Incorporating Hierarchical Semantics in Sparse Autoencoder Architectures

Kiho Park; Mark Muchane; Sean Richardson; Victor Veitch

REVIEW 3 major objections 6 minor 1 cited by

Incorporating Hierarchical Semantics in Sparse Autoencoder Architectures

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A sparse autoencoder with an explicit two-level semantic hierarchy improves reconstruction and interpretability on LLM activations while cutting compute.

desk verdict A clever hierarchical SAE architecture whose headline reconstruction gains are not yet separated from raw capacity—needs matched-capacity baselines before the hierarchy claim lands. read the letter →

arxiv 2506.01197 v1 pith:MXXHK3I7 submitted 2025-06-01 cs.CL cs.AIcs.LG

Mark Muchane , Sean Richardson , Kiho Park , Victor Veitch This is my paper · ORCID

classification cs.CLcs.AIcs.LG

keywords hierarchicalsparseautoencodersdictionarylearningLLMinterpretabilityfeaturesplittingsemantichierarchymixture-of-expertsreconstruction-interpretabilityfrontierGemma2-2B

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Sparse autoencoders (SAEs) learn a small set of human-readable features that reconstruct a model's internal activations, but the learned features usually ignore the fact that concepts come in hierarchies—'corgi' is a kind of 'dog'. This paper introduces the Hierarchical Sparse Autoencoder (H-SAE), an architecture that replaces a flat dictionary with a top level of concept detectors (experts), each attached to a low-dimensional subspace and a small SAE of sub-features. The design builds on the geometric finding that a categorical concept lives in a language model as a parent feature plus a polytope of child features, so a low-level representation is the parent's vector plus a child-in-context vector. Trained on Gemma 2-2B residual streams, the H-SAE improves reconstruction, reduces feature splitting and cross-language duplicate features, keeps or improves interpretability, and adds little computational overhead because only activated experts run their sub-SAE. The paper's conclusion is that semantic hierarchy can be learned by an SAE and that exploiting it improves the reconstruction-interpretability frontier.

What carries the argument

The load-bearing object is the two-level mixture-of-experts decomposition defined in Eq. (3.1): a standard top-level SAE chooses a sparse set of parent features, and each parent owns a down-projector $\Pi^{\text{down}}_j$ into a subspace of dimension $s$, a low-level SAE $SAE^j_1$ with TopK1 (only the largest sublatent is kept), and an up-projector $\Pi^{\text{up}}_j$ back to the full space. The mechanism enforces hierarchy by routing: a sublatent can only activate when its parent is in the top-$k$, matching the prior geometric result that 'corgi' is 'dog' plus 'corgi in the context of dog'. Because only $k$ experts are activated, the forward cost is dominated by the top-level encoder term $O(m_{\text{top}}d)$ instead of scaling with the full effective dictionary of $m_{\text{top}}\times m_{\text{low}}$ atoms, which is the source of the computational efficiency gain.

What would settle it

Take a residual-stream position where two known child concepts of one expert should both be strongly active (for example a token that is simultaneously about 'divorce' and 'engagement' under a 'marriage' expert) and compare H-SAE reconstruction error against a variant that keeps the top two sublatents per expert; if the top-1 model loses one child or its error spikes while the top-2 model does not, the single-vertex assumption is the cause.

Watch

Extended reading notes

Core claim

The central discovery is that an SAE whose architecture explicitly mirrors hierarchical concept geometry can learn that hierarchy and outperform a flat SAE of comparable effective size. On a given input, the model selects the $k$ strongest top-level features, reconstructs their contribution $Dz$, and for each selected expert projects the input into a low-dimensional subspace, runs a one-sublatent SAE, and adds the projected child reconstruction back: $H\text{-SAE}(x)=\sum_{j\in\text{TopK}_k(z)}(z_j d_j+\Pi^{\text{up}}_j SAE^j_1(\Pi^{\text{down}}_j x))$. This decomposition implements the paper's premise that a child concept is represented as the parent concept plus the child in the context of the parent. Experimentally, the H-SAE with 64 sublatents per expert is on par with a standard SAE with 32,000 top-level features while costing about a quarter of the compute, and it shows less feature absorption and more language-shared features than the baseline. Interpretability checks show coherent parents such as 'airports' with children 'US airport', the token 'airport', and 'airport size', and the authors conclude that hierarchy is learned rather than imposed.

Load-bearing premise

The architecture assumes a subordinate concept lives at exactly one vertex of its parent's polytope, so each active expert keeps only one sublatent; if one input needs several children of the same parent at once, the model has no way to express it.

