Pith. sign in

REVIEW 3 major objections 5 minor 37 references

Building Fast, Evaluating Slow: Pipeline Choices Dominate Autointerpretability Score Variance

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Evaluation pipeline choices, not SAE architecture, dominate autointerpretability score variance.

desk verdict A genuinely useful empirical study of SAE evaluation instability whose headline variance-decomposition claim outruns its statistics; R2 and R3 stand, R1 needs rework before the paper can support field-wide conclusions. read the letter →

arxiv 2607.19386 v1 pith:YCMUDDJB submitted 2026-07-01 cs.LG cs.CL

Sinie van der Ben , Neele Roch , Anna Hedström , Mennatallah El-Assady This is my paper · ORCID
classification cs.LGcs.CL
keywords sparseautoencodersautointerpretabilityevaluationreliabilityvariancedecompositionintraclasscorrelationcoefficientfeaturerankingstabilityLLM-as-judgeSAEinterpretability
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

Autointerpretability scores—where one language model explains what a sparse-autoencoder feature detects and a second language model scores the explanation—are the field's main instrument for comparing sparse autoencoders (SAEs). This paper tries to establish that those scores mostly reflect pipeline choices rather than the features or architectures they are meant to compare. Across four metrics, two large language models, and three SAE architectures, the authors test whether the combined variance from corpus, random example draw, explainer, and phrasing exceeds the variance attributable to SAE architecture; they report it does for every metric. A sympathetic reader would care because if this holds, published cross-paper comparisons of SAEs based on such scores may be comparing evaluation setups instead of architectures.

What carries the argument

The carrying object is a variance-components model fit by restricted maximum likelihood. For score $s_{ijklm}$ of feature $i$ under corpus $j$, draw $k$, explainer $l$, and architecture $m$, the model writes $s_{ijklm} = \mu + \alpha_i + \beta_j + \gamma_k + \delta_l + \epsilon_m$ with independent mean-zero random effects, giving total variance $\sigma^2_{\text{total}} = \sigma^2_{\alpha} + \sigma^2_{\beta} + \sigma^2_{\gamma} + \sigma^2_{\delta} + \sigma^2_{\epsilon}$. The test is whether $\sigma^2_{\text{method}} = \sigma^2_{\beta} + \sigma^2_{\gamma} + \sigma^2_{\delta} + \sigma^2_{\text{phrasing}}$ exceeds $\sigma^2_{\epsilon} = \sigma^2_{\text{arch}}$. Two supporting measures track within-feature instability: the intraclass correlation coefficient (the proportion of total variance that is stable between features) and top-k Jaccard lift (overlap of top-k feature sets normalized by chance). The variance decomposition gives the global verdict; the ICC and Jaccard lift localize which pipeline axis breaks which metric.

What would settle it

A fully crossed experiment that scores the same set of features across three SAE architectures and all pipeline axes (corpora, draws, explainers, phrasings) in one design would settle the claim: if the architecture variance component reaches or exceeds the summed pipeline components, the central claim collapses. A lighter check would be to recompute the decomposition with a different scorer model, since the paper fixes one scorer.

Watch

Extended reading notes

Core claim

The central claim is that the reliability precondition for cross-paper autointerpretability comparison fails: for simulation, detection, fuzzing, and purity scores on both Pythia-160M and Apertus-8B, methodological variance ($\sigma^2_{\text{method}} = \sigma^2_{\text{corpus}} + \sigma^2_{\text{draw}} + \sigma^2_{\text{explainer}} + \sigma^2_{\text{phrasing}}$) exceeds architectural variance ($\sigma^2_{\text{arch}}$). The architecture term contributes less than 7% of total variance in most cases, while corpus and draw dominate for simulation and purity. Each metric has its own failure profile: detection is the most stable, fuzzing is unreliable under every variation, and simulation and purity are highly sensitive to which random examples are drawn. Mean scores look stable across conditions, but top-k feature rankings have near-chance overlap across corpus and draw conditions, and explanation similarity stays high even when scores move substantially, so similarity cannot serve as a cheap diagnostic. The paper concludes that autointerpretability scores should not be treated as a stable basis for cross-SAE comparison without running a Stability Check and reporting the Minimum Reporting Checklist.

