acceptodds
Under review as a conference paper at ICLR 2027

Symmetry, Not Dimension: When Sparse-Coding Atoms in a Feature Plane Are Identifiable

Abstract

Sparse autoencoders trained from different seeds disagree, which concurrent work attributes to feature manifolds of dimension ; in the sharpest analysis, from an isotropy assumption. For a feature plane, the sparse-coding objective under a rigid rotation of its atoms is a circular cross-correlation between the angular density and a configuration-dependent cost kernel, flat along the rotation iff the density's relevant Fourier coefficients vanish, else pinned with computable curvature. Under , dimension is necessary for that flatness, not sufficient. On densities from models' own rings, twelve-seed dictionaries meet predictions the density condition makes and symmetry cannot: agreement at an installed harmonic (strongest in of Gemma-2-2B tilts at no symmetry explains; chance ), and its loss when one Fourier coefficient is removed from a spread density, including the number circles of Llama-3.1-8B and Pythia-6.9B, but not from clustered rings. Only ReLU+ seeds reliably sit near their landscape's minimum; TopK's in-plane optimum at is flat on full-support densities. On released dictionaries and our own, ablating a dictionary's plane for a concept transfers across dictionaries in all cells, its most selective latent in , failing where dictionaries disagree on its item. Making the data exactly symmetric in one plane removes the seeds' agreement on its atoms (matched atoms apart) and reinstalling one harmonic restores it; a plane found in natural text without probes transfers too. Claims should attach to the plane, not the atom. Code, records and some trained dictionaries are in the supplement; all will be public at https://github.com/xxx/xxx upon acceptance.

Primary area:
interpretability and explainable AI
Keywords:
sparse autoencoders, dictionary learning, identifiability, feature geometry, rotational symmetry, mechanistic interpretability, circular convolution

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_sjS6OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper is broadly concerned with whether sparse autoencoders (SAEs) learn a canonical set of dictionary atoms when there exist feature manifolds in activations, under different distributions that activations may be sampled from on those manifolds. The authors mostly consider activations in a 2d plane, and are interested in the conditions under which there is a unique optimal set of dictionary directions under different distributions that activations are sampled from within that plane ρ(θ)\rho(\theta). In particular though, they study a more narrow question, which is: for a fixed set of dictionary directions, whether rotating those directions by φ\varphi changes the dictionary learning loss. They find a particular criterion for when rotations by φ\varphi won't change the loss, which is defined in terms of both ρ(θ)\rho(\theta) and the arrangement of the dictionary atoms. Interestingly, this criterion says that the activations need not be isotropic (ρ(φ)\rho(\varphi) constant) for the dictionary learning loss to not depend on the angle of the atoms. They do some experiments training SAEs on synthetic data to show this analysis applies to real SAEs. My overall takeaway from the paper is that the manner in which sparse autoencoders (or dictionary learning algorithms generally) "tile" feature manifolds depends on the particular way that activation vectors are distributed on that manifold in a somewhat nuanced manner.

