acceptodds
Under review as a conference paper at ICLR 2027

How Optimality Structures Sparse Dictionaries: Theory for Interpreting SAE Representations

Abstract

Sparse Autoencoders (SAEs) have found success parsing neural network representations into interpretable concepts, providing a basis for understanding and control. However, what exactly SAEs extract and, hence, the scientific conclusions we can draw from them are not obvious. In short, if your SAE behaves strangely, does that reflect interesting neural network behaviour or an SAE-imposed distortion? Towards answering this, we use dictionary learning identifiability results to derive constraints that optimal dictionary learning features must satisfy. For example, an optimal feature will never turn on only while another is active. We use these conditions to explain various SAE oddities - hierarchical splitting & absorption, which features can be left in the residuals, dense antipodal features, and infinite feature splitting - simply as properties imposed by the dictionary learning objective. Finally, these constraints are diagnostic: real SAEs pass when measured on the dataset on which they were trained, but increasingly fail as the test dataset becomes more `distant'. In sum, we hope to provide theoretical tools to explain puzzling SAE patterns, allowing more principled inferences about internal model behaviour.

Primary area:
interpretability and explainable AI
Keywords:
Sparse Autoencoders, Mechanistic Interpretability, Dictionary Learning, Representation Learning, Neural Coding

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

This paper proposes a unified optimality framework to theoretically explain feature splitting and feature absorption for SAEs. It derives necessary local-optimality conditions on learned SAE features and residuals. It also analyzes the wide-dictionary limit, showing that increasingly wide SAEs tend to refine representations toward ray-based partitions of the data. Empirical examples are provided to aid the understanding of the theoretical conclusions.

Critical strengths
  • This paper investigates an important problem of theoretically explaining feature splitting and feature absorption for SAEs.

  • This paper proposes a unified optimality framework. Particularly, the feature–residual constraints shown in Section 5 seem novel enough.

  • The theoretical conclusions are justified through empirical examples.

Critical weaknesses
  • The presentations of the theoretical results are a bit unclear. It would be necessary to formulate them clearly as assumptions, theorems, and proofs.

  • The assumption of no optimization gap is not clearly listed as the other assumptions. And empirical justifications on whether this optimization gap is small enough are lacking.

  • Although "relating and generalising the findings", some of the theoretical insights are already discussed in prior related works, which somewhat limits the novelty of the proposed theoretical framework. E.g. Chanin et al. (2024) "show mathematically that if the low level feature activates only when the high level feature is active, feature absorption decreases the loss" and Cui et al., (2026) discuss that their theoretical results "can partly explain the feature absorption phenomenon because non-overlapping sparse features are more recoverable". (The relation with the latter one should also be discussed in Section 4.)

Questions for authors
  • How large is the optimization gap empirically?

  • Can the theoretical analysis explain what kind of parent features are more likely to split than others under a given number of latent dimensions? E.g. is a parent feature with more diverse child features more likely to split? Such a result could further strengthen the paper, as some of the theoretical insights are already known from prior related works.

5 comments
AuthorsOpenReview ↗

Response 1/2

We thank the reviewer for their review and constructive suggestions.

W1: Formal statements

We agree. All results are derived in full in the appendices, but they are not formatted as theorem–proof statements. We will restate the main results as formal propositions with explicit assumptions for a camera-ready version.

W2: The optimal-code assumption

Thank you; we agree this should be stated alongside the other assumptions. We now list it explicitly in Section 3: codes are optimal for the dictionary, so there is no optimisation (amortisation) gap. We also refer back to it wherever we discuss the theory's limitations.

Q1: How large is the optimisation gap?

There is a measurable gap between the codes produced by a standard SAE encoder and optimal sparse codes. O'Neill et al. (2024) study it in detail, and show that more sophisticated inference, such as iterative sparse coding or more expressive encoders, outperforms standard SAE encoders at modest extra compute. For our purposes the relevant question is whether the structure we predict survives the gap, and the evidence suggests it does. The GPT-2 SAEs we analyse have standard ReLU encoders, yet show no violations of our conditions on their training distribution. In our new toy experiments, ReLU SAEs with jointly trained encoders reach the same absorbed solutions as idealised dictionary learning, while independent features are left in place.

