Pith. sign in

REVIEW 2 major objections 5 minor 43 references

Statistically Grounded Sparse-Feature Interventions for Activation-Space Control in Large Language Models

T0 review · 2 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read Consensus-ranking sparse features by three classical statistics and weighting them by Cohen's d steers LLM behavior measurably — but control is sharply localized, and raw shifts overstate quality-preserving success.

desk verdict Honest, useful SAE-steering paper whose main claim (raw shift overstates usable steering) holds up, but the selection-to-injection site mismatch is a real gap that needs addressing before the interpretability story lands. read the letter →

arxiv 2607.19364 v1 pith:SQNN7NB4 submitted 2026-06-05 cs.AI cs.CL

Oshayer Siddique , J. M Areeb Uzair Alam , Md Jobayer Rahman Rafy , Syed Rifat Raiyan , Hasan Mahmud , Md Kamrul Hasan This is my paper
classification cs.AIcs.CL
keywords activationsteeringsparseautoencodersfeaturerankingBordaconsensusCohen'sdmutualinformationquality-conditionedevaluationlanguagemodelinterpretability
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

This paper tries to establish that transparent, optimization-free activation steering can work: select sparse autoencoder (SAE) features — the sparse, individually interpretable directions into which residual-stream activations decompose — by a consensus of three classical statistics (F-test, mutual information, Cohen's d), weight the steering vector by Cohen's d, and inject it at the last token. Swept across three Gemma-family models, four behavioral domains, and 356 layer–strength configurations, the method shifts target attributes measurably, with a peak +1.16 primary-score gain on logical correctness in the largest model. The broader finding is about how steering should be measured: raw attribute movement overstates usable control, since plain success rates exceed quality-preserving 'clean success' by up to 28 percentage points, and usable steering is sharply localized by model, domain, layer, and strength. A sympathetic reader would care because this supplies an auditable, statistics-only baseline for a fragmented technique, and because it argues that any steering evaluation — including those from optimization-based methods — should report quality-conditioned success alongside raw shift.

What carries the argument

The machinery is the consensus-ranked feature direction, built in three steps: a six-condition quality filter on log1p-transformed SAE activations (activity ≥1%, positive Cohen's d with bootstrap lower bound above zero, |d| ≥ 0.2, MI bootstrap CV < 0.5, FDR q < 0.05 on at least one test); an unweighted two-tier Borda consensus over three complementary statistics — F-test (linear separation), KSG mutual information (nonlinear/threshold dependence), Cohen's d (signed, variance-standardized separation) — with no learned weights; and a steering vector equal to the Cohen's-d-weighted sum of selected decoder rows, normalized and injected at the last token with activation-norm scaling α. Cohen's d

What would settle it

Take the LOGIC-domain steering in Gemma 2 9B at layer 19 (the +1.16 setting) and ablate the selected features one by one at inference — zero out each top-ranked SAE feature's activation, or patch in activations from an unsteered forward pass — and measure whether the primary-score delta collapses. If removing the single highest-weight feature leaves the shift essentially intact, the statistical ranking is selecting correlates, not causes. A complementary control is a label-permutation test: if running the full filter-and-consensus pipeline on shuffled contrast labels still yields 'steering' wi

Watch

Extended reading notes

Core claim

The paper's central claim: a fully auditable, optimization-free sparse-feature steering pipeline works. A six-condition filter prunes the 16,384-feature space; survivors are ranked by F-test, KSG mutual information, and Cohen's d, merged by an unweighted two-tier Borda consensus; the steering direction is the Cohen's-d-weighted sum of selected decoder rows, motivated by Fisher-LDA under approximate feature decorrelation. The method shifts target attributes measurably — logical correctness reaches +1.16 primary delta in Gemma 2 9B — but the broader finding is the raw-versus-clean gap: raw win rate overstates quality-preserving clean success by 4.1–28 percentage points. Steering is model- and

