Pith. sign in

REVIEW 4 major objections 5 minor 30 references

CircuitSteer: Geometrically Aligned Multi-Layer Steering via Sparse Autoencoder Circuits

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read CircuitSteer claims that multi-layer steering becomes fluency-preserving when SAE features are selected by geometric alignment of decoder directions, and reports being the only method with valid interventions on all eight model–dataset…

desk verdict A genuinely new multi-layer SAE steering recipe whose own ablation table undercuts its 'alignment is necessary' claim, but solid enough to merit a serious referee. read the letter →

arxiv 2608.05732 v1 pith:AMR5IFCE submitted 2026-08-06 cs.LG

Mehrshad Saadatinia , Parsa Razmara , Ardalan Aryashad , Ali Abbasi , Seyedarmin Azizi This is my paper · ORCID
classification cs.LG
keywords sparseautoencodersmulti-layersteeringactivationfeaturecircuitsgeometricalignmentcontrastivespecificityinterpretabilityfluencypreservation
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

The paper proposes CircuitSteer, a training-free method for controlling LLM behavior by intervening on multiple layers at once. It claims that existing single-layer steering methods fail because semantic concepts are spread across layers, and that naive multi-layer combinations fail because the same concept is encoded in geometrically inconsistent directions at different depths. CircuitSteer selects cross-layer "feature flow circuits" by requiring both co-activation and aligned SAE decoder directions, then adds a dense averaged vector at every circuit layer. Across two model families and four behaviors, it is the only method that keeps fluent text while reducing target behavior on every configuration. If the claim holds, geometric alignment turns multi-layer steering from a fluency-destroying operation into a reliable inference-time control mechanism.

What carries the argument

The feature flow circuit $C=(V,E)$ is the central object: nodes are SAE features indexed by layer and unit, and a directed edge $(l,i)\rightarrow(l+1,j)$ is included when the two features co-activate on an input ($c_{l,i}>\tau_{\mathrm{act}}$, $c_{l+1,j}>\tau_{\mathrm{act}}$) and their decoder directions satisfy $\cos(d_{l,i},d_{l+1,j})>\tau_{\mathrm{sim}}=0.10$. A contrastive specificity score $S=P(\cdot|D^+)-P(\cdot|D^-)$ then keeps only edges that appear more often on target prompts than on contrast prompts. Each layer's dense steering vector is the mean decoder direction $\frac{1}{|F_l|}\sum_{i\in F_l} d_{l,i}$, applied simultaneously at every circuit layer. This machinery converts multi-point intervention from a compounding perturbation into a coherent directional shift, because aligned directions reinforce one another across depth.

What would settle it

Run the pipeline on a new model with the alignment sign flipped: require $\cos(d_{l,i},d_{l+1,j}) < -\tau_{\mathrm{sim}}$ instead of above the threshold. If target-behavior reduction survives with preserved fluency, geometric alignment is not necessary. Alternatively, use activation patching to test each retained edge individually; if most edges have near-zero causal effect on the target behavior, the co-activation-plus-cosine indicator is not tracking causal flow.

Watch

Extended reading notes

Core claim

The central claim is that requiring geometric alignment of sparse-autoencoder decoder directions before adding any cross-layer edge is what makes multi-layer steering work. The paper constructs a target-specific subcircuit by keeping edges whose features co-activate on target prompts and whose decoder directions have cosine similarity above $\tau_{\mathrm{sim}}=0.10$, then scores edges by contrastive specificity $P(\cdot|D^+)-P(\cdot|D^-)$. For each circuit layer it averages the decoder directions of participating features into one dense steering vector and applies all vectors simultaneously under a single coefficient $\lambda$. The paper reports that this is the only method with a valid fluency-preserving operating point on all eight model–dataset configurations, and that ablations removing or reversing alignment collapse fluency, particularly on sycophancy.

