their method across different modalities, model families, SAE types, and hierarchical concepts, which inspires confidence in all the conclusions of the paper.
Critical weaknesses

The main weakness of the method is its limited practical applicability. DAR requires a ready-made feature hierarchy as input, and in practice this involves two non-trivial tasks. First, the hierarchy has to come from somewhere: the authors use either the benchmark of Chanin et al., where the absorbing latents have already been identified, or external taxonomies such as ImageNet and WordNet, which simply do not exist for most domains. Second, each concept in the hierarchy has to be mapped to a specific SAE latent, which is hard in itself: a suitable feature may not exist at all, or the concept may be spread across several latents. The paper does not describe how this mapping is done for ImageNet and WordNet, even though all results depend on it directly. In essence, the method solves only the second half of the problem and assumes that the first and hardest half has already been solved. The work would be much stronger if the authors also proposed a way to discover such hierarchies directly in a trained SAE, for example from the feature co-activation matrix. This would give a complete pipeline: first automatically find the feature hierarchy, then remove absorption within it using DAR. In its current form, the contribution feels incomplete, and the method is applicable only to a small number of manually specified groups.

I am open to discussion. If the authors can show that I have misunderstood something or that my concerns are wrong, I will be happy to respond and reconsider my evaluation.

Questions for authors

How exactly are ImageNet and WordNet concepts mapped to SAE latents? What criterion is used, can one concept map to several latents, and what happens when no suitable latent exists? How many concepts could not be mapped, and how many groups were used in each setting?

Reviewer_7jUMOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates feature absorption of SAEs, a phenomenon that a specific child latent captures part of the signal associated with its parent, leaving the parent remaining inactive. Existing methods attempt to mitigate feature absorption by retraining a new SAE, whereas in this paper, the authors propose DAR, a novel post-hoc method for reparing already-trained SAE features.

Critical strengths
  • This paper investigates an important aspect of feature absorption. Instead of retraining new feature representations, this paper solves this problem through a post-hoc approach.

  • Theoretical justifications are provided for the parent–child reparameterization part of the DAR method.

  • The experiments are comprehensive and prove the effectiveness of the proposed method.

Critical weaknesses
  • A limitation of DAR is its assumption that a reliable parent-child hierarchy is available in advance. I think it would be valuable to further investigate whether DAR remains effective when such a hierarchy is unavailable, or at least only noisily estimated (for example, when parent-child relations are inferred approximately from feature semantics using an LLM). In particular, the robustness of DAR to incorrect or uncertain hierarchy edges would help clarify its practical applicability.

  • In Section 3, the authors use the notation β≤0\beta \leq 0 in the modification of zpz_p and dcd_c, which is less readable than using an α=−β≥0\alpha = -\beta \geq 0. Intuitively, to fix feature absorption, zpz_p should be modified larger whereas dcd_c should be modified smaller.

  • Section 3 lacks intuitive explanations on the estimation of β\beta.

  • There is only one subsection (4.1) in Section 4. The subsection title should be removed. Plus, I think a theoretical justification of β\beta estimation would be necessary. I haven't carefully check the theoretical results in the appendix. If there are some related results, then they should be moved to Section 4.

Questions for authors
  • Is kk fixed? According to Section 3, it seems that all parent nodes have the same number of child node. If so, how does DAR deal with real data where parent nodes have different number of childern? If not, the notations should be fixed.

  • Why is β\beta estimated using Eq.(11)? Is there any theoretical justifications? Why is yy also centered in this part?

Reviewer_SN4AOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper concerns feature absorption in sparse autoencoders (SAEs), in which more specific features (such as a dog) activate (absorb) the attribution of more general features (such as an animal). These feature relations in an encoder are usually linked hierarchically (an animal is an object but also a word starting with letter a). One popular view in the literature is that feature absorption in SAEs hinders their utility for interpretability and should therefore be reduced. This work proposes a post-hoc method for disentangling such hierarchical absorption relations in the already trained SAEs. Experiments aim to demonstrate that this method reduces absorption without degrading the SAE's reconstruction performance.

Critical strengths

One strength of this work is its well-founded motivation (reducing absorption in SAEs) and a specific idea for research, i.e. framing it as an optimization problem and fixing it post-hoc.

Critical weaknesses

My general impression is that this work is unfinished.

  1. The method (and its evaluation) requires defining an ad hoc hierarchy on the SAE latents, which is a major drawback.
  2. The experimental setup relies on quite obsolete model architectures (GPT-2, Gemma 2, CLIP) which have already been studied extensively, making the findings less relevant and interesting. To move the field forward and improve the paper's informativeness, I would advise interpreting more recent models such as Gemma 4 or SigLIP 2 instead. Experimental results are quite limited: Tables 1-3 resemble a proof-of-concept demonstration, while Figures 1-2 show two anecdotal examples without much insight.
  3. Presentation requires major revisions to pass the bar for publication, e.g.:
    • Abstract mentions the AUROC result that does not appear in the main body of the paper.
    • Difference in reconstruction is given an excessive number of times in-text.
    • Metrics used should be clearly described in the paper, not in the appendix.
    • Tables: Why are some elements bold? For example, DAR is bolded instead of the best score in Table 3.
    • Figures: Text and shapes are not aligned properly.
    • Figure 1: Why are there horizontal and vertical ellipses "..."?
    • L155: citation for Top-k SAE after mentioning Batch Top-k
    • L158: "Short" with textt
    • L274: Eqs. equation
    • L317/L319: citep instead of citet
Questions for authors

I don't think answers to any question could affect my assessment of this work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.