Load-bearing premise

The load-bearing premise is that features ranked by statistical association on a contrastive selection set are the right causal handles to inject at the last token at inference — the paper's own limitations state that the causal validation is statistical, not mechanistic, and that associational ranking alone does not establish that activating these features causes the observed shift; if the statistics select features correlated with domain vocabulary or style rather than feat

Editorial extensions

If this is right

  • Steering evaluations should report quality-conditioned success (clean success) alongside raw shift; raw primary-win rates overstate usable control by up to 28 percentage points, so headline deltas alone are misleading.
  • The optimal intervention layer is not universal: no monotone relation exists between layer depth and effect (Spearman ≈ 0 across 344 configurations), so layer must be treated as a per-domain, per-model hyperparameter.
  • Steering strength interacts non-monotonically with effect — the largest logic shift occurs at the smallest α (0.1) — and large α pushes activations out of distribution, producing repetition and incoherence that further steering cannot recover.
  • Multi-layer, budgeted composition should be judged on clean-success reliability rather than raw shift: on the strongest domain it lifts clean success from 24% to 31% while dampening the raw delta, and on 9 of 12 cells it dilutes the best single-layer shift.
  • In the tested settings, 600 gradient steps of a supervised steering objective from the Cohen's-d initialization reach similar projection behavior, indicating the optimization-free direction sits near the practical ceiling; learned steering objectives add validation burden without evident gain here.

Reading between the lines

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

  • One testable extension the authors leave implicit: because selection is purely associational, the method's interpretability — feature labels cluster on conditional-logic, implication, and causation directions for LOGIC — could reflect surface markers of the contrast corpus rather than mechanisms upstream of the behavior; per-feature ablation or activation patching would separate the two.
  • A transfer prediction: the paper's near-zero depth–effect correlation across the sweep suggests that no canonical 'steering layer' exists for other model families either; a matching layer sweep on a non-Gemma model with a public sparse autoencoder would confirm whether the localization is architectural or an artifact of Gemma's training.
  • A safety artifact that follows from the paper's own dual-use warning: since flipping the sign of d steers toward the anti-target, one could build a steering-audit tool that prints the top contributing features and their human-readable labels for any deployed vector, making inversion attempts detectable.
  • A compositional hypothesis: if cumulative perturbation budget, not raw shift, governs reliability, then multi-attribute steering — which the paper leaves open — could inherit the multi-layer budget result, so additive α-budgeting across attributes with clean success as the objective is the natural next experiment rather than naive vector addition.
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.

Referee Report

2 major / 5 minor

Summary. The paper proposes a transparent SAE-based activation-steering pipeline: a six-condition reliability filter, a three-statistic Borda consensus (F-test, KSG mutual information, Cohen's d) for feature ranking, and a Cohen's-d-weighted combination of SAE decoder rows as the steering direction. The method is evaluated across three Gemma models, four behavioral domains, and 356 layer–strength configurations, with a three-judge scoring protocol and human adjudication. The authors report that the method produces measurable domain-specific shifts, that the strongest logical-correctness configuration reaches Δp=+1.16 in Gemma 2 9B, and — more importantly — that raw attribute movement substantially overstates quality-preserving control. They also report that no single configuration passes a strict FDR-corrected joint significance test, and they frame strong configurations as Pareto-efficient candidates with bootstrap support. The paper is explicitly positioned as a transparent, optimization-free baseline rather than a claim of state-of-the-art steering.