Load-bearing premise

The load-bearing premise is that co-activation plus a decoder-direction cosine above $\tau_{\mathrm{sim}}=0.10$ approximates causal cross-layer feature flow, even though individual edges are only candidate indicators.

Editorial extensions

If this is right

  • If correct, multi-layer SAE circuit steering is a reliable inference-time control method that requires no gradient updates or weight changes.
  • The same discovery hyperparameters transfer to a new model family and SAE type (Qwen3.5-27B with TopK SAEs) with no retuning, suggesting geometry-based circuit selection is portable across architectures.
  • Behaviors that single-layer methods miss entirely, such as sycophancy on Gemma, become steerable with measurable reduction and near-baseline perplexity.
  • Refusal, a generation-time decision, can be reduced from 89% to 0% by extracting features at the decision point rather than at input encoding.
  • Steering remains selective: MMLU and GSM8K scores move by at most a few points under strong steering, indicating the intervention stays on the target circuit.

Reading between the lines

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

  • If the alignment criterion is a proxy for feature reuse across layers, then the cosine threshold may be measuring how consistently a concept is re-encoded in the residual stream; a testable consequence is that the threshold should correlate with a model's depth-wise representation stability.
  • The contrastive specificity score could be applied to automated circuit discovery for other discrete behaviors, such as chain-of-thought deception or hallucination, where input-level contrastive vectors are known to be weak.
  • Because the paper only claims candidate indicators, a natural next step is to replace the cosine proxy with a patching-based edge score and check whether steering vectors and transfer behavior survive.
  • The decision-point extraction of refusal suggests that the same method could red-team alignment by mapping where refusal is committed, a safety-relevant use the paper itself flags.
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

4 major / 5 minor

Summary. The paper proposes CircuitSteer, a training-free multi-layer steering method that discovers cross-layer SAE feature circuits via co-activation and decoder-direction cosine similarity, synthesizes per-layer dense steering vectors, and intervenes simultaneously at multiple layers. It claims to be the only method producing fluency-preserving valid operating points on all eight model-dataset configurations and that geometric alignment of decoder directions is necessary for stable multi-layer steering. These claims rest on Table 1, the component ablations in Table 2, and supporting robustness experiments in the appendices.

Significance. If the central claims held, the paper would make a useful empirical contribution to inference-time control of LLMs: it ships code, includes extensive ablations (threshold sensitivity, seed stability, contrastive-set sensitivity, layer-count controls, capability-tax measurements), and reports transfer to a third model family with a different SAE type. Those are genuine strengths. However, the main explanatory claim is contradicted by the paper's own Table 2, and the headline comparison in Table 1 is weakened by test-set lambda selection and an author-defined validity window. The empirical contribution is potentially salvageable, but the current claims are not supported by the reported evidence.