Load-bearing premise

The central result assumes that the variance in scores can be cleanly divided into independent contributions from pipeline choices and from SAE architecture, even though the architecture effect is measured in a separate experiment while the pipeline effects are measured without varying architecture.

Editorial extensions

If this is right

  • Cross-architecture comparisons made with one corpus and one random draw cannot reliably attribute score differences to architecture; the observed effect sizes are small and inconsistent in direction.
  • Feature selection applications such as steering, circuit analysis, and manual inspection inherit the ranking instability: the top-k set on one draw or corpus can be near-chance relative to another.
  • Multi-metric evaluation is necessary because no single metric is reliable everywhere: detection is most stable, fuzzing unreliable across conditions, and simulation and purity are sensitive to draw.
  • Explanation similarity (cosine or BERTScore) cannot be used as a diagnostic for score instability, since it stays high even when scores vary.
  • A Stability Check (ICC at least 0.5 and top-k Jaccard lift at least 2x) is the minimum bar before interpreting scores as architectural evidence.

Reading between the lines

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

  • If the finding generalizes, many existing single-corpus, single-draw SAE comparisons in the literature would need to be treated as pipeline-dependent; the paper itself does not re-evaluate prior results.
  • A natural extension is to vary the scorer model and the SAE layer, which the paper fixed; including those axes might make the pipeline-variance dominance even larger.
  • The proposed Stability Check could be run prospectively as a reporting standard: authors would report ICC and Jaccard lift alongside mean scores, and readers could discount comparisons that fail the thresholds.
  • The dissociation between stable explanation similarity and unstable scores suggests the noise enters through the held-out evaluation sequences, pointing toward the scoring stage rather than the explanation stage as the main site of measurement error.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper studies the reliability of autointerpretability scores for sparse autoencoder (SAE) features, where an LM explains a feature and a second LM scores the explanation. The authors systematically vary four methodological axes (evaluation corpus, random draw of activating examples, explainer LLM, and explanation paraphrasing) plus SAE architecture, across two base models (Pythia-160M, Apertus-8B) and four metrics (simulation, detection, fuzzing, purity). They report three main findings: R1) methodological variance collectively exceeds architectural variance; R2) each metric has a distinct instability profile, with detection most stable and fuzzing unreliable; R3) top-k feature rankings are unstable across corpus and draw conditions even when mean scores are stable. They also show that explanation similarity is not a diagnostic for score instability and propose a Stability Check and a Minimum Reporting Checklist for practitioners.

Significance. If the central claims hold, this is a timely and practically important result for the SAE interpretability community, because cross-paper and cross-architecture comparisons of autointerpretability scores are common and often used as evidence in the SAE usefulness debate. The paper's strengths include a broad experimental grid (two models, four metrics, four method axes plus architecture), explicit LLM prompts and protocol details, honest limitation statements, and a concrete, low-cost Stability Check that many groups could adopt. The paper also reports a clean falsifiable observation: explanation semantic similarity stays high even when scores are unstable, ruling out a cheap diagnostic. However, the headline R1 claim rests on a variance decomposition fitted to a non-crossed, partially unmatched design, and the manuscript does not provide enough statistical detail to establish that the comparison in Eq. (3) is identifiable. That issue is load-bearing and needs to be addressed before the main claim can be accepted.