Significance. If the findings hold, the paper makes a useful empirical and methodological contribution: it provides a fully transparent, optimization-free SAE-feature steering recipe, and it documents a large configuration sweep showing that steering success is highly localized by model, domain, layer, and strength. The paper also makes a valuable negative point — raw primary-score deltas are not a reliable proxy for usable, quality-preserving control. The release of code and data, the explicit treatment of statistical assumptions, and the honest reporting of the strict-test failure are strengths that make the empirical claims easy to check and extend. However, the causal-handle interpretation is not yet fully supported because of a mismatch between the feature-selection site and the injection site.

major comments (2)
  1. [§2.3 vs. §2.6] Feature selection is computed on max-pooled log1p activations over the full contrastive text (Z_D in §2.3), while the steering vector is injected only at the last-token position (§2.6: h_{:,-1,:} ← h_{:,-1,:} + α||h_{:,-1,:}|| δ̂h). The paper's stated intervention is 'activating these features at inference,' but there is no evidence that the selected features are separable, or even active, at the injection site. If a feature's class-discriminative signal is concentrated in an earlier content token, its decoder row may have little effect at the last token, or may activate a different feature set than the one audited via Neuronpedia (Table 10). This is a necessary condition for the causal-handle interpretation, and it is checkable from the authors' own contrastive data: compute the three selection statistics at the last-token position (or at least report per-token max vs. last-token agreem
  2. [§4.6 / Appendix N.5 / Abstract] The manuscript correctly reports that no configuration survives the strict joint FDR-corrected paired t / Wilcoxon / sign-test criterion, and it frames headline cells as Pareto candidates. However, the Abstract still presents '+1.16 in Gemma 2 9B' as a headline result without this caveat, and the bootstrap CI for that cell is [+0.33, +1.96] (Appendix N.5) — a wide interval. Since the per-configuration claim is not FDR-significant, the abstract and Section 4.1 should either carry the strict-test caveat explicitly or report the strongest configuration as a descriptive Pareto candidate only. This is not a request for new analysis; it is a reporting-consistency issue for the paper's most prominent number.
minor comments (5)
  1. [Appendix B.4] The Fisher LDA derivation contains a duplicated paragraph: the two-sentence 'Fisher LDA maximizes J(w)...' block appears twice verbatim. Remove the duplicate.
  2. [References] The reference list appears to duplicate entries for the same work: Panickssery et al. 2023 and Rimsky et al. 2024 are the same CAA paper, and Zou et al. 2023a/2023b are the same Representation Engineering preprint. Consolidate.
  3. [Table 10] The Neuronpedia audit is shown only for one domain, one model, and one layer (Gemma 2 2B, LOGIC, layer 12). A representative audit for at least one additional model/domain, or an explicit statement that the audit is illustrative only, would strengthen the interpretability claim.
  4. [Appendix M / Eq. (1)] The F-statistic formula in Appendix M is visually garbled by line breaks; the numerator and denominator should be typeset clearly. Also, the text says 'F-statistic and Cohen's d are mean-based' but Cohen's d is a standardized mean difference; consider phrasing.
  5. [§2.4] The six-condition quality filter includes the OR-combined FDR rule (C6) alongside per-statistic FDR control. The text explains this is intentionally not a global FDR guarantee, but the caption of Table 12 says 'Direction significance test' for Cohen's d via Welch's t; clarify that the bootstrap lower-bound condition (C3), not Welch's t, is the directional gate.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the steering vector is built only from contrastive-selection statistics, and all claimed effects are measured on disjoint held-out prompts.