Editorial extensions

If this is right

A 64-sublatent H-SAE matches a 32,000-feature standard SAE on reconstruction at about one quarter of the compute, so hierarchy buys fidelity and scale at the same time.
The reconstruction gain appears across all tested model sizes: H-SAE beats the standard SAE on one minus explained variance and one minus CE loss at 8k, 16k, and 32k top-level latents.
H-SAE features show less absorption on the first-letter probe and lower cross-lingual set differences, indicating that the learned atoms stay less redundant and more composable.
Because only active experts are used in each forward and backward pass, per-batch memory and compute scale with activated parameters rather than total parameters.
Visualized experts are semantically organized, such as a 'marriage' parent with 'divorce', 'engagement', and 'marriage' sublatents, so the hierarchy provides a readable map of concept relationships.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper's TopK1-per-expert choice means an input needing two sibling sublatents at once cannot be represented; testing a TopK_r variant with $r>1$ would directly reveal whether this restriction is a hidden capacity limit.
The unembedding ablations show that whitening is necessary for meaningful features; applying the same causal-inner-product whitening to residual streams, which the paper does not do, is a natural way to test whether the reported gains grow further.
If the learned parent-child structure is causal, parent-level interventions should affect all children while child-level interventions stay local; the paper does not run such intervention tests, but they would be a sharp validation of the hierarchy.
Since the forward cost is dominated by the top-level encoder, a three-level version of the architecture should be nearly as cheap as the two-level version; that extension is implied by the design but not explored.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: the H-SAE is a genuinely new architecture and the paper is worth engaging, but the central empirical claim—that hierarchy, not extra capacity, drives the reconstruction gains—is not currently supported. The stress-test note is right: Fig. 3 measures model size in thousands of top-level latents, while H-SAE also has 16 or 64 sublatents per expert. At 16k top-level, H-SAE-64 has roughly a million decoder atoms versus 16k for the flat SAE. Of course it reconstructs better. The efficiency claim suffers the same problem: the O(m_top d) framing hides the fact that the effective dictionary is m_top * m_low. Without a flat SAE matched on total atoms or FLOPs, or a Switch SAE with the same expert and sublatent counts, you cannot attribute the improvement to hierarchy.

That said, there is a real contribution here. Translating the parent-feature plus polytope geometry from Park et al. into a mixture-of-experts SAE with per-expert sublatents is new, and it is not just an objective tweak like Matryoshka or EWG. The architecture is clean, the paper is honest about what it does not know, and the qualitative examples (Figures 1, 4, 6, 7) do show plausible two-level structure. Code is provided. The authors also explicitly admit the orthogonality and L1 terms are not necessary for the reported effects—that is refreshing, even if it undercuts the stated objective.

The softer spots: no error bars anywhere, so the margins in Figures 3 and 5 might be noise. The TopK1 restriction per expert (Eq. 3.1) is a real expressivity limit and is not ablated; the paper justifies it only by the geometry claim. The cross-lingual redundancy result is suggestive but uses a Gemini translation pipeline with no discussion of translation artifacts. And the hierarchical-semantics evidence is mostly qualitative; the systematic absorption test is indirect.

The citation pattern is fine. The two self-cited prior papers do establish the geometric motivation, and the paper does not lean on them dishonestly.

Verdict: this deserves a serious referee and a major revision. The right fix is not cosmetic: add a flat SAE matched on total dictionary size, a Switch SAE matched on expert/sublatent count, and a matched-FLOP comparison. If the hierarchy advantage survives those controls, this becomes a solid contribution. If it does not, the paper still documents a useful efficiency-expressivity trade-off, but the title claim needs to change.

I would take this to reading group and cite it as related work, but I would not yet cite it as evidence that hierarchical structure improves reconstruction.

Referee Report

3 major / 6 minor

Summary. The paper proposes a hierarchical sparse autoencoder (H-SAE) in which a top-level SAE routes inputs to expert-specific low-level SAEs, aiming to explicitly represent the hierarchical structure of concepts. The authors apply the architecture to Gemma 2-2B residual streams and report that it improves reconstruction (explained variance and downstream CE loss) and interpretability (feature absorption and cross-lingual redundancy) relative to a standard TopK SAE, while also providing computational efficiency gains. The paper includes qualitative visualizations of learned hierarchies, ablations on unembedding data, and public code. The central claim is that modeling hierarchy itself improves the reconstruction-interpretability frontier, but the main experimental comparison is confounded by differences in total dictionary capacity.