major comments (3)
  1. [C.4 / Table 14 / Eq. (7)-(9)] The central hypothesis in Eq. (3) is tested using variance components estimated from a design that is neither fully crossed nor matched at the architecture level. Method components (corpus, draw, explainer, phrasing) are estimated on BatchTopK features within Experiments 1 and 2, whereas sigma2_arch is estimated in Experiment 3 from three architectures whose features are not matched across architectures and where all method factors are fixed. Section C.4 states only that 'partially overlapping subsets' are used and gives no fitting procedure, identifiability argument, or sensitivity analysis. Since the architecture variance component is identified from differences among only three architecture group means, the narrow confidence intervals in Table 14 (e.g., arch = 2% [1,5]) are not self-evidently plausible without additional detail. As written, R1's claim that methodological variance exceeds architectural variance is not established by the reported analysis.
  2. [B.8.1 / B.8.2 (Fuzzing)] The Fuzzing metric is described as token-level, with an AUC over token-level predictions, but the scoring prompt in B.8.2 asks the LLM to return one binary label per example ('return 1 if the sentence is correctly labeled or 0 if the tokens are mislabeled'). The paper never states how token-level predictions are obtained from these sequence-level labels. Without this detail, the fuzzing scores underlying Tables 2 and 14 and the R2 claim that fuzzing is unreliable across all conditions cannot be reproduced or interpreted.
  3. [Section 3 (R3) / Eq. (4) / Table 7] The top-k Jaccard lift analysis compares rankings across conditions with different numbers of evaluable features. Table 7 shows that the corpus condition substantially changes the feature pool, reducing the number of features with ten activating examples by about half on Apertus-8B and producing different overlap sets across corpora. The chance baseline in Eq. (4) assumes a fixed pool of size N, but when the eligible feature set is condition-dependent, the expected Jaccard under random selection is not k/(2N-k). The paper does not specify whether the lift is computed only on the common feature intersection or on the union of condition-specific pools; if the latter, R3's near-chance results may be an artifact of a misspecified chance model rather than genuine ranking instability.
minor comments (5)
  1. [Section 3, R1] The main text says the architecture contribution on Apertus-8B has a peak of 11% for DETECTION, but Table 14 reports sigma2_arch = 15% [9,23] for Apertus-8B DETECTION; the text and table should be reconciled.
  2. [B.8.1 (Purity)] The text states 'The last metric, PURITY, does not depend on any LLM', but the immediately following description and the scoring prompt in B.8.2 describe PURITY as rated by a scorer LLM; one of these statements is incorrect.
  3. [Table 2 / Section 3, R2] Negative ICC values are interpreted as 'less consistent than chance'; this phrasing is statistically inaccurate, since negative estimates can arise from sampling variability in a one-way random-effects model. The results should be described as poor reliability rather than worse-than-chance consistency.
  4. [D.3 / D.2] The manuscript contains incomplete sentences: Section D.3 ends with 'The time for one experiment on PYTHIA-160M is roughly .' and Section D.2 ends with 'limitatios'; these should be completed or removed.
  5. [Section 4.1 / Table 3] The Stability Check thresholds (ICC at least 0.5, Jaccard lift at least 2x) are presented as prescriptive enough to 'pass' or 'fail' a metric. The authors acknowledge in Section 5 that these are pragmatic cutoffs, but the main-text presentation should more prominently frame them as heuristics rather than validated standards.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the variance-decomposition and stability results are empirical observations, not derivations from fitted parameters or self-cited theorems.

full rationale

The paper's central claims (R1-R3) are empirical measurements: ICC, Top-k Jaccard lift, and REML variance components are computed from experimental score data, and the paper never fits a parameter to a subset of outcomes and then re-presents the same quantity as a prediction. The variance decomposition (Eqs. 7-10) estimates separate random-effect components from distinct experiments; the inequality sigma2_method > sigma2_arch is a comparison of fitted components, not a relation imposed by construction. The Stability Check thresholds (ICC >= 0.5, lift >= 2x) are pragmatic cutoffs citing standard reliability conventions (Koo & Li; Kuncheva) and are not fed back into the variance analysis. No self-citations are load-bearing; the prior-work citations (e.g., SAEBench, [27], [28]) supply datasets, metrics, and prompting procedures but do not justify the paper's conclusions by appeal to the same authors' authority. The acknowledged non-crossed design and three-architecture comparison raise a statistical validity/identifiability question, but that is a correctness concern, not circularity; the paper itself flags the partially crossed design in Appendix C.4 and limits the scope of the claim in Section 5.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical or latent entities are postulated. The Stability Check and Minimum Reporting Checklist are procedural recommendations, not invented entities. The key free parameters are the activation thresholds and the 10-example inclusion rule, both hand-chosen and consequential for feature selection. The main axioms concern the statistical identifiability of the variance decomposition, the transfer of standard ICC thresholds to LLM-scored metrics, the reliability of LLM output parsing, and the absence of scorer-pipeline interaction effects.