full rationale

The derivation chain is not circular. Feature selection (Section 2.4) uses F-test, KSG MI, and Cohen's d computed only on the contrastive selection set (Z_D and y_D), and the steering vector (Section 2.5) is a Cohen's-d-weighted combination of decoder rows. This is a construction, not a prediction from the same data. The paper's headline effects (primary deltas, clean success) are evaluated on 100 held-out completion prompts per domain that are disjoint from the selection set (Section 3: 'For evaluation we draw 100 held-out completion prompts per domain that are disjoint from the selection set'), with baselines (CAA, RePe, Top PC, ITI) compared under the same protocol. The dual role of Cohen's d (selection and weighting) is an explicit design choice, not a fitted parameter renamed as a prediction. The Fisher-LDA connection is explicitly approximate ('should be interpreted as Fisher-motivated under approximate decorrelation and comparable-variance assumptions, rather than as an exact Fisher-LDA optimum'), so no exact identity is claimed. The multi-layer budget being 'by construction' milder is an acknowledged design constraint, not a hidden circularity. The only substantive gap—that selection statistics are computed on max-pooled activations over the full text while injection is at the last token—is an internal-validity limitation, which the paper partly acknowledges in its Limitations ('Causal validation is statistical, not mechanistic... it does not, on its own, establish that activating these features at inference causes the observed behavioral shift'); this is not a circular reduction because the paper does not define steering effectiveness in terms of the selection statistic. No load-bearing self-citations or imported uniqueness theorems appear; references to Gemma Scope, Neuronpedia, and prior steering methods are external sources. The empirical claims are self-contained against held-out evaluation, so the circularity score is 0.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The pipeline's main degrees of freedom are the sweep hyperparameters (layer, α, K) and the six quality-filter thresholds. The causal-relevance and judge-validity assumptions are acknowledged limitations. No new entities are introduced.

free parameters (5)
  • Steering strength α = swept {0.1, 0.2, 0.3, 0.5, 0.7, 1.0, 1.5, 2.0}; best cells vary (e.g., 0.1 for Gemma 2 9B LOGIC, 2.0 for Gemma 3 4B POLI
    Injection magnitude is a manually swept hyperparameter; headline results depend on selecting the best α per cell.
  • Intervention layer ℓ = probed layers: Gemma 2 2B {12,13,16,19,23}, 9B {19,26,31,38}, Gemma 3 4B {9,17,22,29}; best layer varies per model/domai
    Layer selection is an empirical grid choice; the paper's non-universality claim rests on this grid.
  • Top-K feature count K = 16, 24, or 32
    Number of SAE features composing the steering vector is a hyperparameter over which results are reported.
  • Six-condition quality-filter thresholds = activity ≥1% and ≥5 rows; Cohen's d > 0 with lower bootstrap CI > 0; |d| ≥ 0.2; MI bootstrap CV < 0.5; FDR q < 0.05 (OR-
    These explicit thresholds set the candidate pool; |d|≥0.2 was calibrated downward from 0.5 to stabilize the pool.
  • Multi-layer selection weights = w_l = E*_l / Σ E*_l over 2-3 layers per cell; α_total swept {0.1,...,1.0}
    Multi-layer layers and weights are chosen from single-layer evaluation outcomes on the same held-out prompts, not from a separate validation set.
assumptions (6)
  • domain assumption JumpReLU SAE features on post-MLP residual streams are sufficiently monosemantic/interpretable that decoder rows can be used as residual-stream steering directions.
    Assumed from Gemma Scope/Neuronpedia prior work (§2.3, Appendix I); the paper does not independently validate feature causality.
  • domain assumption Features ranked by F-test, KSG MI, and Cohen's d on max-pooled log1p contrastive activations are the right causal handles for last-token injection.
    Explicitly flagged as unproven in Limitations: 'Causal validation is statistical, not mechanistic.'
  • domain assumption Selected SAE features are approximately decorrelated within class and have comparable pooled variances, making Cohen's-d weighting approximately Fisher-LDA.
    Stated approximation in Appendix B.4; sparsity does not guarantee exact decorrelation.
  • domain assumption The three-judge protocol (Gemini 2.5 Flash/Pro, GPT-5.4) with human adjudication of disagreements yields valid primary/quality scores with no systematic shared miscalibration.
    Acknowledged in Limitations as an unhandled risk if all three judges share bias.
  • standard math F-test, KSG MI estimator, Benjamini-Hochberg FDR, and bootstrap procedures are valid at N=1600 with 16,384 features.
    Standard statistical tools; power and CI derivations in Appendix B.1/M.
  • domain assumption Log1p transform and max-pooling over tokens preserve the behaviorally relevant activation signal better than alternatives.
    Method choice justified by sparsity/heavy tails; not independently established.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Statistically Grounded Sparse-Feature Interventions for Activation-Space Control in Large Language Models." pith.science (2026). https://pith.science/paper/SQNN7NB4