Significance. If the hierarchy-specific benefits were established, the H-SAE would be a meaningful contribution to SAE interpretability and scalability, providing an architecture that jointly captures coarse and fine concepts and reduces compute. The paper ships code and includes some ablations, which is a strength. However, the current evidence does not separate the effect of hierarchy from the effect of a much larger effective dictionary realized through cheap routing, so the significance of the empirical contribution is presently conditional on additional controlled comparisons.

major comments (3)

[§4, Figure 3] The reconstruction comparison in Figure 3 is not capacity-matched. The H-SAE with 64 sublatents per expert at 16k top-level latents has 16k + 16k×64 = 1,040,000 total decoder atoms, while the standard SAE baseline at 16k has only 16,000 atoms; the x-axis labels both as '16k' because it counts only top-level latents. No flat SAE with m = mtop×mlow features, no Switch SAE [Mud+24] with matched expert/sublatent counts, and no matched-FLOP baseline is reported, so the observed reconstruction gains cannot be attributed to the hierarchical structure rather than to the 16–65× larger effective dictionary. This confound is load-bearing for the paper's central claim that hierarchy itself improves reconstruction.
[§3.1, §4] The computational-efficiency claim is expressed relative to a baseline with only mtop features. The cost formula in Section 3.1 is dominated by O(mtop d), but the H-SAE actually has mtop × mlow atoms; the sentence in Section 4 that an H-SAE with 64 sublatents is 'on par with the standard SAE with 32 top-level features, despite having 1/4 the compute cost' compares an architecture with roughly 32×64 = 2048 effective atoms against one with 32. A matched-capacity flat SAE or a Switch SAE with the same total number of atoms/FLOPs would be required to substantiate the efficiency benefit, and Appendix A.3 also states that no detailed wall-time analysis was conducted.
[§3.1, Eq. (3.1)] The low-level SAE uses TopK1 per expert (Eq. 3.1 and Algorithm 1), so an input that genuinely requires two or more children of the same parent cannot be represented. This is a strong inductive bias justified only by the geometry claim that a subordinate concept is a single vertex of the parent polytope. The paper provides no ablation with larger per-expert k (e.g., TopK2 or TopK4) and no quantitative test of whether such inputs occur in the residual stream distribution. This limits the expressivity of the architecture and should be tested before the hierarchy claim can be taken as general.

minor comments (6)

[All figures] Figure 3 has no error bars or multiple seeds; all comparisons appear to be single runs, which makes it impossible to assess the variability of the reported gains.
[Figure 5] The cross-lingual set-difference results in Figure 5b are reported as point estimates without variance or significance testing; the differences (e.g., 8.448 vs 9.772) may not be meaningful without such information.
[Figure 5a] The metric 'Mean Absorption Fraction Score' is not formally defined in the text; please provide a precise definition or a reference for its computation.
[Appendix B] The ablations of the orthogonality and L1 losses are performed on the unembedding matrix and are qualitative; the authors explicitly state that the losses do not seem necessary for hierarchy, yet they are retained in the main experiments. Please clarify whether the main results depend on these losses or reconcile the ablation with the final model choice.
[Figure 3, general] The x-axis label 'Model Size (Thousands of Top-Level Latents)' is misleading because it conflates the standard SAE's full dictionary with the H-SAE's top-level only; consider reporting total dictionary size or relabeling the axis.
[Throughout] There are several typos: 'auxillary' in Appendix A.3, 'Wheras' in the Figure 4 caption, and 'shitzu' in the Introduction (likely 'shih tzu').

Circularity Check

2 steps flagged · score 4.0 of 10

Hierarchical activation is imposed by the architecture, and the key geometric motivation is carried by self-citations; the empirical reconstruction and cross-lingual results remain substantive, though partly confounded by unmatched dictionary size.

self definitional [Section 3.1, Algorithm 1 and Eq. (3.1); claim in Section 4.3]
"for j∈K do ▷ Only process activated experts ... zj← LeakyReLUα(Ejxsub_j )"

Under Eq. (3.1) and Algorithm 1, a low-level sublatent is computed only after its expert j has already been selected by the high-level TopK operation, so sublatent activation is definitionally conditional on parent-expert activation. The paper's Section 4.3 conclusion that 'hierarchical semantics are clearly emerging'—with low-level features active only when the high-level topic is relevant—is therefore not an independent discovery of a two-level activation hierarchy: that nesting is built into the forward pass by construction. What is genuinely learned is the semantic content and grouping of the sublatents within each expert, not the parent-child gating itself.
self citation load bearing [Section 3.1, 'Hierarchical Geometry'; Section B.1]
"Our main inspiration follows Park et al. [Par+24], who find that the representations of categorical concepts in language models have a specific geometric structure."

