pith. sign in

arxiv: 2605.28149 · v1 · pith:QAI2FWGDnew · submitted 2026-05-27 · 💻 cs.LG

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

Bartosz Wieciech , Zmnako Awrahman , Marcin Czelej , Victor Hugo Jaramillo Velasquez , Wioletta Stobieniecka This is my paper

Pith reviewed 2026-06-29 14:13 UTC · model grok-4.3

classification 💻 cs.LG
keywords sparse autoencodersgated SAEsanticorrelated featuresbipolar latentsLLM interpretabilityBi-Jump-ReLUsign-aware gating
0
0 comments X

The pith

Half-width sign-aware gated sparse autoencoders match or exceed full-width gated SAEs on reconstruction while cutting dead feature rates by up to 500x.

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

The paper proposes Sign-Aware Gated SAEs to let dictionary elements represent both positive and negative versions of the same underlying concept instead of forcing separate non-negative latents for anticorrelated pairs. It introduces a polarity-sensitive gate, a signed-magnitude encoder path, and an auxiliary reconstruction loss that together enable bipolar sharing through a Bi-Jump-ReLU activation. Across six hookpoint-model cells from Pythia-1B and SmolLM3-3B, a model at width H strictly Pareto-dominates a standard gated SAE at width 2H on three cells and stays within 0.025 R^2 on the rest while lowering dead fraction by 0.35-0.62 absolute. The symmetric variant that ties the two radii performs indistinguishably from the untied version and is recommended as default.

Core claim

The central claim is that sign-awareness combined with auxiliary supervision realizes bipolar sharing—one latent encoding both signs along a shared direction—while remaining parameter-efficient. On real LLM activations the half-width SA-GSAE at H strictly Pareto-dominates the full-width Gated SAE at 2H over the entire L0 sweep on three of six cells and matches R^2 within 0.025 (max gap -0.008) on the remaining three while cutting dead fraction by 0.35-0.62 absolute; sweep-geomean dead-fraction reductions are 100x-500x on MLP-output and Pythia-1B resid cells. Ablations confirm the two-sided gate and auxiliary loss are load-bearing, tying the radii is sufficient, and full-width SA-GSAE exhibit

What carries the argument

The Bi-Jump-ReLU activation, which uses a polarity-sensitive gate to select support on either sign and a signed-magnitude path to avoid L1 shrinkage while an auxiliary reconstruction term prevents gate collapse.

If this is right

  • Most latents in MLP-output hookpoints carry both polarities.
  • The auxiliary loss is required; removing it causes collapse.
  • Tying positive and negative radii yields |Delta R^2| of only 0.0015.
  • Bipolar structure concentrates in a small set of top latents on attention cells.
  • Full-width SA-GSAE shows reproducible reconstruction collapse on SmolLM3-3B resid that half-width avoids.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The efficiency pattern may appear in larger models if the same anticorrelation structure exists.
  • Fewer but bipolar latents could simplify downstream interpretability analyses.
  • Attention versus MLP differences suggest layer-specific feature polarity statistics worth mapping.
  • Testing whether the same auxiliary loss suffices when dictionary size grows by another factor of four would clarify scaling.

Load-bearing premise

The auxiliary reconstruction loss is sufficient to prevent gate collapse in general settings and the observed Pareto dominance will generalize beyond the six specific hookpoint-model combinations tested.

What would settle it

Running the same sweep on a seventh hookpoint or new model and finding either gate collapse to learning rate 0.27 with 98 percent dead latents despite the auxiliary loss, or loss of the half-width Pareto dominance on reconstruction metrics.

Figures

Figures reproduced from arXiv: 2605.28149 by Bartosz Wieciech, Marcin Czelej, Victor Hugo Jaramillo Velasquez, Wioletta Stobieniecka, Zmnako Awrahman.