@misc{pith2026260719364,
  author       = {Pith},
  title        = {Pith review of: Statistically Grounded Sparse-Feature Interventions for Activation-Space Control in Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SQNN7NB4}},
  note         = {Machine review of arXiv:2607.19364}
}
abstract

Activation steering offers a lightweight alternative to fine-tuning for behavioral control of large language models, but SAE-based steering methods often rely on learned steering objectives or single-criterion feature selection. We introduce a transparent SAE-feature steering pipeline that first applies a six-condition reliability filter, then ranks sparse features through an unweighted Borda consensus over three complementary statistics: $F$-test, KSG mutual information, and Cohen's $d$. The resulting steering direction is constructed as a Cohen's-$d$-weighted combination of SAE decoder rows, providing an optimization-free direction motivated by Fisher-LDA under approximate SAE-feature decorrelation. Across three Gemma-family models, four behavioral domains, and 356 layer-strength configurations, the method produces measurable domain-specific shifts while revealing a substantial gap between raw attribute movement and quality-preserving generation. In the strongest configuration, logical-correctness steering reaches a primary-score delta of $+1.16$ in Gemma~2 9B; however, our broader finding is that usable steering is highly localized by model, domain, layer, and strength. These results argue that activation-steering evaluations should report quality-conditioned success alongside raw behavioral shift. Our code and data are available at https://github.com/Oshayer-Siddique/LLM-Steering-Using-SAE.

Figures

Figures reproduced from arXiv: 2607.19364 by the authors.

Figure 1
Figure 1. Three independent statistical tests rank SAE [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An overview of our end-to-end pipeline. SENTIMENT and MORAL. Per-domain semantics appear in Appendix C. 2.3 SAE Feature Extraction We use Gemma Scope JumpReLU SAEs (Lieberum et al., 2024; Rajamanoharan et al., 2024) for Gemma 2 and Gemma Scope 2 (Google DeepMind, 2025) for Gemma 3 4B, all of width DSAE=16,384 and hooked into the post-MLP residual stream. For each text we compute H(ℓ) (xi) ∈ R T ×dmodel , apply the S… view at source ↗
Figure 3
Figure 3. Layer–α steering landscape across the three models (rows) and four behavioral domains (columns). Each cell reports the mean primary delta ∆p (steered−baseline) for one (layer, α) configuration; blue indicates positive shift, red negative, saturation tracking magnitude. Two patterns are immediately visible: (i) the operative layer is domain- and model-specific—LOGIC peaks at L19 in Gemma 2 9B but L17 in Gemma 3 4B; M… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Primary-shift versus quality-preservation across all 344 (layer, [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: Per-cell gap between best raw primary-win SR (the proportion of prompts on which the steered output [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 12 linked inside Pith

  1. [1]

    Steering

    Rimsky, Nina and Gabrieli, Nick and Schulz, Julian and Tong, Meg and Hubinger, Evan and Turner, Alexander Matt , booktitle =. Steering. 2024 , pages =

    2024

  2. [2]

    arXiv preprint arXiv:2308.10248 , year =

    Steering Language Models with Activation Engineering , author =. arXiv preprint arXiv:2308.10248 , year =

    arXiv

  3. [3]

    Representation Engineering: A Top-Down Approach to

    Zou, Andy and Phan, Long and Chen, Sarah and Campbell, James and Guo, Phillip and Ren, Richard and Pan, Alexander and Yin, Xuwang and Mazeika, Mantas and Dombrowski, Ann-Kathrin and others , journal =. Representation Engineering: A Top-Down Approach to. 2023 , url =

    2023

  4. [4]

    arXiv preprint arXiv:2406.11717 , year =

    Refusal in Language Models Is Mediated by a Single Direction , author =. arXiv preprint arXiv:2406.11717 , year =

    arXiv

  5. [5]

