Pith. sign in

REVIEW 15 cited by

Learning Multi-Level Features with Matryoshka Sparse Autoencoders

Not yet reviewed by Pith; the record is open.

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

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2503.17547 v1 pith:JD4C457K submitted 2025-03-21 cs.LG cs.AI

Bart Bussmann , Noa Nabeshima , Adam Karvonen , Neel Nanda This is my paper · ORCID
classification cs.LGcs.AI
keywords featuresconceptsdictionariessaesmatryoshkasizesparsewhile
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

Sparse autoencoders (SAEs) have emerged as a powerful tool for interpreting neural networks by extracting the concepts represented in their activations. However, choosing the size of the SAE dictionary (i.e. number of learned concepts) creates a tension: as dictionary size increases to capture more relevant concepts, sparsity incentivizes features to be split or absorbed into more specific features, leaving high-level features missing or warped. We introduce Matryoshka SAEs, a novel variant that addresses these issues by simultaneously training multiple nested dictionaries of increasing size, forcing the smaller dictionaries to independently reconstruct the inputs without using the larger dictionaries. This organizes features hierarchically - the smaller dictionaries learn general concepts, while the larger dictionaries learn more specific concepts, without incentive to absorb the high-level features. We train Matryoshka SAEs on Gemma-2-2B and TinyStories and find superior performance on sparse probing and targeted concept erasure tasks, more disentangled concept representations, and reduced feature absorption. While there is a minor tradeoff with reconstruction performance, we believe Matryoshka SAEs are a superior alternative for practical tasks, as they enable training arbitrarily large SAEs while retaining interpretable features at different levels of abstraction.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 15 Pith papers

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

  1. Beyond the Hard Budget: Sparsity Regularizers for More Interpretable Top-k Sparse Autoencoders

    cs.LG 2026-06 unverdicted novelty 7.0 of 10

    Sparsity regularizers applied before Top-k selection in SAEs improve monosemanticity and make reconstruction robust to inference-time k across vision models and datasets.

  2. Sign-Aware Gated Sparse Autoencoders: Modeling Anticorrelated Features with Bi-Jump-ReLU Activations

    cs.LG 2026-05 conditional novelty 7.0 of 10

    SA-GSAE with Bi-Jump-ReLU enables one latent to encode both polarities of anticorrelated features, Pareto-dominating or matching full-width gated SAEs while reducing dead latents by up to 500x on some LLM hookpoints.

  3. Incorporating Hierarchical Semantics in Sparse Autoencoder Architectures

    cs.CL 2025-06 conditional novelty 7.0 of 10

    A two-level mixture-of-experts sparse autoencoder models parent and child concepts together, improving reconstruction and reducing feature redundancy on Gemma 2-2B activations compared to flat top-k SAEs.

  4. PolySAE: Modeling Feature Interactions in Sparse Autoencoders via Polynomial Decoding

    cs.LG 2026-02 conditional novelty 6.0 of 10

    A sparse autoencoder decoder with quadratic and cubic feature interactions on a shared low-rank subspace improves average probing F1 by ~8%, yet the causal steering claim in the abstract is not demonstrated.

  5. Mechanistic Interpretability of Antibody Language Models Using SAEs

    cs.LG 2025-12 unverdicted novelty 6.0 of 10

    TopK SAEs uncover biologically meaningful latent features in antibody language models without guaranteeing causal steering, whereas Ordered SAEs provide reliable generative control at the cost of complex activation patterns.

  6. Sparse but Wrong: Incorrect L0 Leads to Incorrect Features in Sparse Autoencoders

    cs.LG 2025-08 conditional novelty 6.0 of 10

    Incorrect L0 makes sparse autoencoders mix correlated features rather than disentangling them, and a decoder projection metric can identify the correct L0.

  7. Insights into a radiology-specialised multimodal large language model with sparse autoencoders

    cs.LG 2025-07 conditional novelty 6.0 of 10

    Applying Matryoshka sparse autoencoders to a radiology-specialised multimodal LLM reveals a minority of interpretable clinical features, while steering them produces unreliable and often off-target report changes.

  8. Position: Mechanistic Interpretability Should Prioritize Feature Consistency in SAEs

    cs.LG 2025-05 conditional novelty 6.0 of 10

    A position paper proposing feature consistency, measured by PW-MCC, as a core SAE evaluation criterion, with evidence that TopK SAEs achieve high consistency on LLM activations.

  9. The Origins of Representation Manifolds in Large Language Models

    cs.LG 2025-05 conditional novelty 6.0 of 10

    A proof and empirical check that cosine similarity in language model representations encodes the intrinsic geometry of features through shortest paths on manifolds.

  10. Stable and Steerable Sparse Autoencoders with Weight Regularization

    stat.ML 2026-03 conditional novelty 5.0 of 10

    L2 weight regularization in TopK SAEs increases cross-seed feature overlap and roughly doubles measured steering success on Pythia-70M, at the cost of collapsing most latents to zero.

  11. Sealing The Backdoor: Unlearning Adversarial Text Triggers In Diffusion Models Using Knowledge Distillation

    cs.CV 2025-08 unverdicted novelty 5.0 of 10

    SKD-CAG erases adversarial text triggers from diffusion models by distilling the model's own clean outputs through cross-attention guidance, claiming 100% and 93% removal for pixel and style backdoors.

  12. BlueGlass: A Framework for Composite AI Safety

    cs.AI 2025-07 conditional novelty 5.0 of 10

    BlueGlass provides composite AI safety infrastructure; its case studies on object-detection VLMs reveal dataset trade-offs, a decoder-layer phase transition in probe accuracy, and SAE-discovered concepts including spu...

  13. Train One Sparse Autoencoder Across Multiple Sparsity Budgets to Preserve Interpretability and Accuracy

    cs.LG 2025-05 conditional novelty 5.0 of 10

    HierarchicalTopK trains a single sparse autoencoder that reconstructs transformer activations well at many sparsity levels, matching or beating separate per-level models.

  14. Factual Self-Awareness in Language Models: Representation, Robustness, and Scaling

    cs.CL 2025-05 conditional novelty 5.0 of 10

    Language models encode a linear, decodable signal in their residual stream that predicts whether an upcoming factual recall will be correct.

  15. Towards Atoms of Large Language Models

    cs.CL 2025-09 reject novelty 4.0 of 10

    The authors define 'atoms' as sparse, near-orthogonal directions in LLM representations under a data-adaptive inner product, and show threshold-activated sparse autoencoders can recover them with about 99.9% reconstru...

Pith tools