W3: Relation to prior work (Chanin et al., Cui et al.)

Thank you; we agree these relations deserve a fuller discussion, and we have added a comparison with both works, including Cui et al., to Section 4. In short, we believe our results are a significant advance on existing SAE theory, permitting much more flexible application, as our range of applications shows.

As acknowledged in the paper, the core interaction between the data’s co-activation structure and L1 optimisation that produces hierarchical absorption and splitting was previously explored by Chanin et al. However, their result comes from a specific toy model built to explain exactly this phenomenon. Even for absorption, eq. (6) goes further. Chanin et al. analyse one constructed path towards absorption in a toy with four orthogonal features of fixed magnitude. Eq. (6) shows that a dictionary in which one latent fires only when another does cannot be optimal, for any feature geometry and magnitude distribution, and it can be checked on trained SAEs. More broadly, our conditions apply to any dataset, and we use them to study not just hierarchical phenomena but range-independence, dense antipodal feature pairs, the structure of the residuals, and unbounded splitting in the wide limit. This is a significant extension of Chanin et al.’s results.

The price of this generality is that we idealise the encoder away. Tang et al. and Cui et al. (2026) model the encoder explicitly, but their results assume independently sampled sparse features, whereas ours make no assumptions about the data. This is what allows us to study hierarchical features directly. As discussed in our response to Reviewer dmvY and in new text in Section 4 and Appendix G, Tang et al. offer an explanation based on spurious (partial) minima that we do not think is sufficient.

Cui et al. note in their App. B that their recovery theorems hold only for features that are extremely sparse and do not overlap, and suggest this may partly explain absorption. In their example, however, a parent feature is defined as a positive combination of its children, so the two overlap in two ways at once: they point in similar directions, and the parent is always active with the child. Their analysis cannot say which of these causes absorption, and their theorems assume that features are active independently, which rules out a child that only appears with its parent. Our eq. (6) isolates the second effect: if one feature is only ever active when another is, a dictionary with a separate latent for each cannot be optimal, whatever directions the features point in (unless the two latents point the same way). Our new toy experiments (App. G) confirm this. We keep the features’ directions and frequencies fixed and change only whether the child appears independently or only with its parent; L1 dictionary learning and L1 SAEs absorb only in the second case. Cui et al.’s intuition therefore agrees with ours, and eq. (6) makes it precise and identifies the cause, something that lies outside their theoretical setting.

To our knowledge, no previous theory accounts for this range of SAE phenomena in a single framework, or provides a diagnostic that can be checked on trained SAEs. Hindupur et al. (2025) study the complementary question of what different encoders can represent; combining the two approaches is a natural next step.

AuthorsOpenReview ↗

Response 2/2

Q2: Which parents split first?

This is a good question. Our theory is a necessary set of conditions that every optimum, local or global, must satisfy. This is what says that a parent and child feature simply cannot co-exist at an optimum. However, when there are fewer SAE neurons than features, different clusterings of child features can form local minima and our theoretical results merely state that they are allowed, rather than ranking them. Enumerating which clustering is globally optimal is a separate problem that is non-trivial.

A simple extension is to then calculate the value of the loss at each of these local minima, and check what that says about which features will split. In short, subdividing a parent into two is a gain when it contains children that point in different directions and occur often, matching your intuition that diverse children lead a hierarchy to split in narrower SAEs than more homogenous ones. We tried a little set of toy experiments that demonstrate this point which we include in the anonymised git repo and are happy to include if interesting enough. In short, they show that adding latents to the dictionary often but not always splits according to a greedy heuristic such as ‘split the latent that can be subdivided to maximal gain’. That said, the full problem is a combinatorial clustering search, matching the observation that at small SAE widths the problem is non-convex and making statements about the global optimum is challenging.

Thank you for the detailed rebuttal. It partly addressed my concerns. However, I have two concerns remain:

W1: I appreciate the proposed formalization of the theoretical results, but would prefer to see it incorporated into the manuscript during rebuttal rather than deferred to camera-ready.