Figure 1
Figure 1. Figure 1: Bi-Jump-ReLU is zero inside a learnable dead zone [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Dead-feature fraction vs log(L0) across the three mid-depth hookpoints (columns: mlp_out, attn, resid). Top row: Pythia-1B (layer 8/16). Bottom row: SmolLM3-3B (layer 18/36). Error bars show ±SE over 3 seeds on both axes [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Loss Recovered (LR) vs log(L0) across the three mid-depth hookpoints (columns: mlp_out, attn, resid). Top row: Pythia-1B (layer 8/16). Bottom row: SmolLM3-3B (layer 18/36). Error bars show ±SE over 3 seeds on both axes. LR saturates near 1 at residual-stream hookpoints and has a noise-dominated ceiling on attn/SmolLM3-3B (baseline gap 0.037 nats); see benchmark setup for the LR-not-cross-hookpoint-comparab… view at source ↗
Figure 4
Figure 4. Figure 4: Protocol A (Polarity Dial) across the full sparsity sweep. Left: reconstruction MSE vs. [PITH_FULL_IMAGE:figures/full_fig_p020_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Protocol B (feature-direction anomalies) at [PITH_FULL_IMAGE:figures/full_fig_p022_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Protocol C qualitative geometry. Blue points mark the ground-truth feature directions on [PITH_FULL_IMAGE:figures/full_fig_p023_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Bi-jump histogram for a representative Sign-Aware latent in Protocol C. The distribution [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Protocol C quantitative curve (superposition tolerance, [PITH_FULL_IMAGE:figures/full_fig_p024_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Thresholded Pair Consolidation Rate vs superposition in Protocol C (mean [PITH_FULL_IMAGE:figures/full_fig_p024_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Robustness sweep: Pair Consolidation Rate vs within-pair correlation [PITH_FULL_IMAGE:figures/full_fig_p025_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Empirical CDFs of per-latent calibration slopes [PITH_FULL_IMAGE:figures/full_fig_p027_11.png] view at source ↗
read the original abstract

Sparse Autoencoders (SAEs) extract interpretable features from Large Language Models, but standard variants enforce non-negativity, forcing separate latents for diametrically opposed concepts (e.g., "pressure too high" vs. "pressure too low") and wasting dictionary capacity when features are anticorrelated. We propose the Sign-Aware Gated SAE (SA-GSAE): two-sided gated sparsity with signed magnitude and auxiliary supervision. A polarity-sensitive gate selects support on either sign, a signed-magnitude path avoids L1 shrinkage, and an auxiliary reconstruction prevents gate collapse. Bipolar sharing - one latent encoding both signs along a shared direction - is realised via a new Bi-Jump-ReLU activation; parameter accounting shows sign-awareness stays parameter-efficient even when anticorrelated pairs are rare. On real LLM activations across three mid-depth hookpoints on Pythia-1B and SmolLM3-3B (6 cells, 3 seeds), a half-width SA-GSAE at width H strictly Pareto-dominates a full-width Gated SAE at 2H over the entire swept L0 overlap on 3 of 6 cells (both MLP-output hookpoints and resid-mid/Pythia-1B); on the remaining 3 it matches R^2 within 0.025 (max gap -0.008) while cutting dead fraction by 0.35-0.62 absolute. Sweep-geomean dead-fraction reductions are ~100x-500x on MLP-output cells and Pythia-1B resid, ~2x-4x on attention cells and SmolLM3-3B resid. Ablations show the two-sided gate and auxiliary loss are load-bearing (no auxiliary collapses LR to 0.27, 98% dead); tying r_i^+ = r_i^- is indistinguishable (|Delta R^2| = 0.0015), and we recommend this symmetric variant as default. MLP-output gains come from most latents carrying both polarities; on attention, bipolar structure concentrates in a small set of top latents. Full-width SA-GSAE exhibits a reproducible reconstruction collapse at SmolLM3-3B resid that the half-width entirely avoids.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 3 minor

Summary. The manuscript introduces Sign-Aware Gated Sparse Autoencoders (SA-GSAE) that employ a Bi-Jump-ReLU activation to model anticorrelated features via two-sided gated sparsity, signed magnitudes, and an auxiliary reconstruction loss to avoid gate collapse. It claims that a half-width SA-GSAE at dictionary size H strictly Pareto-dominates a full-width Gated SAE at 2H over the full L0 overlap on 3 of 6 tested cells (MLP-output hookpoints and Pythia-1B resid-mid) while matching R² within 0.025 and cutting dead fraction by 0.35-0.62 absolute on the remaining 3 cells; sweep-geomean dead-fraction reductions reach 100x-500x on MLP-output and Pythia-1B resid cells. Ablations confirm the auxiliary loss and two-sided gate are required, with a reproducible collapse noted for full-width SA-GSAE on SmolLM3-3B resid-mid.

