Pith. sign in

REVIEW 45 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
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). Sign in to comment.

Forward citations

Cited by 45 Pith papers

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

  1. Expander Sparse Autoencoders: Parameter-Efficient Dictionaries for Mechanistic Interpretability

    cs.LG 2026-07 conditional novelty 8.0 of 10

    Expander SAEs apply left-d-regular expander masks to TopK SAEs, learning only dn decoder parameters instead of mn and tracing a storage-fidelity frontier that reaches 293x compression with 84% retained performance on ...

  2. Descriptive Collision in Sparse Autoencoder Auto-Interpretability: When One Explanation Describes Many Features

    cs.LG 2026-05 accept novelty 8.0 of 10

    Many distinct SAE features share identical explanations, with the average annotation resolving only 70% of feature identity in a large annotated dataset.

  3. WriteSAE: Sparse Autoencoders for Recurrent State

    cs.LG 2026-05 unverdicted novelty 8.0 of 10

    WriteSAE introduces sparse autoencoders with rank-1 matrix atoms for recurrent state updates, allowing replacement tests that outperform deletion on 92.4% of positions and a formula predicting logit changes with R²=0.98.

  4. WriteSAE: Sparse Autoencoders for Recurrent State

    cs.LG 2026-05 unverdicted novelty 8.0 of 10

    WriteSAE decomposes recurrent model cache writes into substitutable atoms with a closed-form logit shift, achieving high substitution success and targeted behavioral installs on models like Qwen3.5 and Mamba-2.

  5. WriteSAE: Sparse Autoencoders for Recurrent State

    cs.LG 2026-05 unverdicted novelty 8.0 of 10

    WriteSAE is the first sparse autoencoder that factors decoder atoms into the native d_k x d_v cache write shape of recurrent models and supplies a closed-form per-token logit shift for atom substitution.

  6. Turn-Averaged SAEs for Feature Discovery and Long-Context Attribution

    cs.CL 2026-06 unverdicted novelty 7.0 of 10

    Turn-averaged SAEs reconstruct average activations over conversation turns to represent high-level turn characteristics with a fixed number of features, simplifying long-context interpretability compared to per-token SAEs.

  7. 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.

  8. Rational Sparse Autoencoder

    cs.LG 2026-06 unverdicted novelty 7.0 of 10

    RSAE replaces fixed SAE encoder activations (ReLU, JumpReLU, TopK) with trainable rational functions, initialized from baselines and fine-tuned to improve reconstruction and downstream metrics on language-model residu...

  9. VFUSE: Virulent Feature Understanding with Sparse autoEncoders

    cs.LG 2026-06 unverdicted novelty 7.0 of 10

    VFUSE applies sparse autoencoders to diffusion-transformer activations in RoseTTAFold3 and RFDiffusion3 to find monosemantic features that detect hazardous protein designs with AUROC up to 0.84.

  10. Subspace-Aware Sparse Autoencoders for Effective Mechanistic Interpretability

    cs.LG 2026-06 conditional novelty 7.0 of 10

    SASA replaces single-vector decoders in SAEs with learned subspaces plus block sparsity and nuclear-norm regularization, proving that a single group becomes the global minimizer once block size meets intrinsic dimensi...

  11. 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.

  12. WriteSAE: Sparse Autoencoders for Recurrent State

    cs.LG 2026-05 unverdicted novelty 7.0 of 10

    WriteSAE factors sparse autoencoder decoder atoms to the native d_k x d_v cache write shape in recurrent models, provides a closed-form logit shift, and demonstrates high success in atom substitution and behavioral ed...

  13. Tensor Product Representation Probes Reveal Shared Structure Across Linear Directions

    cs.LG 2026-05 unverdicted novelty 7.0 of 10

    Linear probes for Othello board states factor into tensor-product structure with square and color embeddings composed by a binding matrix, from which the linear probes can be directly recovered.

  14. Monosemanticity in Recommender Systems

    cs.IR 2026-06 unverdicted novelty 6.0 of 10

    Matryoshka Sparse Autoencoders applied to matrix-factorization embeddings from the Amazon Fashion dataset recover hierarchical monosemantic features that align with metadata and permit targeted intervention.

  15. Monosemanticity in Recommender Systems

    cs.IR 2026-06 unverdicted novelty 6.0 of 10

    Matryoshka Sparse Autoencoders applied to matrix-factorization embeddings recover hierarchical, metadata-aligned features that permit targeted intervention on gender-associated neurons.

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

    cs.LG 2026-06 conditional novelty 6.0 of 10

    Pre-selection off-support ℓ1 and ℓ1/ℓ2 regularizers improve Top-k SAE monosemanticity and concentration on vision foundation models at no reconstruction cost.

  17. Do Sparse Autoencoders Learn Meaningful Concept Hierarchies?

    cs.LG 2026-06 unverdicted novelty 6.0 of 10

    Sparse autoencoders provide a basis for sensible concept hierarchies on visual data but are undermined by hard and soft feature absorption.

  18. Critical Percolation as a Synthetic Data Model for Interpretability

    cs.LG 2026-06 unverdicted novelty 6.0 of 10

    Critical percolation clusters embedded in high dimensions, combined with taxonomic latent variables, form an analytically tractable synthetic data model whose ground-truth hierarchy can be linearly decoded from networ...

  19. ICA Lens: Interpreting Language Models Without Training Another Dictionary

    cs.LG 2026-06 unverdicted novelty 6.0 of 10

    ICALens applies an optimized ICA workflow to LLM activations and recovers compact interpretable directions that match or exceed public SAEs on SAEBench probing and perturbation tasks without per-layer dictionary training.