W2/Q1: Clarifying the optimal-code assumption is helpful, but satisfying necessary optimality conditions does not demonstrate a small optimization gap. Direct empirical quantification would be more convincing.

I maintain my rating of 3 for now

Reviewer_dmvYOpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
2(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper derives local optimality conditions for dictionary learning and uses them to show that several puzzling SAE behaviors, hierarchical feature splitting and absorption, residual structure, dense antipodal features, and infinite splitting, are largely consequences of the dictionary learning objective itself rather than necessarily reflecting the internal structure of neural networks.

Critical strengths

Provides a clean, unifying theoretical framework that explains multiple otherwise disparate SAE oddities through the same set of optimality constraints, reducing the risk of mistaking tool-induced artifacts for genuine network behavior.

Critical weaknesses

The major results are derived for vanilla dictionary learning. Real SAEs, however, typically include architectural constraints (e.g., TopK, JumpReLU, Gated...) and an amortization gap between the ideal sparse codes and the encoder’s outputs (Compute Optimal Inference and Provable Amortisation Gap in Sparse Autoencoders, A Unified Theory of Sparse Dictionary Learning in Mechanistic Interpretability: Piecewise Biconvexity and Spurious Minima). If these constraints are taken into account, many of the “must-split / must-absorb” pressures may no longer be strictly binding, potentially allowing hierarchical features to coexist stably and thereby weakening the claim that the observed phenomena are primarily forced by dictionary-learning optimality.

Questions for authors
  1. This work (A Unified Theory of Sparse Dictionary Learning in Mechanistic Interpretability: Piecewise Biconvexity and Spurious Minima) shows that the SAE optimization landscape contains many spurious local minima that systematically produce feature absorption. How do you distinguish between the phenomena being forced by the optimality conditions of the objective versus arising from optimization getting stuck in such bad local minima?
  2. Your optimality conditions are derived for the ideal dictionary learning problem. Given that real SAEs use constrained encoders (e.g., TopK, JumpReLU) and often fail to reach the corresponding sparse codes due to amortization gaps, to what extent do these conditions still explain the feature splitting and absorption observed in practice?
4 comments
AuthorsOpenReview ↗

Response 1/3

Thank you for your time spent reviewing our work.

In short, yes, as acknowledged in the paper, there remain theoretical gaps between the object of our theory and real SAEs. Having said that, we argue that our theoretical work is a significant advance on the existing approaches, at times extending, at times complementing the present literature. We argue these advances, while not a complete picture, present an intuitive and useful lens through which to understand SAE phenomenology. We will now argue this in two parts. Part 1 addresses your Q1 and Part 2 your Q2.

Part 1: Insufficient comparison and study of a restrictive setting relative to Tang et al. (2025)

We thank the reviewers for prompting us to set out the similarities and differences between our approach and that of Tang et al. (arXiv:2512.05534, v6). In short, we do not believe Tang et al. offer a sufficient explanation of hierarchical absorption, as their theoretical assumptions preclude exactly the kind of data we are trying to explain, and we add a small supporting appendix to demonstrate this point.

We would first like to highlight how different the two settings are. We study the optima of a classic L1 dictionary-learning problem. Our main limitation is that we idealise the encoder away: codes are optimal for each input rather than produced by a particular architecture. In this setting we show that hierarchical distributional structure (a child feature active only when its parent is) means a dictionary with separate parent and child latents cannot be optimal, instead absorption and splitting can be, generalising Chanin et al.’s analysis, and explaining the observed absorption and splitting as a result of the objective-data interaction.