The architecture is explicitly constructed to instantiate the geometric account of [Par+24], and the whitening preprocessing in Section B.1 is justified by [PCV24]. Both cited works share authors with the present paper (Park and Veitch), and neither is independently verified within this manuscript. The design choice—parent feature plus per-parent low-rank subspace and child vertices—thus leans on a self-citation chain. However, the reconstruction curves and cross-lingual redundancy measurements are new empirical results against an external model and benchmark, so this self-citation is load-bearing for the architecture's motivation but does not by itself determine the main empirical findings.

full rationale

The paper contains no fitted-parameter-renamed-as-prediction circularity: the H-SAE is trained on Gemma 2-2B residual streams, and the reported reconstruction, absorption, and cross-lingual divergence numbers are measured externally rather than derived from the training objective. The strongest circularity is partial and structural: because Algorithm 1 only executes an expert's low-level SAE after that expert has been selected by the high-level TopK, the observed parent-child activation dependence is guaranteed by the forward pass, so citing it as evidence that 'semantic hierarchy can be learned' is partly self-definitional. The genuinely empirical component is whether the learned expert/sublatent groupings are semantically coherent, which the paper supports with qualitative visualizations. The self-citations [Par+24] and [PCV24] motivate the architecture and preprocessing but are not independently checked here; they lower the evidentiary independence of the motivational chain without voiding the external reconstruction results. The absence of a flat SAE matched on total atoms or FLOPs is a real confound for the causal claim that hierarchy, rather than added capacity, drives the improvements, but a missing control is a correctness concern, not circularity. Overall, the central empirical claims are not reduced to their inputs by construction, so the score is 4 rather than higher.

Assumptions & free parameters 3 free parameters · 3 assumptions · 1 invented entities

The central empirical claims rest mainly on a domain assumption about LLM concept geometry (from the authors' own prior work), an untested TopK1 modeling restriction, and the use of Euclidean geometry on residual streams. The three loss weights (beta, lambda1, lambda2) are hand-chosen and partly admitted to be nonessential.

free parameters (3)

beta (top-level reconstruction weight) = 0.1
Weights the auxiliary top-level reconstruction loss Ltop_recon in Eq. 3.3; chosen by hand with the stated reason that the top-level SAE alone has less capacity than the full H-SAE.
lambda1 (orthogonality penalty) = 0.1
Weights the bi-orthogonality penalty in Eq. 3.4; the authors' own ablations (Appendix B) find it unnecessary for interpretability or hierarchy, yet it is kept in main runs to reduce dead latents.
lambda2 (L1 sparsity) = 0.001
Weights the sparsity penalty on non-top-k latents; ablations show it is not necessary for the reported hierarchical or interpretability effects.

assumptions (3)

domain assumption Categorical concepts in LLMs are represented by a parent feature plus a low-rank subspace containing a polytope of child features (Park et al., 2024).
Invoked in Section 3.1 to justify the three-part H-SAE design; the prior work is by two of the present authors, so the architecture inherits any limitations of that theory without independent verification here.
ad hoc to paper A subordinate concept is captured by exactly one vertex of the parent polytope, so one active sublatent per expert suffices (TopK1).
Eq. 3.1 and Algorithm 1 set the per-expert low-level SAE to TopK1; contexts that require multiple simultaneous child features of the same parent cannot be represented, and this restriction is not tested against alternatives.
domain assumption The Euclidean inner product is the appropriate geometry for reconstruction on residual stream activations; no whitening is applied.
Losses in Eq. 2.2 and 3.3 use Euclidean distance on normalized residual vectors; Appendix B.3 shows Euclidean geometry is inadequate on unembeddings and requires whitening, but the paper does not test whether whitening changes the residual-stream results.

invented entities (1)

Expert-attached sublatent hierarchy (top-level experts with per-expert low-level SAEs and projection operators)
purpose: Models the parent-child concept structure so that child features activate only when their parent is active, improving reconstruction and interpretability.
The two-level structure is built into the architecture rather than discovered; supporting evidence is the paper's own visualizations and metrics, with no external falsifiable prediction or independently verified handle.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Incorporating Hierarchical Semantics in Sparse Autoencoder Architectures." pith.science (2026). https://pith.science/paper/MXXHK3I7

@misc{pith2026250601197,
  author       = {Pith},
  title        = {Pith review of: Incorporating Hierarchical Semantics in Sparse Autoencoder Architectures},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MXXHK3I7}},
  note         = {Machine review of arXiv:2506.01197}
}

read the original abstract

Sparse dictionary learning (and, in particular, sparse autoencoders) attempts to learn a set of human-understandable concepts that can explain variation on an abstract space. A basic limitation of this approach is that it neither exploits nor represents the semantic relationships between the learned concepts. In this paper, we introduce a modified SAE architecture that explicitly models a semantic hierarchy of concepts. Application of this architecture to the internal representations of large language models shows both that semantic hierarchy can be learned, and that doing so improves both reconstruction and interpretability. Additionally, the architecture leads to significant improvements in computational efficiency.