Significance. If the reported Pareto dominance holds, the method improves SAE parameter efficiency for bipolar features without doubling width. The manuscript supplies concrete, multi-seed numbers (3 seeds, 6 cells) on R² and dead fractions, explicit ablation results (LR drops to 0.27 and 98% dead without auxiliary loss), and acknowledgment of a collapse case, enabling direct falsification. These elements strengthen the empirical contribution relative to prior gated SAE baselines.

major comments (2)
  1. [Ablations] Ablations section: while the auxiliary reconstruction loss is shown to be load-bearing (no-auxiliary LR=0.27, 98% dead), the manuscript also reports a reproducible full-width SA-GSAE reconstruction collapse on the SmolLM3-3B resid-mid cell. This indicates the auxiliary term's robustness may be setting-dependent, which is load-bearing for the claim that SA-GSAE reliably achieves the reported dominance without collapse.
  2. [Results] Results section (6-cell comparison): the strict Pareto dominance on 3/6 cells and the 'entire swept L0 overlap' claim rest on the intersection of evaluated L0 ranges; the manuscript should explicitly state the pre-specified criteria used to define this overlap and confirm that L0 targets were not selected after inspecting per-cell outcomes.
minor comments (3)
  1. [Methods] Methods: the Bi-Jump-ReLU activation and the precise parameter count for the sign-aware mechanism should be stated with an explicit equation or table rather than prose description alone.
  2. [Figures] Figures: the Pareto-front plots would benefit from explicit annotation of the L0 overlap interval used for each cell to allow readers to verify the dominance claim without re-deriving the ranges.
  3. [Discussion] The manuscript should add a short limitations paragraph addressing generalization beyond the six hookpoint-model combinations tested.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive assessment and recommendation for minor revision. The comments highlight important points on robustness and experimental transparency, which we address below with clarifications and planned revisions.

read point-by-point responses

  1. Referee: [Ablations] Ablations section: while the auxiliary reconstruction loss is shown to be load-bearing (no-auxiliary LR=0.27, 98% dead), the manuscript also reports a reproducible full-width SA-GSAE reconstruction collapse on the SmolLM3-3B resid-mid cell. This indicates the auxiliary term's robustness may be setting-dependent, which is load-bearing for the claim that SA-GSAE reliably achieves the reported dominance without collapse.

    Authors: We agree that the observed collapse of the full-width variant on the SmolLM3-3B resid-mid cell indicates that auxiliary-loss robustness can be width- and setting-dependent. The manuscript already reports this collapse explicitly and notes that the half-width SA-GSAE avoids it entirely on that cell. Our primary claims and Pareto-dominance results concern the half-width configuration (the recommended default), where the auxiliary loss remains load-bearing per the reported ablation. We will expand the discussion section to explicitly address this setting dependence, clarify that the half-width variant is the focus of the efficiency claims, and note the full-width collapse as a known limitation of the wider configuration. revision: yes

  2. Referee: [Results] Results section (6-cell comparison): the strict Pareto dominance on 3/6 cells and the 'entire swept L0 overlap' claim rest on the intersection of evaluated L0 ranges; the manuscript should explicitly state the pre-specified criteria used to define this overlap and confirm that L0 targets were not selected after inspecting per-cell outcomes.

    Authors: The L0 overlap is the intersection of the discrete L0 targets at which both SA-GSAE and Gated SAE models were trained and evaluated across the fixed sweep ranges (L0 targets chosen in advance based on standard SAE literature ranges of approximately 5–200). These targets were not adjusted post-hoc after inspecting per-cell results. We will add an explicit statement in the Results section defining the overlap criterion as this pre-specified intersection and confirming the a-priori selection of L0 targets. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical results stand on direct benchmarks

full rationale

The paper proposes the SA-GSAE architecture (two-sided gated sparsity, Bi-Jump-ReLU, auxiliary reconstruction) and reports its performance via direct experimental comparison against Gated SAE baselines on six specific hookpoint-model cells. No equations or derivations are presented that reduce the reported R^2, dead-fraction, or Pareto-dominance outcomes to quantities defined by the paper's own fitted parameters or self-citations. Ablations confirm the auxiliary loss is required, but this is an independent empirical check rather than a self-referential loop. The central claims remain falsifiable against external models and layers outside the tested set.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 2 invented entities