Tang et al.’s main advantage over our work is that they model the encoder weights directly. But in other respects, their setting is more restrictive, particularly its data assumptions: their results concern an approximate loss that is only accurate for very sparse variables whose sparsity patterns are independent. To state their argument more precisely, Tang et al. study a reconstruction objective. They assume features are active independently of one another, each in at most a fraction 1/n of inputs, where n is the number of features. Independence is required for eq. 35 of their proof of Thm 3.1, as their limitations section acknowledges. Under these assumptions they approximate the loss by its terms from inputs in which exactly one feature is active (Thm 3.1). They then state that every realisable polysemantic activation pattern that partitions the features contains a partial minimum with suboptimal loss (Thm 3.6) [1]. Their theorem 3.8 then shows that if an encoder can make one latent respond to a whole group of child features, it can also make that latent drop one child, which then gets its own latent (the absorption pattern). This is a statement about which activation patterns an encoder can produce. Combined with Thm 3.6, the implication is that absorbed patterns are spurious (partial) minima, so absorption reflects optimisation getting stuck on the route to a better solution. The reviewer reasonably asks us to distinguish this optimisation argument from our objective-based explanation.

Our main concern is that their theory does not model the hierarchical effects we are reporting. Theorems 3.2–3.6 concern the approximate loss, which Thm 3.1 justifies only for independent, very sparse features. Hierarchical data breaks this assumption: a child that is active only with its parent co-occurs with it as often as the child occurs at all. Their own definition of hierarchical features, definition 3.2, assumes that the parent and child are each never active alone, something that their approximate loss never sees since it only measures performance on points in which a single feature is active. As such, even within the paper’s own logic, the explanation rests on an approximate loss whose assumptions hierarchical data violate, and no evidence is given that the approximation still holds.

AuthorsOpenReview ↗

Response 2/3

We empirically tested the premise of this argument directly: that the un-absorbed dictionary is a minimum, with absorption a separate, worse minimum. In a new Appendix G, we study vanilla SAEs, L1 dictionary learning and TopK SAEs initialised as a set of ground truth hierarchical features and trained by gradient descent. We find that for vanilla SAEs and dictionary learning the un-absorbed dictionary is not a local minimum, descent moves to absorption, as predicted by the theory, and that this holds for orthogonal and non-orthogonal features, with fixed or variable magnitudes. With independent features of the same frequencies, the ground truth stays in place. For TopK SAEs the ground truth is likewise not a minimum whenever more than k features co-occur. Tang et al. report experiments using only independent features, while, consistent with our experiments, Chanin et al. find that absorption is learned reliably in toy models, including TopK SAEs, and appears in every LLM SAE they test.

In summary, we do not believe Tang et al. offer a sufficient explanation of the absorption of child features that are active only with their parent. Their theoretical assumptions exclude this setting, and we show that in the true loss the un-absorbed dictionary is not a minimum at all for L1 dictionary learning and L1 SAEs. We have added a paragraph to section 4 to explain these links, corrected our related work section, and added an appendix with these experiments. We apologise for not including this comparison initially.

[1] A note on Thm 3.6. We believe the step from eq. 64 to 65 is an error, and that the statement fails in general. As an example, their Example 3.5 with unequal feature weights M = (2, 1) does not satisfy the zero gradient conditions for the encoder weights. In fact, alternating minimisation changes the activation pattern of the system, and the latent drops the weaker feature. Although our main argument does not depend on this, the theorem can be repaired under additional conditions, for example linearly independent features whose pairwise dot products are positive, and we have reported this to the authors.

Part 2: What the theory says and doesn’t say about real SAEs

We agree with the reviewers that, as discussed in our original submission, our theory says essentially nothing directly about the effects of different nonlinearities or encoder architectures. Ideally future work will expand these results, e.g. to the TopK architecture.

We nevertheless think the results are useful, for the reason solvable toy models usually are. Toy models of superposition (Elhage et al., 2022) and normative models of sensory coding (Olshausen & Field, 1996) do not predict any particular network exactly; their value lies in isolating a mechanism and showing what follows. Our model isolates the ingredient every SAE shares: fitting the data as a sparse linear combination of learnt directions. We then present a series of implications. If a phenomenon arises simply from this objective structure then that is informative: it gives a plausible explanation that does not depend on the details of the architecture or optimiser. If it cannot be explained in this framework then that is also informative, highlighting other parts of the problem.

In summary, from one family of optimality conditions we account for a range of reported behaviours: hierarchical splitting and absorption; which features can be left in the residuals; dense antipodal latent pairs; and unbounded splitting as width grows, which we relate to ray clustering around the geometric median. The same conditions act as a diagnostic on trained SAEs: GPT-2 SAEs show no violations on their training distribution, and increasingly many as the test data move to other datasets or layers.