free parameters (3)
  • Activation threshold for Pythia-160M features = 0.1
    Chosen by hand to yield roughly 250 active features per condition; determines which features enter the analysis and thus affects all ICC and variance estimates (Appendix B.3).
  • Activation threshold for Apertus-8B features = 100
    Chosen to match the activation scale of Apertus-8B; affects which features are included and the comparability of the two model analyses (Appendix B.3).
  • Minimum of 10 activating examples per feature = 10
    Adopted from prior work (refs 27, 28) to generate explanations; excludes many features, e.g., 51% on Apertus-8B (Table 7), introducing selection bias in the corpus experiment (B.3.1).
assumptions (4)
  • domain assumption Variance components are independent and identifiable from a partially crossed design
    The REML model in C.4 assumes independent mean-zero random effects, but the architecture term is estimated from a separate experiment (Exp. 3) while method terms come from within-architecture experiments, so the variance components may not be cleanly separable or comparable.
  • domain assumption ICC interpretation thresholds (>=0.75 good, 0.50-0.75 moderate, <0.50 poor) apply to these score distributions
    These thresholds from Koo and Li (2016) are standard in reliability research; the paper notes lower values can be acceptable depending on context (C.1), so applying them to LLM-produced scores is an assumption.
  • ad hoc to paper LLM outputs are reliably parsed into the required JSON and Python-list structures
    The scoring prompts require the scorer to return JSON arrays or Python lists; no parsing-failure analysis is reported. If malformed outputs are silently dropped, the score distributions would be shifted.
  • domain assumption The scorer model (Gemini Flash) does not interact with the studied variation axes
    Scorer variation is excluded by design, but the scorer is also the paraphraser in Exp. 2c, which could inflate paraphrase stability because paraphraser and scorer share representational biases (B.6).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Building Fast, Evaluating Slow: Pipeline Choices Dominate Autointerpretability Score Variance." pith.science (2026). https://pith.science/paper/YCMUDDJB

@misc{pith2026260719386,
  author       = {Pith},
  title        = {Pith review of: Building Fast, Evaluating Slow: Pipeline Choices Dominate Autointerpretability Score Variance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YCMUDDJB}},
  note         = {Machine review of arXiv:2607.19386}
}
read the original abstract

Cross-paper comparison of sparse autoencoder (SAE) interpretability often relies on autointerpretability scores. In this evaluation pipeline, a language model (LM) explains each feature, and another LM scores the explanation. For these comparisons to be meaningful, scores must reflect stable properties of the features rather than confounding aspects of the evaluation pipeline. Through systematic experiments across four metrics (simulation, detection, fuzzing, purity), two models (Pythia-160M, Apertus-8B), and four axes of methodological variation, we show that this assumption does not hold. Specifically, we find that R1) methodological variance collectively exceeds architectural variance across all metrics and tested models; R2) each metric exhibits a distinct instability profile, with detection being the most stable and fuzzing unreliable across all conditions; R3) top-k feature rankings do not stay consistent across corpus and draw conditions, masking per-feature instability behind stable mean scores; a failure that cannot be detected by monitoring explanation similarity alone. These findings suggest that cross-paper comparisons based on autointerpretability scores may reflect pipeline differences rather than architectural differences, with implications for the ongoing debate on SAE utility. More broadly, unreliable evaluation slows progress in interpretability research at a time when reliable tools for understanding AI systems are needed. To support evaluation, we contribute a variance decomposition approach, a Stability Check, and a Minimum Reporting Checklist.

Figures

Figures reproduced from arXiv: 2607.19386 by the authors.