    Transformer Circuits Thread , year =

    Toy Models of Superposition , author =. Transformer Circuits Thread , year =

  6. [6]

    Transformer Circuits Thread , year =

    Towards Monosemanticity: Decomposing Language Models with Dictionary Learning , author =. Transformer Circuits Thread , year =

  7. [7]

    Scaling Monosemanticity: Extracting Interpretable Features from

    Templeton, Adly and Conerly, Tom and Marcus, Jonathan and Lindsey, Jack and Bricken, Trenton and Chen, Brian and Pearce, Adam and Citro, Craig and Ameisen, Emmanuel and Jones, Andy and others , journal =. Scaling Monosemanticity: Extracting Interpretable Features from. 2024 , url =

    2024

  8. [8]

    Proceedings of the International Conference on Learning Representations (ICLR) , year =

    Sparse Autoencoders Find Highly Interpretable Features in Language Models , author =. Proceedings of the International Conference on Learning Representations (ICLR) , year =

Show all 43 references
  1. [9]

    Proceedings of the International Conference on Learning Representations (ICLR) , year =

    Scaling and Evaluating Sparse Autoencoders , author =. Proceedings of the International Conference on Learning Representations (ICLR) , year =

  2. [10]

    Jumping Ahead: Improving Reconstruction Fidelity with

    Rajamanoharan, Senthooran and Lieberum, Tom and Sonnerat, Nicolas and Conmy, Arthur and Varma, Vikrant and Kram. Jumping Ahead: Improving Reconstruction Fidelity with. arXiv preprint arXiv:2407.14435 , year =

    arXiv

  3. [11]

    Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on

    Lieberum, Tom and Rajamanoharan, Senthooran and Conmy, Arthur and Smith, Lewis and Sonnerat, Nicolas and Varma, Vikrant and Kram. Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on. BlackboxNLP Workshop at EMNLP 2024 , year =

    2024

  4. [12]

    2025 , howpublished =

    Announcing. 2025 , howpublished =

    2025

  5. [13]

    2025 , pages =

    He, Zirui and Jin, Mingyu and Shen, Bo and Payani, Ali and Zhang, Yongfeng and Du, Mengnan , journal =. 2025 , pages =

    2025

  6. [14]

    Enhancing

    Wang, Anyi and Wu, Xuansheng and Shu, Dong and Ma, Yunpu and Liu, Ninghao , journal =. Enhancing. 2025 , url =

    2025

  7. [15]

    Controllable

    Fang, Yi and Wang, Wenjie and Xue, Mingfeng and Deng, Boyi and Xu, Fengli and Liu, Dayiheng and Feng, Fuli , journal =. Controllable. 2026 , url =

    2026

  8. [16]

    Physical Review E , volume =

    Estimating Mutual Information , author =. Physical Review E , volume =. 2004 , doi =

    2004

  9. [17]

    1988 , edition =

    Statistical Power Analysis for the Behavioral Sciences , author =. 1988 , edition =

    1988

  10. [18]

    Journal of the Royal Statistical Society: Series B (Methodological) , volume =

    Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing , author =. Journal of the Royal Statistical Society: Series B (Methodological) , volume =

  11. [19]

    Proceedings of the International Conference on Learning Representations (ICLR) , year =

    Discovering Latent Knowledge in Language Models Without Supervision , author =. Proceedings of the International Conference on Learning Representations (ICLR) , year =

  12. [20]

    2025 , howpublished =

    2025

  13. [21]

    OpenAI , year =

    Language Models Can Explain Neurons in Language Models , author =. OpenAI , year =

  14. [22]

    arXiv preprint arXiv:2408.00118 , year =

    Gemma 2: Improving Open Language Models at a Practical Size , author =. arXiv preprint arXiv:2408.00118 , year =

    arXiv

  15. [23]

    arXiv preprint arXiv:2503.19786 , year =

    Gemma 3 Technical Report , author =. arXiv preprint arXiv:2503.19786 , year =

    arXiv

  16. [24]

    arXiv preprint arXiv:2411.02193 , year =

    Improving Steering Vectors by Targeting Sparse Autoencoder Features , author =. arXiv preprint arXiv:2411.02193 , year =

    arXiv

  17. [25]

    Are Sparse Autoencoders Useful?