major comments (4)
  1. [Section 5.3, Table 2] Table 2 contradicts the central claim that geometric alignment is necessary. At the paper's own fixed coefficients (Gemma λ=-3, Llama λ=-1.5), SINGLELAYER achieves Δ=0.193 with PPL=1.04 on Llama-Sycophancy, outperforming the full method's Δ=0.128 with PPL=1.49; NOGEO achieves Δ=0.088 with PPL=0.97 on Gemma-RTP, outperforming the full method's Δ=0.075 with PPL=1.06. The text states that 'on Llama the full method is the only one to combine meaningful reduction with preserved fluency,' which is directly contradicted by the table. Consequently, the conclusion 'All three ingredients prove necessary' is unsupported; at the reported operating points, geometric alignment is not necessary for the best result in two of the four columns.
  2. [Section 4.1, Table 1] The main comparison uses test-set lambda selection. The setup states that 'steering coefficients λ∈{−4,...,2} are swept and the best λ within 0.01 ≤ PPL ≤ 1.5 by argmax Δ is reported,' and the evaluation is performed on the 100 held-out prompts. Selecting λ on the same evaluation data biases every Table 1 cell toward the largest Δ a method can achieve on that particular test set and invalidates the claim that CircuitSteer is 'the only method' with a valid operating point. A proper protocol would fix λ on a separate validation set, or report the full coefficient sweep for all eight cells as Figure 3 does for RTP and Sycophancy.
  3. [Section 4.1, Table 1] The validity window [0.01, 1.5] is too permissive to support the fluency-preservation claim. The paper itself describes normalized perplexity values 'near zero' as reflecting repetitive or collapsed text, yet Table 1 reports many operating points with PPL between 0.1 and 0.4 (e.g., RepE Llama-Emotion 0.37, ITI Llama-Sycophancy 0.356, SAE-SSV Llama-RTP 0.34, CircuitSteer Llama-RTP 0.51). Such values indicate heavily degenerate or repetitive output under the paper's own definition, so calling all points in this window 'fluency-preserving' is not credible, and the window appears to be set so as to admit the authors' results while excluding baselines.
  4. [Section 3.2, Eq. (3)] The edge indicator assumes that co-activation plus decoder-direction cosine above τ_sim approximates causal cross-layer feature flow. The paper explicitly labels these as 'candidate indicators' and states that causality is inferred from downstream ablations rather than from individual edges. However, the abstract and Section 6 present the discovered circuit as 'the subcircuit responsible for a target behavior' and the intervention as targeted at the mechanism. Without per-edge causal validation (e.g., activation patching), the dense vectors of Eq. (6) cannot be claimed to target the behavior's mechanism, and the transparency contribution is not established. The authors should either add edge-level causal validation or substantially soften the mechanistic claims.
minor comments (5)
  1. [Section 3.2, Eq. (6)] The averaging in Eq. (6) does not specify whether the decoder directions d_{l,i} are normalized before averaging; since decoder directions have different norms, this affects the scale of v_l and therefore the interpretation of the shared coefficient λ.
  2. [Figure 1] The axis label 'Mean Normalised Perplexity (ppl↓ better)' combined with the annotation 'higher steering' is confusing; clarify which direction on each axis is better and what the arrow denotes.
  3. [Appendix F, Table 11] Table 11 reports behavioral reduction for all methods but reports normalized perplexity only for CircuitSteer, so the claim that CircuitSteer preserves fluency relative to CAA at this scale cannot be verified from the table.
  4. [Appendix C] The main text says 'One-sided 95% bootstrap lower bounds' but Table 7 describes 'per-prompt resampling'; make the resampling procedure explicit so the reader can assess the test.
  5. [Section 4.1] The description of the validity-window selection says it was chosen on 'data disjoint from our evaluation set' and later 'by human inspection of the same outputs'; the relation between these two statements is unclear and should be clarified.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the steering pipeline is empirically evaluated on held-out prompts; the only blemishes are a minor non-load-bearing self-citation and lambda selection on the evaluation set, which is selection pressure rather than circular reasoning.

full rationale

The derivation chain is not circular. Circuit edge selection (Eq. 3) requires co-activation and decoder-direction cosine, specificity scoring (Eq. 4) is contrastive, and the steering vector (Eq. 6) is an average of decoder directions; none of these definitions presuppose the outcome metrics Delta or PPL. The paper explicitly separates circuit construction (80% of examples) from evaluation (100 held-out prompts), so the reported behavioral reduction is not the fitted quantity. The main 'only method' claim is contingent on the empirically chosen validity window and on lambda being selected by argmax Delta within that window on the evaluation set; this is test-set selection pressure, not a definitional reduction, and the window is applied identically to all methods. There is one minor self-citation (Aryashad et al., 2025, cited in Section 5.1 for task-level evaluation) that is not load-bearing, and several other coauthor citations in Related Work are contextual. The paper's own Section 3.2 candidly states that the edge indicators are 'candidate indicators' and that causality is inferred from ablations, which is a limitation but not a circularity. A separate correctness concern, not a circularity, is that Section 5.3's conclusion 'All three ingredients prove necessary' is contradicted by Table 2, where SINGLELAYER (Delta=0.193, PPL=1.04) and NOGEO (Delta=0.088, PPL=0.97) beat the full method on Llama-Sycophancy and Gemma-RTP, respectively; this undermines the necessity claim but does not make the derivation self-referential.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central pipeline rests on hand-set thresholds, a test-set-selected coefficient, an empirical validity window, and the core causal approximation. None of these are derived; they are engineering choices. The invented-entity count is zero because the circuit is a graph over existing SAE features.