Figure 1
Figure 1. Overview of Automated Interpretability Pipeline with the insights location. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Variance decomposition across five experiments for [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The Jaccard lift for all four metrics across all conditions. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Cohen’s d effect sizes between SAE architecture pairs on Pythia-160M (top) and Apertus-8B (bottom). We interpret Cohen’s d effect sizes as follows: d = 0.2 indicates a small ef￾fect, d = 0.5 a medium effect, and d = 0.8 a large effect [31]. Most effects are small to ne…
Figure 5
Figure 5. Figure 5: Metric score trajectories of 20 features across three paraphrases. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Per-feature metric scores for Gemini Flash vs.GPT4o-mini; each point represents one feature. thereby which features appear interpretable. CORPUS variation also determines which features can be evaluated at all. The number of Pythia-160M features with at least one activ…
Figure 7
Figure 7. Figure 7: Activation diagnostics for features, considering all four datasets. The first column of figures [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: ICC heatmap with the delta of the mean for each corpus and metric [PITH_FULL_IMAGE:figures/full_fig_p030_8.png]
Figure 9
Figure 9. Figure 9: Corpus heatmap with the delta of the mean for each corpus combination and metric [PITH_FULL_IMAGE:figures/full_fig_p032_9.png]
Figure 10
Figure 10. Figure 10: The score across draws for each feature. One line represents one feature. Both [PITH_FULL_IMAGE:figures/full_fig_p033_10.png]
Figure 11
Figure 11. Figure 11: Mean metric scores for unrestricted and W10 (10-word-constrained) explanations, per [PITH_FULL_IMAGE:figures/full_fig_p033_11.png]
Figure 12
Figure 12. Figure 12: The score of the original paraphrase vs. the average of the other paraphrases. [PITH_FULL_IMAGE:figures/full_fig_p034_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 26 canonical work pages

  1. [1]

    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, et al. Pythia: a suite for analyzing large language models across training and scaling. In Proceedings of the 40th International Conference on Machine Learning, pages 2397–2430, 2023

    2023

  2. [2]

    Language models can explain neurons in language models, 2023

    Steven Bills, Nick Cammarata, Dan Mossing, Henk Tillman, Leo Gao, Gabriel Goh, Ilya Sutskever, Jan Leike, Jeff Wu, and William Saunders. Language models can explain neurons in language models, 2023

    2023

  3. [3]

    Towards monosemanticity: Decomposing language models with dictionary learning.Transformer Circuits Thread, 2023

    Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Con- erly, Nick Turner, Cem Anil, Carson Denison, Amanda Askell, Robert Lasenby, Yifan Wu, Shauna Kravec, Nicholas Schiefer, Tim Maxwell, Nicholas Joseph, Zac Hatfield-Dodds, Alex Tamkin, Karina Nguyen, Brayden McLean, Josiah E Burke, Tristan Hume, Shan Carter, Tom Henighan, and ...

    2023

  4. [4]

    Learning multi-level features with matryoshka sparse autoencoders

    Bart Bussmann, Noa Nabeshima, Adam Karvonen, and Neel Nanda. Learning multi-level features with matryoshka sparse autoencoders. InProceedings of the Forty-second International Conference on Machine Learning, 2025

    work page 2025

  5. [5]

    Improving steering vectors by targeting sparse autoencoder features.arXiv preprint arXiv:2411.02193, 2024

    Sviatoslav Chalnev, Matthew Siu, and Arthur Conmy. Improving steering vectors by targeting sparse autoencoder features.arXiv preprint arXiv:2411.02193, 2024

    arXiv 2024

  6. [6]

    Sparse but wrong: Incorrect L0 leads to incorrect features in sparse autoencoders.arXiv preprint arXiv:2508.16560, 2025

    David Chanin and Adrià Garriga-Alonso. Sparse but wrong: Incorrect L0 leads to incorrect features in sparse autoencoders.arXiv preprint arXiv:2508.16560, 2025

    arXiv 2025

  7. [7]

    A is for absorption: Studying feature splitting and absorption in sparse autoencoders, 2024

    David Chanin, James Wilken-Smith, Tomáš Dulka, Hardik Bhatnagar, and Joseph Bloom. A is for absorption: Studying feature splitting and absorption in sparse autoencoders, 2024

    work page 2024

  8. [8]

    GitHub Code dataset

    CodeParrot. GitHub Code dataset. Hugging Face Datasets, 2022. Accessed: March 2025

    work page 2022

Show all 37 references
  1. [9]