The central performance claims rest on the new architectural components and the empirical observation that anticorrelated features exist in the tested activations; no independent evidence is supplied for the new activation function outside the reported runs.

free parameters (2)
  • dictionary width H
    Half-width and full-width choices used for Pareto comparison; values not numerically specified in abstract.
  • L0 sparsity targets
    Swept values used to generate overlap curves; exact schedule not given.
axioms (2)
  • domain assumption LLM activations contain anticorrelated feature pairs that can be usefully represented by a single bipolar latent
    Invoked to motivate the Bi-Jump-ReLU design and the claim that sign-awareness stays parameter-efficient.
  • ad hoc to paper The auxiliary reconstruction loss prevents gate collapse without introducing other distortions
    Stated as load-bearing in the ablation section of the abstract.
invented entities (2)
  • Bi-Jump-ReLU activation no independent evidence
    purpose: Realize bipolar sharing so one latent encodes both signs along a shared direction
    New activation function introduced to avoid separate latents for opposite concepts.
  • Sign-aware gated sparsity mechanism no independent evidence
    purpose: Select support on either sign while avoiding L1 shrinkage on magnitude
    Core of the SA-GSAE architecture.

pith-pipeline@v0.9.1-grok · 5975 in / 1692 out tokens · 47819 ms · 2026-06-29T14:13:40.543047+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 31 canonical work pages · 10 internal anchors

  1. [1]

    Mechanistic interpretability, variables, and the importance of interpretable bases

    Chris Olah. Mechanistic interpretability, variables, and the importance of interpretable bases. https: //www.transformer-circuits.pub/2022/mech-interp-essay, 2022

    2022

  2. [2]

    Zoom in: An introduction to circuits.Distill, 2020

    Chris Olah, Nick Cammarata, Ludwig Schubert, Gabriel Goh, Michael Petrov, and Shan Carter. Zoom in: An introduction to circuits.Distill, 2020. URLhttps://distill.pub/2020/circuits/zoom-in/

    2020

  3. [3]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. InAdvances in Neural Information Processing Systems, 2017. URL https://papers.neurips.cc/paper/7181-attention-is-all-you-need. pdf

    2017

  4. [4]

    An interpretability illusion for BERT.arXiv preprint arXiv:2104.07143, 2021

    Tolga Bolukbasi, Adam Pearce, Ann Yuan, Andy Coenen, Emily Reif, Fernanda Viégas, and Martin Wattenberg. An interpretability illusion for BERT.arXiv preprint arXiv:2104.07143, 2021

    work page arXiv 2021

  5. [5]

    Toy Models of Superposition

    Nelson Elhage et al. Toy models of superposition.arXiv preprint arXiv:2209.10652, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  6. [6]

    The Linear Representation Hypothesis and the Geometry of Large Language Models

    Kiho Park, Yo Joong Choe, and Victor Veitch. The linear representation hypothesis and the geometry of large language models.arXiv preprint arXiv:2311.03658, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  7. [7]

    Michaud, Isaac Liao, Wes Gurnee, and Max Tegmark

    Joshua Engels, Eric J. Michaud, Isaac Liao, Wes Gurnee, and Max Tegmark. Not all language model features are one-dimensionally linear.arXiv preprint arXiv:2405.14860, 2024

    work page arXiv 2024

  8. [8]

    Olshausen and David J

    Bruno A. Olshausen and David J. Field. Emergence of simple-cell receptive field properties by learning a sparse code for natural images.Nature, 381(6583):607–609, 1996

    1996

  9. [9]

    Regression shrinkage and selection via the lasso.Journal of the Royal Statistical Society: Series B (Methodological), 58(1):267–288, 1996

    Robert Tibshirani. Regression shrinkage and selection via the lasso.Journal of the Royal Statistical Society: Series B (Methodological), 58(1):267–288, 1996