Critical strengths
  • The broader question motivating this work, of how sparse autoencoders behave when there exist feature manifolds in activations, is interesting and topical.
  • I appreciate the mathematical approach that the authors take to this question. They define a clear setup and prove a result that was cool/surprising to me about how dictionary learning behaves in the presence of manifolds (though see caveats below).
Critical weaknesses
  • Overall, I found the paper to be very difficult to read in a manner that I think is characteristic of AI systems being used extensively in the writing process. The abstract and intro are quite dense. The whole paper is also extremely long. I know that these are not grounds for rejection since AI use was properly disclosed by the authors, but I feel I must point out that this has substantially degraded my experience reading the paper. I had to rely extensively on AI systems myself in order to understand any of the paper's results and contributions. I really wish that the authors had spent more time on the writing/presentation of their results, and overall I recommend clear rejection on this basis. I will not describe all places where the paper is unclear, but I'll highlight some phrases that are blant, below. These all evoke jargon that is not defined yet and so are meaningless to readers when they are read:
    • "agreement at an installed harmonic (strongest in 28 of 35 Gemma-2-2B tilts at q≥2 no symmetry explains; chance 8.8"
    • "cross-seed jitter tracks the density’s circular moment over an 8.6× jitter range"
    • "has its objective deepen from flat to pinned as that symmetry is broken, in the representation (5/5 defect levels) and by dose (3/3; seeds unresolved)"
    • "Trained seeds lock when the density says."
  • The authors prove a relatively narrow result about whether dictionaries are canonical in the presence of feature manifolds. In particular, they show that for a fixed dictionary (in 2d), rotating those atoms doesn't change the dictionary learning loss exactly when a certain criterion is met that has to do with the relationship between the geometry of the dictionary atoms and the probability density of the activations ρ(θ)\rho(\theta). Their result implies that there are some dictionaries where even nonisotropic distributions over activations achieve the same dictionary learning loss (reconstruction + sparsity) no matter the orientation of those dictionary atoms. This is cool! However, if I understand correctly, in general the dictionaries that have this property are not the dictionaries that actually minimize the dictionary learning loss when activations are distributed in this way. So an optimal solution from something like an SAE might actually be canonical, even though this symmetry does indeed exist in the space of dictionaries. The authors acknowledge this on line 059, although I also find the language they use for this caveat to be inscrutable "a flat rotation at an optimum is a continuum of optima, and with free atoms we found one only under isotropy (for L1) and for TopK at k ≥ 2..." In simple terms, this means that the only optimal solutions for a dictionary learner that are genuinely rotation-invariant (and thus non-canonical) are when the activation distribution is indeed isotropic, which I think makes the paper's claims a bit less interesting.
Questions for authors
  • Have I correctly understood the nature of your most important contributions, as you see them? There are other results in the paper that are more empirical, including ablations on the days of the week representation. I'd be happy to have some discussion of these if you think they should change my outlook on the paper.
  • Related to the above, in your own words, what do you think are the most important takeaways from the paper? What should the interpretability community learn from your work? What have you learned from doing this work? Please write this yourself in plain language.
Reviewer_rhovOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper is studying why sparse autoencoders trained from different seeds can learn different individual features. The authors primary argument is that feature dimensionality alone is not enough to explain rotational ambiguity. Instead the activation distribution and sparse coding objective determine whether particular orientation are preferred. So for a fixed atom configuartion in a two dimensional feature subspace, they provide a Fourier based conditon for whether rotating the atoms changes the loss. They test this idea with experiments using controlled data, various coding objectives, feature planes from pretrained language models, and cross-dictionary interventions. An interesting finding is that the interventions on concept associated planes reproduce across dictionaries more consistently than interventions on their most selective individual latents.

Critical strengths
  • The paper links features identifiability with the activation density and the cost kernel. This gives a criterion for when rotation is flat and a way to quantify how strongly it is pinned, rather than treating all feature planes as equally ambiguous.
  • Authors test the idea across controlled data, pretrained models feature rings, coding rules, and dictionary transfer interventions. They are honest about the failures as well. That makes the claim more informative.
Critical weaknesses

W1: Although the introduction identifies SAE reproducibility as the motivating problem, the manuscript moves quickly into technical details without properly explaining what problem the paper solves and why solving that problem is important. The presentation moves among the rotation landscape of a fixed configuration, agreement among trained atoms, and reproducibility of interventions, despite these being distinct questions.

W2: The Fourier criterion characterizes only the rigid rotation of a specified dictionary configuration. The paper states that this is “one mode of the solution set” and that the criterion does not determine which configuration training selects or whether an amortized encoder reaches a preferred orientation. The empirical limitations make this distinction important. Lock order prediction remains weak even when incorporating the trained kernel which only provides a small improvement (Section 5.1), the harmonic removal test fails on clustered rings, and some coding rules disagree despite pinned exact code optima (Section 4.3). These results do not invalidate the mathematical identity, but also doesn't establish it's empirical predictiveness for realistic SAEs.

Questions for authors
  • Can the authors provide more detail on the situation that if you have a naturally trained SAE, can the proposed criterion predict in advance which feature planes will show strong atom-level agreement across random seeds?
  • Does the proposed criterion predicts cross seed atom agreement better than decoder geometry, or a simple density statistic?
  • Across the independent seeding, what advantage remains for planes after semantic item matching and comparison with the similarly constructed rank one direction? A consolidated account would clarify whether the practical gain concerns dimension, aggregation, or semantic alignment.
  • What evidence shows that rotational disagreement causes atom level interventions to transfer poorly? Do the existing experiments demonstrate a corresponding change in intervention reproducibility when atom alignment is removed and restored?
Reviewer_5vJQOpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper studies rotational ambiguity of SAE dictionary atoms within two-dimensional feature representations. The authors argue that dimensionality alone does not determine whether dictionary orientations are identifiable, and derive a Fourier condition relating rotational invariance of the sparse-coding objective to the angular activation distribution. The experiments test this relationship on synthetic distributions and periodic concepts in language models. The paper also reports that concept-associated subspace interventions are more reproducible across SAEs than individual latent interventions.