  20. MM-Matryoshka: Towards Budget-Elastic Visual Document Retrieval via a 2D Multimodal Matryoshka Training Framework

    cs.CV 2026-06 unverdicted novelty 6.0 of 10

    MM-Matryoshka is a 2D Matryoshka training framework enabling budget-elastic ColPali-style multi-vector visual document retrieval along dimension and layer without separate models per budget.

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

    cs.LG 2026-05 conditional novelty 6.0 of 10

    A half-width sign-aware gated sparse autoencoder matches full-width Gated SAE reconstruction on six LLM hookpoints while cutting dead features by 0.35–0.82 absolute at matched sparsity.

  22. Matryoshka Concept Bottleneck Models

    cs.LG 2026-05 unverdicted novelty 6.0 of 10

    MCBM nests concepts by relevance and redundancy to support multi-level inference in a single model, claiming logarithmic reduction in intervention cost with monotonic performance gains.

  23. Matryoshka Concept Bottleneck Models

    cs.LG 2026-05 unverdicted novelty 6.0 of 10

    MCBM nests concepts hierarchically to enable multi-granularity inference in a single model, reducing expected intervention costs to O(log K) with monotonic performance gains.

  24. Are Sparse Autoencoder Benchmarks Reliable?

    cs.LG 2026-05 unverdicted novelty 6.0 of 10

    An audit of SAEBench reveals that Targeted Probe Perturbation and Spurious Correlation Removal metrics fail reliability tests and should not be used to evaluate sparse autoencoders.

  25. The Rate-Distortion-Polysemanticity Tradeoff in SAEs

    cs.LG 2026-05 unverdicted novelty 6.0 of 10

    SAEs exhibit a rate-distortion-polysemanticity tradeoff where monosemanticity increases rate and distortion, with optimal polysemanticity set by feature co-occurrence probabilities in the data.

  26. Do Language Models Encode Knowledge of Linguistic Constraint Violations?

    cs.CL 2026-05 unverdicted novelty 6.0 of 10

    Sparse autoencoder analysis of language model activations finds limited evidence for a unified set of features detecting linguistic constraint violations.

  27. Do Language Models Encode Knowledge of Linguistic Constraint Violations?

    cs.CL 2026-05 unverdicted novelty 6.0 of 10

    Sparse autoencoder features in language models do not satisfy joint falsification criteria for unified grammatical violation detectors across linguistic phenomena.

  28. Tree SAE: Learning Hierarchical Feature Structures in Sparse Autoencoders

    cs.LG 2026-05 unverdicted novelty 6.0 of 10

    Tree SAE learns hierarchical feature pairs in sparse autoencoders by combining activation coverage with a new reconstruction condition, outperforming prior methods on hierarchy detection while remaining competitive on...

  29. Tree SAE: Learning Hierarchical Feature Structures in Sparse Autoencoders

    cs.LG 2026-05 unverdicted novelty 6.0 of 10

    Tree SAE learns hierarchical feature structures by combining activation coverage with a new reconstruction condition, outperforming prior SAEs on hierarchical pair detection while matching state-of-the-art benchmark p...

  30. From Tokens to Concepts: Leveraging SAE for SPLADE

    cs.IR 2026-04 conditional novelty 6.0 of 10

    Replacing SPLADE's MLM vocabulary with SAE-learned semantic concepts achieves comparable retrieval performance with improved efficiency.

  31. From Tokens to Concepts: Leveraging SAE for SPLADE

    cs.IR 2026-04 unverdicted novelty 6.0 of 10

    SAE-SPLADE substitutes SPLADE's backbone vocabulary with SAE-derived semantic concepts and matches standard SPLADE performance with better efficiency on in- and out-of-domain tasks.

  32. Dictionary-Aligned Concept Control for Safeguarding Multimodal LLMs

    cs.LG 2026-04 unverdicted novelty 6.0 of 10

    DACO curates a 15,000-concept dictionary from 400K image-caption pairs and uses it to initialize an SAE that enables granular, concept-specific steering of MLLM activations, raising safety scores on MM-SafetyBench and...

  33. Improving Robustness In Sparse Autoencoders via Masked Regularization

    cs.LG 2026-04 unverdicted novelty 6.0 of 10

    Masked regularization in sparse autoencoders disrupts token co-occurrences to reduce feature absorption, enhance probing, and narrow OOD gaps across architectures and sparsity levels.

  34. 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.

  35. Mechanistic Interpretability of Antibody Language Models Using SAEs

    cs.LG 2025-12 conditional novelty 6.0 of 10

    In p-IgGen, TopK SAE features correlate with IGHJ4 but do not steer it, while two Ordered SAE latents predictably raise and lower IGHJ4 output.

  36. 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.

  37. 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.

  38. Anatomy of Post-Training: Using Interpretability to Characterize Data and Shape the Learning Signal

    cs.LG 2026-06 unverdicted novelty 5.0 of 10

    A new pipeline uses interpretability to characterize concepts in preference data and shape rewards via feature or data interventions during LM post-training.

  39. Probing for Representation Manifolds in Superposition

    cs.LG 2026-05 unverdicted novelty 5.0 of 10

    Introduces the Manifold Probe to discover representation manifolds in superposition and demonstrates causal steering on time concepts in Llama 2-7b.

  40. 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.

  41. Mechanistic Interpretability of Antibody Language Models Using SAEs

    cs.LG 2025-12 unverdicted novelty 5.0 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.

  42. Safe-SAIL: Towards a Fine-grained Safety Landscape of Large Language Models via Sparse Autoencoder Interpretation Framework

    cs.LG 2025-09 unverdicted novelty 5.0 of 10

    Safe-SAIL supplies a pre-explanation metric and segment-level simulation to interpret 1758 safety SAE features across pornography, politics, violence, and terror, with public models and tools released.

  43. 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.

  44. 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...

  45. 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