    1996

  10. [10]

    K-SVD: An algorithm for designing overcomplete dictionaries for sparse representation.IEEE Transactions on Signal Processing, 54(11):4311–4322, 2006

    Michal Aharon, Michael Elad, and Alfred Bruckstein. K-SVD: An algorithm for designing overcomplete dictionaries for sparse representation.IEEE Transactions on Signal Processing, 54(11):4311–4322, 2006

    2006

  11. [11]

    Online dictionary learning for sparse coding

    Julien Mairal, Francis Bach, Jean Ponce, and Guillermo Sapiro. Online dictionary learning for sparse coding. InProceedings of the 26th International Conference on Machine Learning (ICML), pages 689–696, 2009

    2009

  12. [12]

    Lee and H

    Daniel D. Lee and H. Sebastian Seung. Learning the parts of objects by non-negative matrix factorization. Nature, 401(6755):788–791, 1999

    1999

  13. [13]

    Learning fast approximations of sparse coding

    Karol Gregor and Yann LeCun. Learning fast approximations of sparse coding. InProceedings of the 27th International Conference on Machine Learning (ICML), pages 399–406, 2010

    2010

  14. [14]

    Towards monosemanticity: Decomposing language models with dictionary learning

    Trenton Bricken et al. Towards monosemanticity: Decomposing language models with dictionary learning. https://transformer-circuits.pub/2023/monosemantic-features/index.html, 2023

    2023

  15. [15]

    Sparse Autoencoders Find Highly Interpretable Features in Language Models

    Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. Sparse autoencoders find highly interpretable features in language models.arXiv preprint arXiv:2309.08600, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  16. [16]

    Scaling and evaluating sparse autoencoders

    Leo Gao, Tom Dupré la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Ilya Sutskever, Jan Leike, and Jeffrey Wu. Scaling and evaluating sparse autoencoders.arXiv preprint arXiv:2406.04093,

    work page internal anchor Pith review Pith/arXiv arXiv

  17. [17]

    URLhttps://cdn.openai.com/papers/sparse-autoencoders.pdf

  18. [18]

    Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2

    Tom Lieberum et al. Gemma scope: Open sparse autoencoders everywhere all at once on gemma 2.arXiv preprint arXiv:2408.05147, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  19. [19]

    Llama Scope: Extracting millions of features from Llama-3.1-8b with sparse autoencoders.arXiv preprint arXiv:2410.20526,

    Zhengfu He et al. Llama scope: Extracting millions of features from llama-3.1-8b with sparse autoencoders. arXiv preprint arXiv:2410.20526, 2024

    work page arXiv 2024

  20. [20]

    Compute optimal inference and provable amortisation gap in sparse autoencoders.arXiv preprint arXiv:2411.13117, 2024

    Charles O’Neill, Alim Gumran, and David Klindt. Compute optimal inference and provable amortisation gap in sparse autoencoders.arXiv preprint arXiv:2411.13117, 2024

    work page arXiv 2024

  21. [21]

    On the limits of sparse autoencoders: A theoretical framework and reweighted remedy.arXiv preprint arXiv:2506.15963, 2025

    Jingyi Cui, Qi Zhang, Yifei Wang, and Yisen Wang. On the limits of sparse autoencoders: A theoretical framework and reweighted remedy.arXiv preprint arXiv:2506.15963, 2025

    work page arXiv 2025

  22. [22]

    Sparse autoencoders do not find canonical units of analysis.arXiv preprint arXiv:2502.04878, 2025

    Patrick Leask, Bart Bussmann, Michael Pearce, Joseph Bloom, Curt Tigges, Noura Al Moubayed, Lee Sharkey, and Neel Nanda. Sparse autoencoders do not find canonical units of analysis.arXiv preprint arXiv:2502.04878, 2025

    work page arXiv 2025

  23. [23]

    Sparse autoencoders trained on the same data learn different features

    Gonçalo Paulo and Nora Belrose. Sparse autoencoders trained on the same data learn different features. arXiv preprint arXiv:2501.16615, 2025

    work page arXiv 2025

  24. [24]

    AbsTopK: Rethinking sparse autoencoders for bidirectional features.arXiv preprint arXiv:2510.00404, 2025