    Sparse autoencoders find highly interpretable features in language models

    Hoagy Cunningham, Robert Huben, Logan Riggs Smith, Aidan Ewart, and Lee Sharkey. Sparse autoencoders find highly interpretable features in language models. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  2. [10]

    Wikimedia downloads

    Wikimedia Foundation. Wikimedia downloads

  3. [11]

    The pile: An 800gb dataset of diverse text for language modeling, 2020

    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, 2020. 10

    2020

  4. [12]

    Scaling and evaluating sparse autoencoders

    Leo Gao, Tom Dupre la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Ilya Sutskever, Jan Leike, and Jeffrey Wu. Scaling and evaluating sparse autoencoders. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  5. [13]

    Apertus: Democratizing open and compliant LLMs for global language environments, 2025

    Alejandro Hernández-Cano, Alexander Hägele, Allen Hao Huang, Angelika Romanou, Antoni- Joan Solergibert, Barna Pasztor, Bettina Messmer, Dhia Garbaya, Eduard Frank ˇDurech, Ido Hakimi, Juan García Giraldo, Mete Ismayilzada, Negar Foroutan, Skander Moalla, Tiancheng Chen, Vinko...

    2025

  6. [14]

    Open source automated interpretability for sparse autoencoder features

    Caden Juang, Gonçalo Paulo, Jonathan Drori, and Nora Belrose. Open source automated interpretability for sparse autoencoder features. EleutherAI Blog, July 2024

    2024

  7. [15]

    Are sparse autoencoders useful? a case study in sparse probing

    Subhash Kantamneni, Joshua Engels, Senthooran Rajamanoharan, Max Tegmark, and Neel Nanda. Are sparse autoencoders useful? a case study in sparse probing. InForty-second International Conference on Machine Learning, 2025

    2025

  8. [16]

    SAEBench: A comprehensive benchmark for sparse autoencoders in language model interpretability

    Adam Karvonen, Can Rager, Johnny Lin, Curt Tigges, Joseph Isaac Bloom, David Chanin, Yeu-Tong Lau, Eoin Farrell, Callum Stuart McDougall, Kola Ayonrinde, Demian Till, Matthew Wearden, Arthur Conmy, Samuel Marks, and Neel Nanda. SAEBench: A comprehensive benchmark for sparse au...

    2025

  9. [17]

    A guideline of selecting and reporting intraclass correlation coefficients for reliability research.Journal of chiropractic medicine, 15(2):155–163, 2016

    Terry K Koo and Mae Y Li. A guideline of selecting and reporting intraclass correlation coefficients for reliability research.Journal of chiropractic medicine, 15(2):155–163, 2016

    2016

  10. [18]

    Sanity checks for sparse autoencoders: Do saes beat random baselines?arXiv preprint arXiv:2602.14111, 2026

    Anton Korznikov, Andrey Galichin, Alexey Dontsov, Oleg Rogov, Ivan Oseledets, and Elena Tutubalina. Sanity checks for sparse autoencoders: Do saes beat random baselines?arXiv preprint arXiv:2602.14111, 2026

    2026

  11. [19]

    A stability index for feature selection

    Ludmila I Kuncheva. A stability index for feature selection. InArtificial intelligence and applications, pages 421–427, 2007

    2007

  12. [20]

    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. In Vera Demberg, Kentaro Inui, and Lluís Marquez, editors,Proceedings of the 19th Conference of the European Chapter of the A...

    2026

  13. [21]

    Intraclass correlation–a discussion and demonstration of basic features.PloS one, 14(7):e0219854, 2019

    David Liljequist, Britt Elfving, and Kirsti Skavberg Roaldsen. Intraclass correlation–a discussion and demonstration of basic features.PloS one, 14(7):e0219854, 2019

    2019

  14. [22]

    Towards principled evaluations of sparse autoencoders for interpretability and control

    Aleksandar Makelov, Georg Lange, and Neel Nanda. Towards principled evaluations of sparse autoencoders for interpretability and control. InThe Thirteenth International Conference on Learning Representations, 2025. 11

    2025

  15. [23]