Figures

Figures reproduced from arXiv: 2506.01197 by the authors.

**Figure 1.** A Hierarchical Sparse Autoencoder architecture learns human interpretable hierarchy. Each box shows 5 of the strongest activating contexts for the feature, all tokens underlined activate the feature while the bold/text weight shows the relative strength of activation. For example, a "marriage" high-level feature and "divorce", "engagement", and "marriage" sublatents. Note that “48 sati svadba" is a Serbian wedding r… view at source ↗

**Figure 2.** The structure of the Hierarchical Sparse Autoencoder architecture. The design combines a top-level encoder-decoder (upper path) that captures general concepts with expert-specific autoencoders (lower paths) that model refined features. For a given input, only a sparse subset of experts is activated, enhancing computational efficiency while maintaining expressive power. 3.1 Architecture Hierarchical Geometry Our main… view at source ↗

**Figure 3.** H-SAEs have better reconstruction performance as measured by explained variance and the downstream CE loss of a Gemma 2-2B using SAE-reconstructed activation vectors. term actually achieves this goal. Nevertheless, we do observe empirically that adding this term does an excellent job of mitigating the “dead atom” phenomena where many features are never used in reconstructions. We use this instead of the auxiliary de… view at source ↗

Figures from the paper (18 more)

**Figure 4.** Figure 4: A comparison decomposition of the same context/embedding in the H-SAE architecture and standard SAE. 9 [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

**Figure 5.** Figure 5: H-SAEs exhibit less feature absorption and fewer redundant features across languages. Hierarchical Semantics Finally, we turn to the question of whether the H-SAE is indeed learning hierarchical semantics. As a small-scale test, we start by running the H-SAE on the une…

**Figure 6.** Figure 6: An H-SAE model learns hierarchical semantics on the unembeddings. Most of the time the low-level feature isn’t highly activated. However, for ‘python’ Features 547 & 1867 and ‘Chicago’ Feature 175 the low-level feature is highly activated and relevant (Feature 175 is U…

**Figure 7.** Figure 7: Two more example H-SAE features 12 [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]

**Figure 8.** Figure 8: Ablation studies on the unembedding matrix show no obvious advantage from the orthogonality or ℓ1 regularizers, though we chose to keep them for our final runs on the embeddings. hypothesis, as we find that the unwhitened unembedding matrix does not yield meaningful f…

**Figure 9.** Figure 9: Ablation studies on the unembedding matrix show no obvious advantage from the orthogonality or ℓ1 regularizers, though we chose to keep them for our final runs on the embeddings. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]

**Figure 10.** Figure 10: Ablation studies on the unembedding matrix show no obvious advantage from the orthogonality or ℓ1 regularizers, though we chose to keep them for our final runs on the embeddings. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]

**Figure 11.** Figure 11: Ablation studies on the unembedding matrix show no obvious advantage from the orthogonality or ℓ1 regularizers, though we chose to keep them for our final runs on the embeddings. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]

**Figure 12.** Figure 12: Ablation studies on the unembedding matrix show no obvious advantage from the orthogonality or ℓ1 regularizers, though we chose to keep them for our final runs on the embeddings. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_12.png]

**Figure 13.** Figure 13: Ablation studies on the unembedding matrix show no obvious advantage from the orthogonality or ℓ1 regularizers, though we chose to keep them for our final runs on the embeddings. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_13.png]

**Figure 14.** Figure 14: Ablation studies on the unembedding matrix show no obvious advantage from the orthogonality or ℓ1 regularizers, though we chose to keep them for our final runs on the embeddings. (a) Baseline (b) No Whitening [PITH_FULL_IMAGE:figures/full_fig_p023_14.png]

**Figure 15.** Figure 15: The causal inner product is necessary for the model to learn meaningful features. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_15.png]

**Figure 16.** Figure 16: The causal inner product is necessary for the model to learn meaningful features. (a) Baseline (b) No Whitening [PITH_FULL_IMAGE:figures/full_fig_p024_16.png]

**Figure 17.** Figure 17: The causal inner product is necessary for the model to learn meaningful features. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_17.png]

**Figure 18.** Figure 18: The causal inner product is necessary for the model to learn meaningful features. (a) Baseline (b) No Whitening [PITH_FULL_IMAGE:figures/full_fig_p025_18.png]

**Figure 19.** Figure 19: The causal inner product is necessary for the model to learn meaningful features. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_19.png]

**Figure 20.** Figure 20: The causal inner product is necessary for the model to learn meaningful features. (a) Baseline (b) No Whitening [PITH_FULL_IMAGE:figures/full_fig_p026_20.png]

**Figure 21.** Figure 21: The causal inner product is necessary for the model to learn meaningful features. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_21.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Laguerre Geometry for Interpreting Large Language Models
cs.AI 2026-07 conditional novelty 6.0 of 10

LLM concepts are Laguerre–Voronoi cells; Geometric Lens reads the exact cell of any hidden vector by isolating residual piecewise-linear flow from cross-token attention transport.