    Xudong Zhu, Mohammad Mahdi Khalili, and Zhihui Zhu. AbsTopK: Rethinking sparse autoencoders for bidirectional features.arXiv preprint arXiv:2510.00404, 2025

    work page arXiv 2025

  25. [25]

    Improving Dictionary Learning with Gated Sparse Autoencoders

    Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Tom Lieberum, Vikrant Varma, János Krámár, Rohin Shah, and Neel Nanda. Improving dictionary learning with gated sparse autoencoders.arXiv preprint arXiv:2404.16014, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  26. [26]

    Richard H. R. Hahnloser, H. Sebastian Seung, and Jean-Jacques Slotine. Permitted and forbidden sets in symmetric threshold-linear networks.Neural Computation, 15(3):621–638, 2003. doi: 10.1162/ 089976603321192103

    2003

  27. [27]

    Symmetric-threshold ReLU for fast and nearly lossless ANN-SNN conversion.Machine Intelligence Research, 20(3):435–446, 2023

    Jianing Han, Ziming Wang, Jiangrong Shen, and Huajin Tang. Symmetric-threshold ReLU for fast and nearly lossless ANN-SNN conversion.Machine Intelligence Research, 20(3):435–446, 2023. doi: 10 10.1007/s11633-022-1388-2

    work page doi:10.1007/s11633-022-1388-2 2023

  28. [28]

    Pythia: A suite for analyzing large language models across training and scaling

    Stella Biderman, Hailey Schoelkopf, Quentin Anthony, Herbie Bradley, Kyle O’Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, Aviya Skowron, Lintang Sutawika, and Oskar van der Wal. Pythia: A suite for analyzing large language models across training and scaling. InProceedings of the 40th International Conferen...

    2023

  29. [29]

    The Pile: An 800GB Dataset of Diverse Text for Language Modeling

    Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, Shawn Presser, and Connor Leahy. The Pile: An 800gb dataset of diverse text for language modeling.arXiv preprint arXiv:2101.00027, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  30. [30]

    SmolLM3: smol, multilingual, long-context reasoner

    Elie Bakouch et al. SmolLM3: smol, multilingual, long-context reasoner. https://huggingface.co/ blog/smollm3, 2025

    2025

  31. [31]

    Openwebtext corpus

    Aaron Gokaslan, Vanya Cohen, Ellie Pavlick, and Stefanie Tellex. Openwebtext corpus. http:// Skylion007.github.io/OpenWebTextCorpus, 2019

    2019

  32. [32]

    Michaud, Max Tegmark, and Christian Schroeder de Witt

    Anish Mudide, Joshua Engels, Eric J. Michaud, Max Tegmark, and Christian Schroeder de Witt. Efficient dictionary learning with switch sparse autoencoders.arXiv preprint arXiv:2410.08201, 2024

    work page arXiv 2024

  33. [33]

    ”Learning multi-level features with matryoshka sparse autoencoders.” arXiv preprint arXiv:2503.17547 (2025)

    Bart Bussmann, Noa Nabeshima, Adam Karvonen, and Neel Nanda. Learning multi-level features with matryoshka sparse autoencoders.arXiv preprint arXiv:2503.17547, 2025

    work page arXiv 2025

  34. [34]

    A is for absorption: Studying feature splitting and absorption in sparse autoencoders

    David Chanin et al. A is for absorption: Studying feature splitting and absorption in sparse autoencoders. arXiv preprint arXiv:2409.14507, 2024

    work page arXiv 2024

  35. [35]

    Feature hedging: Correlated features break narrow sparse autoencoders.arXiv preprint arXiv:2505.11756, 2025

    David Chanin et al. Feature hedging: Correlated features break narrow sparse autoencoders.arXiv preprint arXiv:2505.11756, 2025

    work page arXiv 2025

  36. [36]

    An iterative thresholding algorithm for linear inverse problems with a sparsity constraint.Communications on Pure and Applied Mathematics, 57(11): 1413–1457, 2004

    Ingrid Daubechies, Michel Defrise, and Christine De Mol. An iterative thresholding algorithm for linear inverse problems with a sparsity constraint.Communications on Pure and Applied Mathematics, 57(11): 1413–1457, 2004. doi: 10.1002/cpa.20042

    work page doi:10.1002/cpa.20042 2004

  37. [37]