There is also evidence that these intuitions transfer beyond the idealised setting. SAE encoders do not reach optimal sparse codes (O’Neill et al., 2024), yet the GPT-2 SAEs we test, which have standard affine-ReLU encoders, satisfy conditions derived for the idealised problem. Dense antipodal pairs were reported in TopK and JumpReLU SAEs (Sun et al., 2025), and they disappear when signed activations are allowed (AbsoluteTopK), as expected if the nonnegativity of the code is what forces a dense variable to be split. In the new Appendix G, the hierarchy prediction holds for ReLU SAEs exactly as for idealised dictionary learning. This is how we hope the framework will be used: as a source of predictions about which pressures are present, to be checked in the architecture at hand.

AuthorsOpenReview ↗

Response 3/3

On your concern that encoder constraints might let hierarchical features coexist stably: in our toy experiments, ReLU SAEs do not keep separate parent and child latents, and in TopK SAEs coexistence is stable only when the k budget never binds, in which case both dictionaries have zero loss; absorption is never the worse solution.

Finally, we believe it significantly extends existing SAE theory work. Chanin et al. (2025) analyse one constructed path towards absorption in a four-feature orthogonal toy with fixed magnitudes. We agree with their conclusions; eq. (6) generalises them, showing that nested parent–child supports are never locally optimal for any feature geometry or magnitude distribution, and the same conditions yield the other results above. Tang et al. and Cui et al. (2026) explicitly model the encoder, which we do not, but they prove recovery results only for independently sampled sparse features. Our conditions require no assumptions about how the data were generated, which is what lets us analyse hierarchical, correlated and dense structure, and test trained SAEs directly on real activations. We highlight our differences to Tang et al. above, while Cui et al. discuss hierarchy only informally (their App. B), noting that their recovery theorems need non-overlapping features, which a parent and its children are not, and suggesting that this may partly explain absorption. Our work makes this argument precise, without assuming sparse independent features. To our knowledge, no previous theory accounts for this range of SAE phenomena in a single framework or provides a diagnostic that can be checked on trained SAEs. Hindupur et al. (2025) study the complementary question of what different encoders can represent; combining the two approaches is a natural next step.

We now make more of this motivation explicit in the paper, in Section 3 and the discussion, to try and make the value of our work clearer, and we include a more thorough comparison to existing SAE theory literature (Section 4 and Appendix G).

We are sorry for the delay in posting this rebuttal and hope you find time to engage further with our work.

Thanks for the clarification, which addresses my main concerns. I believe this work would contribute some insight to the community. I have raised my score to 3.

Reviewer_A1GMOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents a study of L1 sparse autoencoders focusing on what structure between learned latents arises when SAE code is taken as optimal. Authors explain feature splitting and absorption as two instances of the same mechanism, characterize the behavior of densely activated latents, studies the residual structure and analyses the infinite-width limit that leads to infinitely increasing feature splitting, also providing the diagnostics on real SAEs on GPT-2. The optimality conditions of the paper are built without assuming how the data is generated.

Critical strengths

S1. The writing is very clear and easy to follow, I found the paper interesting by its own. The only presentational issues I would highlight are described in W5.

S2. A nice fruitful framework is presented that enables explanation of several phenomenas from first principles, and linking between absorption and splitting is somewhat counterintuitive. Its independence of assumptions about how data is generated is a major strength. It generalizes existing work and extends it. Despite critique below, I found experimental design careful and well-executed on its own.

Critical weaknesses

W1. The main problem of material is that it sheds light on L1 SAEs, saying essentially nothing theoretically about what would happen in TopK SAEs where sparsity mechanism is notably different and might lead to different behavior.

W2. The encoder part is underdeveloped. Paper builds on assumption that code is optimal for the dictionary, but essential question is whether joint training of SAE encoder and decoder reaches sufficiently nice solution; it might be outside the scope of the paper, but I think it is important to analyze expressivity and reachability, since joint training properties might significantly differ (e.g. encoder might constrain the code, then the code constrains the dictionary, and vice versa).