Reference graph

Works this paper leans on

25 extracted references · 4 canonical work pages · cited by 1 Pith paper

[2]

Weakly supervised causal representation learning

arXiv: 2203.16437 [stat.ML] (cit. on p. 3). [Bri+23] T . Bricken, A. Templeton, J. Batson, B. Chen, A. Jermyn, T . Conerly , N. Turner, C. Anil, C. Denison, A. Askell, R. Lasenby, Y. Wu, S. Kravec, N. Schiefer, T . Maxwell, N. Joseph, Z. Hatfield-Dodds, A. Tamkin, K. Nguyen, B. McLean, J. E. Burke, T . Hume, S. Carter, T . Henighan, and C. Olah. “Towards ...

work page Pith review arXiv 2023
[5]

arXiv: 2409.14507 [cs.LG] (cit. on p. 1). [Cun+23] H. Cunningham, A. Ewart, L. Riggs, R. Huben, and L. Sharkey. Sparse au- toencoders find highly interpretable features in language models

arXiv
[7]

arXiv: 2406.11944 [cs.LG] (cit. on p. 3). [Eng+25] J. Engels, E. J. Michaud, I. Liao, W . Gurnee, and M. Tegmark.Not all language model features are one-dimensionally linear

arXiv
[8]

Scaling and evaluating sparse autoencoders

arXiv: 2405.14860 [cs.LG] (cit. on p. 3). [Gao+24] L. Gao, T . D. la Tour, H. Tillman, G. Goh, R. Troll, A. Radford, I. Sutskever, J. Leike, and J. Wu. “Scaling and evaluating sparse autoencoders”. arXiv preprint arXiv:2406.04093 (2024) (cit. on pp. 1–3, 7, 8). [Hea+25] T . Heap, T . Lawson, L. Farnik, and L. Aitchison.Sparse Autoencoders Can Inter- pret ...

arXiv 2024
[9]

arXiv: 2503.01822 [cs.LG] (cit. on p. 3). [Joh24] D. D. Johnson. Penzai + treescope: a toolkit for interpreting, visualizing, and editing models as data. 2024 (cit. on p. 15). [Jos+25] S. Joshi, A. Dittadi, S. Lachapelle, and D. Sridhar.Identifiable steering via sparse autoencoding of multi-concept shifts

arXiv 2024
[10]

arXiv: 2502.12179 [cs.LG] (cit. on p. 3). [Kar+25] A. Karvonen, C. Rager, J. Lin, C. Tigges, J. Bloom, D. Chanin, Y. -T . Lau, E. Farrell, C. McDougall, K. Ayonrinde, M. Wearden, A. Conmy, S. Marks, and N. Nanda. Saebench: a comprehensive benchmark for sparse autoencoders in language model interpretability

arXiv
[11]

arXiv: 2503.09532 [cs.LG] (cit. on p. 8). [KG21] P . Kidger and C. Garcia.Equinox: neural networks in jax via callable pytrees and filtered transformations

arXiv
[13]

arXiv: 2408.05147 [cs.LG] (cit. on pp. 1, 3, 7, 8). [Lin+25] J. Lindsey, W . Gurnee, E. Ameisen, B. Chen, A. Pearce, N. L. Turner, C. Citro, D. Abrahams, S. Carter, B. Hosmer, J. Marcus, M. Sklar, A. Templeton, T . Bricken, C. McDougall, H. Cunningham, T . Henighan, A. Jermyn, A. Jones, and J. Batson. On the biology of a large language model. 2025 (cit. o...

arXiv 2025

Show all 25 references

[14]

CITRIS: Causal Identifiability from Temporal Intervened Sequences

arXiv:2306.09643 [cs.LG] (cit. on p. 3). [Lip+22] P . Lippe, S. Magliacane, S. Löwe, Y . M. Asano, T . Cohen, and S. Gavves. “CITRIS: Causal Identifiability from Temporal Intervened Sequences”. en. In:Proceedings of the 39th International Conference on Machine Learning. ISSN: ...