    Kantamneni, Subhash and Engels, Joshua and Rajamanoharan, Senthooran and Tegmark, Max and Nanda, Neel , journal =. Are Sparse Autoencoders Useful?. 2025 , url =

    2025

  18. [26]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Inference-Time Intervention: Eliciting Truthful Answers from a Language Model , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  19. [27]

    The Twelfth International Conference on Learning Representations , year =

    Sparse Autoencoders Find Highly Interpretable Features in Language Models , author =. The Twelfth International Conference on Learning Representations , year =

  20. [28]

    Aligning

    Hendrycks, Dan and Burns, Collin and Basart, Steven and Critch, Andrew and Li, Jerry and Song, Dawn and Steinhardt, Jacob , journal =. Aligning. 2021 , url =

    2021

  21. [29]

    Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics , year =

    Learning Word Vectors for Sentiment Analysis , author =. Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics , year =

  22. [30]

    2022 , url =

    Han, Simeng and Schoelkopf, Hailey and Zhao, Yilun and Qi, Zhenting and Riddell, Martin and Benson, Luke and Sun, Lucy and Zubova, Ekaterina and Qiao, Yujie and Burtell, Matthew and others , journal =. 2022 , url =

    2022

  23. [31]

    Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics , year =

    Political Ideology Detection Using Recursive Neural Networks , author =. Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics , year =

  24. [32]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Attention is All You Need , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  25. [33]

    arXiv preprint arXiv:2407.12404 , year =

    Analyzing the Generalization and Reliability of Steering Vectors , author =. arXiv preprint arXiv:2407.12404 , year =

    arXiv

  26. [34]

    arXiv preprint arXiv:2501.16496 , year =

    Open Problems in Mechanistic Interpretability , author =. arXiv preprint arXiv:2501.16496 , year =

    arXiv

  27. [35]

    arXiv preprint arXiv:2404.15522 , year=

    Towards Systematic Evaluation of Logical Reasoning Ability of Large Language Models , author=. arXiv preprint arXiv:2404.15522 , year=

    arXiv

  28. [36]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP '24) , year =

    Fulay, Suyash and Brannon, William and Mohanty, Shrestha and Overney, Cassandra and Poole-Dayan, Elinor and Roy, Deb and Kabbara, Jad , title =. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP '24) , year =

    2024

  29. [37]

    arXiv preprint arXiv:2312.06681 , year=

    Steering llama 2 via contrastive activation addition , author=. arXiv preprint arXiv:2312.06681 , year=

    arXiv

  30. [38]

    Advances in Neural Information Processing Systems , volume=

    Inference-time intervention: Eliciting truthful answers from a language model , author=. Advances in Neural Information Processing Systems , volume=

  31. [39]

    arXiv preprint arXiv:2310.01405 , year=

    Representation engineering: A top-down approach to ai transparency , author=. arXiv preprint arXiv:2310.01405 , year=

    arXiv

  32. [40]

    arXiv preprint arXiv:2502.02716 , year=

    A unified understanding and evaluation of steering methods , author=. arXiv preprint arXiv:2502.02716 , year=

  33. [41]

    arXiv preprint arXiv:2507.06261 , year =

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities , author =. arXiv preprint arXiv:2507.06261 , year =. doi:10.48550/arXiv.2507.06261 , url =

    doi arXiv

  34. [42]

    2026 , month = mar, howpublished =

    Introducing. 2026 , month = mar, howpublished =

    2026

  35. [43]

    2026 , month = mar, url =

    2026

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.