W3. Despite having theoretical generality, I found toy model empirics presented in a paper somewhat anecdotal - these do not characterise what happens in practice in different regimes sufficiently enough; many examples are 2D while severe overcompleteness, different activation generator properties might alter some conclusions. The theory is built to be generator-independent, but with an assumption of a given optimal code it is hard to know whether it transfers to actual SAE training, especially when talking about different models and datasets in future real-world use (where data generation properties and underlying features might differ and be very complex). Moreover, the real-world SAE study is limited to one family and it would be nice to see if conclusions generalize beyond that.

W4. There is no reproducibility details for toy launches; it would be beneficial to have it so every conclusion is clearly reproducible.

W5. Main writing problem from my point of view is that paper reads as phenomenological study that tells what happens in what conditions without clearly forming consequential general view well enough. This is mostly reading style, but I think that more focused exposition, with clearer "so what?" question carried for each section and throughout the paper would help reader form the general view on what they might do with it and how it changes anything in usual practice or thinking about problems.

Questions for authors

Q1. Do you see any change in any conclusion that would stem from changing the underlying generator in data? For example, taking a Zipf law for feature frequencies distribution, would it alter anything, or do you see it as a default underlying assumption (as with real-world Zipf distributions)? I might overthink the role of data generating process so I am interested in your opinion here.

Q2. How do you see connection of your results to TopK and other encoder designs, including nonlinear ones?

Q3. You cite Sun et al. for 1/2 activation limit of dense latents, but the cited paper does not show this to be an upper limit. Do you think that this is an estimation problem and law of large numbers will cover it?

3 comments
AuthorsOpenReview ↗

Response 1/2

We thank the reviewer for a careful and generous review. Below we explain what the theory does and does not say about encoders (W1, W2, Q2), address the role of the data generator and the toy experiments (W3, Q1), reproducibility (W4) and writing (W5), and correct the error the reviewer identified (Q3).

W1 + 2, Q2: L1 SAE w/ no encoder constraints

We include a more general response about the theoretical gaps and motivation of our model below that we also sent to reviewer dmvY. Specifically on the two points the reviewer raises:

On joint training (W2): our theory characterises the dictionary given optimal codes, and unfortunately training effects are out of scope. In practice, though, the structures we predict do appear in jointly trained SAEs. The GPT-2 SAEs we analyse have standard affine-ReLU encoders and satisfy our conditions on their training distribution, and in our new toy experiments, ReLU SAEs with jointly trained encoders, started from the un-absorbed dictionary, move into absorption exactly as idealised dictionary learning does.

On other encoders (Q2): more sophisticated inference, such as iterative sparse coding or more expressive encoders, outperforms standard SAE encoders at modest extra compute (O’Neill et al., 2024), so their codes approach the optimal codes our theory assumes. Different sparsity mechanisms such as those at work in TopK SAEs are likely more different. Hindupur et al. (2025) study what different encoders can represent; combining their approach with ours is a natural next step.

More general response

We agree with the reviewers that, as discussed in our original submission, our theory says essentially nothing directly about the effects of different nonlinearities or encoder architectures. Ideally future work will expand these results, e.g. to the TopK architecture.

We nevertheless think the results are useful, for the reason solvable toy models usually are. Toy models of superposition (Elhage et al., 2022) and normative models of sensory coding (Olshausen & Field, 1996) do not predict any particular network exactly; their value lies in isolating a mechanism and showing what follows. Our model isolates the ingredient every SAE shares: fitting the data as a sparse linear combination of learnt directions. We then present a series of implications. If a phenomenon arises simply from this objective structure then that is informative, giving a plausible explanation independent of the details of the architecture or optimiser. If it cannot be explained in this framework then that is also informative, highlighting other parts of the problem.

In summary, from one family of optimality conditions we account for a range of reported behaviours: hierarchical splitting and absorption; which features can be left in the residuals; dense antipodal latent pairs; and unbounded splitting as width grows, which we relate to ray clustering around the geometric median. The same conditions act as a diagnostic on trained SAEs: GPT-2 SAEs show no violations on their training distribution, and increasingly many as the test data move to other datasets or layers.