2022 arXiv
[17]

arXiv: 2403.19647 [cs.LG] (cit. on p. 3). [MA25] G. E. Moran and B. Aragam. Towards interpretable deep generative models via causal representation learning

arXiv
[18]

arXiv: 2504.11609 [stat.ML] (cit. on p. 3). [Mud+24] A. Mudide, J. Engels, E. J. Michaud, M. Tegmark, and C. S. de Witt.Efficient dictionary learning with switch sparse autoencoders
[19]

arXiv: 2410.08201 [cs.LG] (cit. on p. 3). [O’B+24] K. O’Brien, D. Majercak, X. Fernandes, R. Edgar, J. Chen, H. Nori, D. Carignan, E. Horvitz, and F . Poursabzi-Sangde.Steering language model refusal with sparse autoencoders

arXiv
[20]

The geometry of categorical and hi- erarchical concepts in large language models

arXiv: 2411.11296 [cs.LG] (cit. on p. 3). [Par+24] K. Park, Y. J. Choe, Y . Jiang, and V . Veitch. “The geometry of categorical and hi- erarchical concepts in large language models”.arXiv preprint arXiv:2406.01506 (2024) (cit. on pp. 2, 4). [PCV24] K. Park, Y. J. Choe, and V ....

2024 arXiv
[21]

arXiv: 2404.16014 [cs.LG] (cit. on p. 3). [Raj+24b] S. Rajamanoharan, T . Lieberum, N. Sonnerat, A. Conmy, V . Varma, J. Kramár, and N. Nanda. Jumping ahead: improving reconstruction fidelity with jumprelu sparse autoencoders

arXiv
[22]

arXiv: 2407.14435 [cs.LG] (cit. on p. 3). [Raj+24c] G. Rajendran, S. Buchholz, B. Aragam, B. Schölkopf, and P . Ravikumar.Learning interpretable concepts: unifying causal representation learning and foundation models

arXiv
[23]

arXiv: 2402.09236 [cs.LG] (cit. on p. 3). [Sch+21] B. Schölkopf, F . Locatello, S. Bauer, N. R. Ke, N. Kalchbrenner, A. Goyal, and Y. Bengio. Towards causal representation learning

arXiv
[24]

Scaling monosemanticity: extracting interpretable features from claude 3 sonnet

arXiv: 2102.11107 [cs.LG] (cit. on p. 3). [SBb22] L. Sharkey, D. Braun, and beren. [interim research report] taking features out of superposition with sparse autoencoders. 2022 (cit. on p. 3). [Tem+24] A. Templeton, T . Conerly, J. Marcus, J. Lindsey, T . Bricken, B. Chen, A. ...

2024 arXiv
[25]

The l1 coefficient is 0.001

• We set the orthogonality penalty and top-level reconstruction coefficients to 0.1. The l1 coefficient is 0.001. For standard SAEs that require an auxiliary loss to prevent dead latents we use a coefficient of 1/30. • We use the adam optimizer with global norm clipping of 0.7...
[2002]

02886 [cs.LG] (cit. on pp. 3, 11). [MSL22] J. Maitin-Shepard and L. Leavitt. Tensorstore for high-performance, scalable array storage. Computer software. 2022 (cit. on p. 15). [Mar+25] S. Marks, C. Rager, E. J. Michaud, Y. Belinkov, D. Bau, and A. Mueller. Sparse feature circu...

2022
[2019]

arXiv: 1811.12359 [cs.LG] (cit. on pp. 3, 11). [Loc+20] F . Locatello, B. Poole, G. Rätsch, B. Schölkopf, O. Bachem, and M. Tschannen. Weakly-supervised disentanglement without compromises

arXiv
[2021]

UNLOCKING HIERARCHICAL CONCEPT DISCOVERY IN LANGUAGE MODELS THROUGH GEOMETRIC REGULARIZATION

arXiv: 2111.00254 [cs.LG] (cit. on p. 15). [LR25] E. Li and J. Ren. “UNLOCKING HIERARCHICAL CONCEPT DISCOVERY IN LANGUAGE MODELS THROUGH GEOMETRIC REGULARIZATION”. In:ICLR 2025 Workshop on Building Trust in Language Models and Applications. 2025 (cit. on p. 3). [Lie+24] T . Li...

2025 arXiv
[2022]

Weakly Supervised Representation Learning with Sparse Perturbations

arXiv: 2206.01101 [cs.LG] (cit. on p. 3). [AHB22b] K. Ahuja, J. S. Hartford, and Y. Bengio. “Weakly Supervised Representation Learning with Sparse Perturbations”. en. Advances in Neural Information Pro- cessing Systems (2022) (cit. on p. 11). [Bra+18] J. Bradbury, R. Frostig, ...