    A fast iterative shrinkage-thresholding algorithm for linear inverse problems.SIAM Journal on Imaging Sciences, 2(1):183–202, 2009

    Amir Beck and Marc Teboulle. A fast iterative shrinkage-thresholding algorithm for linear inverse problems.SIAM Journal on Imaging Sciences, 2(1):183–202, 2009. doi: 10.1137/080716542. URL https://www.tau.ac.il/~becka/FISTA.pdf

    work page doi:10.1137/080716542 2009

  38. [38]

    k-Sparse Autoencoders

    Alireza Makhzani and Brendan Frey. k-Sparse autoencoders.arXiv preprint arXiv:1312.5663, 2013

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  39. [39]

    Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders

    Senthooran Rajamanoharan, Tom Lieberum, Nicolas Sonnerat, Arthur Conmy, Vikrant Varma, János Krámár, and Neel Nanda. Jumping ahead: Improving reconstruction fidelity with JumpReLU sparse autoencoders.arXiv preprint arXiv:2407.14435, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  40. [40]

    BatchTopK sparse autoencoders.arXiv preprint arXiv:2412.06410, 2024

    Bart Bussmann, Patrick Leask, and Neel Nanda. BatchTopK sparse autoencoders.arXiv preprint arXiv:2412.06410, 2024

    work page arXiv 2024

  41. [41]

    SAEBench: A comprehensive benchmark for sparse autoencoders in language model interpretability.arXiv preprint arXiv:2503.09532, 2025

    Adam Karvonen et al. SAEBench: A comprehensive benchmark for sparse autoencoders in language model interpretability.arXiv preprint arXiv:2503.09532, 2025

    work page arXiv 2025

  42. [42]

    Evaluating SAE interpretability without explanations.arXiv preprint arXiv:2507.08473, 2025

    Gonçalo Paulo and Nora Belrose. Evaluating SAE interpretability without explanations.arXiv preprint arXiv:2507.08473, 2025

    work page arXiv 2025

  43. [43]

    Measuring progress in dictionary learning for language model interpretability with board game models.arXiv preprint arXiv:2408.00113, 2024

    Adam Karvonen, Benjamin Wright, Can Rager, Rico Angell, Jannik Brinkmann, Logan Smith, Clau- dio Mayrink Verdun, David Bau, and Samuel Marks. Measuring progress in dictionary learning for language model interpretability with board game models.arXiv preprint arXiv:2408.00113, 2024

    work page arXiv 2024

  44. [44]

    Towards principled evaluations of sparse autoen- coders for interpretability and control.arXiv preprint arXiv:2405.08366, 2024

    Aleksandar Makelov, George Lange, and Neel Nanda. Towards principled evaluations of sparse autoen- coders for interpretability and control.arXiv preprint arXiv:2405.08366, 2024

    work page arXiv 2024

  45. [45]

    Li, Suraj Srinivas, Usha Bhalla, and Himabindu Lakkaraju

    Aaron J. Li, Suraj Srinivas, Usha Bhalla, and Himabindu Lakkaraju. Evaluating adversarial robustness of concept representations in sparse autoencoders.arXiv preprint arXiv:2505.16004, 2025

    work page arXiv 2025

  46. [46]

    Adam: A Method for Stochastic Optimization

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014. URLhttps://arxiv.org/abs/1412.6980. 11 A Background and Related Work A.1 Sparse autoencoders as amortized sparse inference Solving eq. (1) exactly requires an inner optimization to inferz(n) for each sample, often via proximal methods (for exa...

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  47. [47]

    asymmetric signed-axis calibration,

  48. [48]

    directional anomaly detection under a strict latent budget,

  49. [49]

    Polarity Dial

    consolidation of anticorrelated pairs in toy geometry. Unless otherwise stated, we report not only a single operating point, but also sparsity-fidelity trade- offs by sweeping the sparsity strength (e.g., λ) and plotting reconstruction MSE versus average L0 (active latents per example). C.1 Protocol A: The "Polarity Dial" (Asymmetry & Calibration) Protoco...

    1910