There is also evidence that these intuitions transfer beyond the idealised setting. The GPT-2 SAEs we test have standard affine-ReLU encoders, yet satisfy conditions derived for the idealised problem. Dense antipodal pairs were reported in TopK and JumpReLU SAEs (Sun et al., 2025), and they disappear when signed activations are allowed (AbsoluteTopK), as expected if the nonnegativity of the code is what forces a dense variable to be split. In the new Appendix G, the hierarchy prediction holds for ReLU SAEs exactly as for idealised dictionary learning. This is how we hope the framework will be used: as a source of predictions about which pressures are present, to be checked in the architecture at hand.

AuthorsOpenReview ↗

Response (2/2)

Finally, we believe it significantly extends existing SAE theory work. Chanin et al. (2025) analyse one constructed path towards absorption in a four-feature orthogonal toy with fixed magnitudes. We agree with their conclusions; eq. (6) generalises them, showing that nested parent–child supports are never locally optimal for any feature geometry or magnitude distribution, and the same conditions yield the other results above. Tang et al. and Cui et al. (2026) explicitly model the encoder, which we do not, but they prove recovery results only for independently sampled sparse features. Our conditions require no assumptions about how the data were generated, which is what lets us analyse hierarchical, correlated and dense structure, and test trained SAEs directly on real activations. We highlight our differences to Tang et al. in our response to reviewer dmvY, while Cui et al. discuss hierarchy only informally (their App. B), noting that their recovery theorems need non-overlapping features, which a parent and its children are not, and suggesting that this may partly explain absorption. Our work makes this argument precise, without assuming sparse independent features. To our knowledge, no previous theory accounts for this range of SAE phenomena in a single framework or provides a diagnostic that can be checked on trained SAEs. Hindupur et al. (2025) study the complementary question of what different encoders can represent; combining the two approaches is a natural next step.

We now make more of this motivation explicit in the paper, in Section 3 and the discussion, to try and make the value of our work clearer, and we include a more thorough comparison to existing SAE theory literature (Section 4 and Appendix G).

W3, Q1: The role of the data generator, and the toy experiments

This is an important clarification. The conditions hold for every dataset: they are constraints that any optimum of the dictionary learning loss, eq. (1) must satisfy, whatever the data distribution. The examples in the paper apply them to particular distributions of interest, such as hierarchical, rank-one and categorical data. Any other generator, whether strongly overcomplete, Zipfian or otherwise, can be analysed the same way, by asking whether a candidate set of features can satisfy the conditions.

So, to answer Q1 directly, a Zipf law would not change the theory, only the data it is applied to. Frequencies do enter some of our conditions quantitatively, for example in appendix C governing the allowed range of cosine similarities between mutually exclusive features, but the main effects we explore are all regarding the relationship between the joint distribution of variables and their identifiability.

We now state in the introduction that the conditions can be applied to any dataset, and note in Section 4 that the conclusions for range-independent and categorical features hold independently of the dimensionality or superposition of the data.

W4: Reproducibility

Thank you for raising this. The code for the toy experiments is in the anonymised repository, for example 2_variable_toy_check.ipynb and Single_Variable_Prediction.ipynb. We have added the code for the new experiments, and expanded the README to list which file generates each figure.

W5: Actionable writing

We agree and thank the reviewer. The paper's purpose is to provide tools for interpreting SAE structure. When an SAE behaves strangely, our conditions let one ask whether the behaviour is imposed by the dictionary-learning objective, allowing interpretation. This question frames the abstract and introduction, and we have now added a short practical takeaway to the end of each section.

Q3: Sun et al. and the 50% figure

Apologies, this is a real error on our part and thank you for catching it. In their introduction Sun et al. paraphrase the results of earlier papers and present dense latents as those that activate on “10% to 50% of tokens”, which we then used. As you say, they report occasional latents that activate on >50% of tokens. We would attribute this not to sampling insufficiencies but to differences in the models: we study L1 dictionary learning, Sun et al. measure TopK and JumpReLU SAEs. Further, their TopK SAEs are trained on OpenWebText, but densities are measured on C4. As we showed in our original submission, these properties apply to the data on which the SAEs are trained. We have couched this more carefully in our new submission. Instead, we view our work as providing an explanation for why in these kinds of models nonnegativity might lead to dense antipodal features whose exact properties are encoder dependent.