2022 arXiv
[2023]

arXiv: 2309.08600 [cs.LG] (cit. on pp. 1, 3). [Den+25] B. Deng, Y. Wan, Y. Zhang, B. Yang, and F . Feng.Unveiling Language-Specific Features in Large Language Models via Sparse Autoencoders. arXiv:2505.05111 [cs]. 2025 (cit. on p. 8). [DCN24] J. Dunefsky, P . Chlenski, and N. ...

2025 arXiv
[2024]

arXiv: 2412.06410 [cs.LG] (cit. on p. 3). [Bus+25] B. Bussmann, N. Nabeshima, A. Karvonen, and N. Nanda. Learning multi-level features with matryoshka sparse autoencoders

arXiv
[2025]

arXiv: 2503.17547 [cs.LG] (cit. on p. 3). [Cha+24] D. Chanin, J. Wilken-Smith, T . Dulka, H. Bhatnagar, and J. Bloom. A is for absorption: studying feature splitting and absorption in sparse autoencoders

arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [2]

Weakly supervised causal representation learning

arXiv: 2203.16437 [stat.ML] (cit. on p. 3). [Bri+23] T . Bricken, A. Templeton, J. Batson, B. Chen, A. Jermyn, T . Conerly , N. Turner, C. Anil, C. Denison, A. Askell, R. Lasenby, Y. Wu, S. Kravec, N. Schiefer, T . Maxwell, N. Joseph, Z. Hatfield-Dodds, A. Tamkin, K. Nguyen, B. McLean, J. E. Burke, T . Hume, S. Carter, T . Henighan, and C. Olah. “Towards ...

work page Pith review arXiv 2023

[2] [5]

arXiv: 2409.14507 [cs.LG] (cit. on p. 1). [Cun+23] H. Cunningham, A. Ewart, L. Riggs, R. Huben, and L. Sharkey. Sparse au- toencoders find highly interpretable features in language models

arXiv

[3] [7]

arXiv: 2406.11944 [cs.LG] (cit. on p. 3). [Eng+25] J. Engels, E. J. Michaud, I. Liao, W . Gurnee, and M. Tegmark.Not all language model features are one-dimensionally linear

arXiv

[4] [8]

Scaling and evaluating sparse autoencoders

arXiv: 2405.14860 [cs.LG] (cit. on p. 3). [Gao+24] L. Gao, T . D. la Tour, H. Tillman, G. Goh, R. Troll, A. Radford, I. Sutskever, J. Leike, and J. Wu. “Scaling and evaluating sparse autoencoders”. arXiv preprint arXiv:2406.04093 (2024) (cit. on pp. 1–3, 7, 8). [Hea+25] T . Heap, T . Lawson, L. Farnik, and L. Aitchison.Sparse Autoencoders Can Inter- pret ...

arXiv 2024

[5] [9]

arXiv: 2503.01822 [cs.LG] (cit. on p. 3). [Joh24] D. D. Johnson. Penzai + treescope: a toolkit for interpreting, visualizing, and editing models as data. 2024 (cit. on p. 15). [Jos+25] S. Joshi, A. Dittadi, S. Lachapelle, and D. Sridhar.Identifiable steering via sparse autoencoding of multi-concept shifts

arXiv 2024

[6] [10]

arXiv: 2502.12179 [cs.LG] (cit. on p. 3). [Kar+25] A. Karvonen, C. Rager, J. Lin, C. Tigges, J. Bloom, D. Chanin, Y. -T . Lau, E. Farrell, C. McDougall, K. Ayonrinde, M. Wearden, A. Conmy, S. Marks, and N. Nanda. Saebench: a comprehensive benchmark for sparse autoencoders in language model interpretability

arXiv

[7] [11]

arXiv: 2503.09532 [cs.LG] (cit. on p. 8). [KG21] P . Kidger and C. Garcia.Equinox: neural networks in jax via callable pytrees and filtered transformations

arXiv

[8] [13]

arXiv: 2408.05147 [cs.LG] (cit. on pp. 1, 3, 7, 8). [Lin+25] J. Lindsey, W . Gurnee, E. Ameisen, B. Chen, A. Pearce, N. L. Turner, C. Citro, D. Abrahams, S. Carter, B. Hosmer, J. Marcus, M. Sklar, A. Templeton, T . Bricken, C. McDougall, H. Cunningham, T . Henighan, A. Jermyn, A. Jones, and J. Batson. On the biology of a large language model. 2025 (cit. o...

arXiv 2025