    Sparse feature circuits: Discovering and editing interpretable causal graphs in language models

    Samuel Marks, Can Rager, Eric J Michaud, Yonatan Belinkov, David Bau, and Aaron Mueller. Sparse feature circuits: Discovering and editing interpretable causal graphs in language models. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  16. [24]

    Rethinking evaluation of sparse autoencoders through the representation of polysemous words

    Gouki Minegishi, Hiroki Furuta, Yusuke Iwasawa, and Yutaka Matsuo. Rethinking evaluation of sparse autoencoders through the representation of polysemous words. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  17. [25]

    OpenAI, :, Aaron Hurst, Adam Lerer, Adam P. Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, Aleksander M ˛ adry, Alex Baker-Whitcomb, Alex Beutel, Alex Borzunov, Alex Carney, Alex Chow, Alex Kirillov, Alex Nichol, Alex ...

    2024

  18. [26]

    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. InProceedings of the International Conference on Learning Representations, 2026

    2026

  19. [27]

    Automatically interpreting millions of features in large language models

    Gonçalo Santos Paulo, Alex Troy Mallen, Caden Juang, and Nora Belrose. Automatically interpreting millions of features in large language models. InProceedings of the Forty-second International Conference on Machine Learning, 2025

    2025

  20. [28]

    FADE: Why bad descriptions happen to good features

    Bruno Puri, Aakriti Jain, Elena Golimblevskaia, Patrick Kahardipraja, Thomas Wiegand, Woj- ciech Samek, and Sebastian Lapuschkin. FADE: Why bad descriptions happen to good features. In Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar, editors, Findin...

    2025

  21. [29]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer, 2023

    2023

  22. [30]

    Improving sparse decomposition of language model activations with gated sparse autoencoders

    Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Tom Lieberum, Vikrant Varma, Janos Kramar, Rohin Shah, and Neel Nanda. Improving sparse decomposition of language model activations with gated sparse autoencoders. InThe Thirty-eighth Annual Conference on Neural Information ...

    2024

  23. [31]

    Sullivan and Richard Feinn

    Gail M. Sullivan and Richard Feinn. Using effect size—or why the P value is not enough. Journal of Graduate Medical Education, 4(3):279–282, sep 2012

    2012

  24. [32]

    Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, Soroosh Mariooryad, Yifan Ding, Xinyang Geng, Fred Alcober, Roy Frostig, Mark Omernick, Lexi Walker, Cosmin Paduraru, Christina Sorokin, A...

    2024

  25. [33]

    Daniel Freeman, Theodore R

    Adly Templeton, Tom Conerly, Jonathan Marcus, Jack Lindsey, Trenton Bricken, Brian Chen, 16 Adam Pearce, Craig Citro, Emmanuel Ameisen, Andy Jones, Hoagy Cunningham, Nicholas L Turner, Callum McDougall, Monte MacDiarmid, C. Daniel Freeman, Theodore R. Sumers, Edward Rees, Josh...

    2024

  26. [34]

    Persona features control emergent misalignment.arXiv preprint arXiv:2506.19823, 2025

    Miles Wang, Tom Dupré la Tour, Olivia Watkins, Alex Makelov, Ryan A Chi, Samuel Mis- erendino, Jeffrey Wang, Achyuta Rajaram, Johannes Heidecke, Tejal Patwardhan, et al. Persona features control emergent misalignment.arXiv preprint arXiv:2506.19823, 2025

    2025

  27. [35]

    Axbench: Steering LLMs? even simple base- lines outperform sparse autoencoders

    Zhengxuan Wu, Aryaman Arora, Atticus Geiger, Zheng Wang, Jing Huang, Dan Jurafsky, Christopher D Manning, and Christopher Potts. Axbench: Steering LLMs? even simple base- lines outperform sparse autoencoders. InForty-second International Conference on Machine Learning, 2025

    2025

  28. [36]

    Mathematical equations and formulas

    Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. InInternational Conference on Learning Representations, 2020. 17 A Related Work Evaluation GapThe field has recognized that improvements in the spa...

    2020

  29. [37]

    Justification: There is no crowdsourcing nor research with human subjects

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals...

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.