We are sorry for the delay in posting this rebuttal and hope you find time to engage further with our work.

Thank you for your answer, I took some time to read the paper alongside your rebuttal. It clarifies my view of the paper to some extent, but unfortunately I found it not resolving the weaknesses I identified. I see the paper as genuine contribution to the community, and I support the acceptance of the paper; I'm nevertheless see a gap to be filled on empirical side of it and from the generalization point of view, therefore keep my current overall assessment.

In case of further development of this line of work, I would very much appreciate seeing more narration elements like "this is how it behaves in the wild, and this is what we say about it, and here is the validation of our theory on not yet seen setups, and here is why it is important", without too many simplifying assumptions and with orientation towards practice. In case of acceptance, I believe that some things could be done to enhance the narrative and positioning of the camera ready: first, making a glimpse towards further generalization of a theory, covering setups today practice cares about (other SAEs, relaxing assumptions etc.), just so the reader can understand where the development would go and how explanations might change, at least in a form of expectations rather than developed thoughts; second, enhancing the experimental side of it by showing not only illustrative examples but the overall behavior of the SAEs at scale and in different conditions - some of them will demand explanations and motivate the theory, some of them would make explanation inevitable and support the theory. It is for you to decide what would really enhance the paper and will not disorganize it, and I hope I helped improve your work. Good luck with final decision.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Comment

We thank the reviewers for their time. Reviewers appreciated the clarity of the writing and presentation, and the number of phenomena we were able to explain within our framework and we were pleased to see the paper doing what we see as its job, helping to explain why sparse representations display the structure they do.

There were two shared concerns, beyond those we answer directly in comments to reviewers:

First, our theory is specific to L1 dictionary learning, while real SAEs vary in their encoder architecture and nonlinearity. We agree that the theory says little directly about these choices. We argue, however, that it is a useful idealisation of the ingredient every SAE shares, fitting the data as a sparse linear combination of learnt directions. This is shown both by the tractable conclusions it reaches (hierarchical splitting, residual structure, antipodal pairs, and wide convex limit) and by its predictions holding beyond the idealised setting (e.g. figure 2, or a new appendix G where we show our hierarchical conclusions hold for architectures other than L1 dictionary learning). As such, the theory shows how the core assumption of the linear representation hypothesis, that data are a sparse linear combination of features, is by itself enough to shape the representations extracted under that assumption.

Second, there was concern about theoretical novelty: that there might be alternative explanations for feature absorption we did not account for, and that other work might already cover our results. We have significantly extended the discussion of existing SAE theory, in Section 4 and the new Appendix G. In these sections we discuss, first, that our approach is much more general than its closest predecessor, Chanin et al. 2025, and we argue this is what allows us to explain a breadth of phenomena rather than building one toy model for each object of interest. And second, our approach does not make the limiting data assumptions of the other most relevant papers, Tang et al. and Cui et al., which assume independent, very sparse features. This is what lets us study hierarchical structure directly, which do not fit this assumption. Indeed, Appendix G shows that the alternative explanation of hierarchical absorption, based on spurious partial minima and raised by reviewer dmvY as a potential alternative explanation, is not sufficient: in the full loss, the un-absorbed dictionary is not a minimum at all for L1 dictionary learning or L1 SAEs.

Overall, we believe we can largely answer the reviewers' concerns. At the end of the day, some of this is a value judgement: is our theory sufficiently rich to be useful? Of course, ideally it would be less toy, and would explain yet more interesting behaviour. We have what we have, it is a significant extension beyond the literature, and it explains a suite of otherwise potentially puzzling phenomena. We believe this is sufficient to warrant acceptance.

Revisions are highlighted in purple in the main document, and we apologise for our delay responding that limited the interaction time possible with reviewers.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.