Critical strengths
  1. The paper makes a useful conceptual distinction between dimensionality and symmetry in SAE feature identifiability. The Fourier characterization provides a precise condition for rotational ambiguity and shows how the angular structure of the data interacts with the sparse-coding objective. Although the theoretical contribution is limited in scope, it offers a concrete way to study a particular source of feature instability.

  2. The paper provides meaningful empirical evidence for the role of symmetry in SAE instability. The controlled harmonic interventions, including experiments on representations from pretrained language models, show that manipulating distributional symmetry can substantially affect cross-seed atom agreement.

Critical weaknesses
  1. My main concern is with the novelty and significance of the contribution. The paper makes a valid distinction between dimensionality and symmetry, but I find the resulting characterization rather limited. The main result expresses the sparse-coding loss under rigid rotation of a fixed dictionary as a circular cross-correlation, with the flatness condition following from its Fourier decomposition. The role of distributional symmetry in representation identifiability is already well established, and the paper discusses this literature. What remains new is a specific characterization of rotational pinning in a two-dimensional sparse-coding setting. The authors are clear about this scope, but I am not convinced that it provides much additional insight into the identifiability of learned SAE features, particularly since freely optimized dictionaries can change their relative atom configurations independently of the rigid-rotation mode being analyzed.

  2. I am also unconvinced that the experiments establish the practical importance of the proposed mechanism. The controlled harmonic interventions provide evidence that symmetry can influence cross-seed agreement, including on representations extracted from pretrained models. However, this is different from establishing that rotational ambiguity explains a meaningful fraction of the instability observed in naturally trained SAEs. The strongest correspondence between objective-level pinning and learned dictionary agreement occurs for ReLU+L1, while TopK and JumpReLU show substantially weaker results. The authors also report failures in predicting the dominant agreement harmonic, an unresolved contradiction with the observed L0 trend, and a rotation-landscape diagnostic that fails at production scale. The production-scale experiments provide interesting evidence that concept subspaces are more reproducible than individual SAE atoms. However, as the authors acknowledge, these experiments do not establish whether the proposed rotational mechanism explains the observed atom instability. Overall, I find the empirical evidence insufficient to support the broader significance of the contribution.

  3. The paper needs major revision in its presentation. I found it unnecessarily dense and difficult to follow. The writing repeatedly packs numerical results, qualifications, and appendix references into long paragraphs, often at the expense of a coherent argument. The result reads more like an exhaustive compilation of findings than a carefully developed scientific paper. The authors should prioritize a clear account of the contribution and its supporting evidence instead of overwhelming the reader with statistics. In its current form, the presentation makes an already narrow contribution considerably harder to assess.

Questions for authors
  1. How much of the instability in naturally trained SAEs can be attributed to the proposed rotational mechanism? The controlled interventions establish that symmetry can affect atom agreement, but it remains unclear how important this mechanism is relative to other sources of instability. Is there evidence from the naturally occurring representations that could help quantify its contribution?

  2. What predictive value does the Fourier criterion have for independently trained SAE dictionaries? The paper reports that the dominant agreement harmonic cannot be reliably predicted, even when accounting for the learned cost kernel. What can the framework reliably predict about learned atom configurations without experimentally modifying the underlying distribution?

  3. How should the results for TopK and JumpReLU affect the paper’s conclusions? The relationship between objective-level pinning and cross-seed agreement appears considerably weaker for these objectives. Does this indicate that rotational pinning is insufficient to explain learned feature stability, even when the relevant symmetry is known?

  4. What additional insight does the theoretical analysis provide beyond existing symmetry-based identifiability arguments? I understand that the paper deliberately restricts its analysis to rigid rotations, but I would appreciate a more precise account of what the Fourier characterization establishes about SAE feature identifiability that was not already understood.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.