free parameters (6)
  • tau_act = 1.5
    Activation threshold in Eq. (3) for co-activation. Set by hand; paper says fixed a priori and Appendix H shows low sensitivity.
  • tau_sim = 0.10
    Cosine threshold in Eq. (2) for decoder-direction alignment. Set by hand; low value keeps many candidate edges.
  • tau_diff = 0.05
    Contrastive specificity threshold in Eq. (4). Set by hand; controls circuit size.
  • top-K = 30 edges (50 for refusal)
    Number of retained edges per layer pair. Set by hand; Table 3 shows saturation only after K=30 but PPL keeps rising.
  • lambda = best per method-dataset in {-4,...,2}
    Steering coefficient chosen by argmax Delta on held-out evaluation within PPL window. This is fitted to the evaluation set and drives Table 1.
  • validity window = [0.01, 1.5] normalized PPL
    Window selected empirically using an LLM fluency judge and human inspection. Determines which operating points count as valid in Table 1.
assumptions (4)
  • domain assumption SAE decomposition is monosemantic and approximately linear in the residual stream.
    Section 3.1; if SAE features do not correspond to interpretable directions, feature-level circuits are not meaningful.
  • ad hoc to paper Co-activation and decoder-direction cosine above threshold approximate causal cross-layer feature flow.
    Section 3.2 and Eq. (3); the paper itself labels these candidate indicators and relies on downstream ablations for causal evidence.
  • domain assumption Off-the-shelf Gemma-Scope, Llama-Scope, and Qwen-Scope SAEs are faithful at the selected layers.
    Section 4.1; discovery depends on these artifacts.
  • ad hoc to paper The [0.01,1.5] normalized perplexity window plus Delta>=0.02 defines a meaningful steering operating point.
    Section 4.1; chosen empirically, not derived, and used to judge all methods.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CircuitSteer: Geometrically Aligned Multi-Layer Steering via Sparse Autoencoder Circuits." pith.science (2026). https://pith.science/paper/AMR5IFCE

@misc{pith2026260805732,
  author       = {Pith},
  title        = {Pith review of: CircuitSteer: Geometrically Aligned Multi-Layer Steering via Sparse Autoencoder Circuits},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AMR5IFCE}},
  note         = {Machine review of arXiv:2608.05732}
}
read the original abstract

Controlling the behavior of large language models (LLMs) remains a critical challenge for AI alignment. Existing steering methods, such as Contrastive Activation Addition (CAA), typically rely on fixed single-layer interventions derived from aggregate activation differences. These methods impose a single intervention across semantically diverse inputs and often fail to sustain consistent behavioral changes across layers, limiting the effectiveness of the steering. In this work, we introduce CircuitSteer, a novel framework that leverages Sparse Autoencoders (SAEs) to identify and manipulate coherent semantic circuits distributed across multiple layers. By constructing a feature flow circuit based on feature co-activation and the geometric alignment of decoder directions, we isolate the specific multi-layer subcircuits responsible for a target behavior. We then synthesize dense steering vectors from these sparse features and apply multi-point interventions to guide the model's internal semantic trajectory. We evaluate CircuitSteer using contrastive examples across a diverse set of tasks, including toxicity, emotion-intensity, sycophancy, and refusal, spanning two model families. Across all models and datasets, CircuitSteer is the only method to consistently produce fluency-preserving interventions; competing methods either sacrifice text quality or lack coverage, failing entirely on complex behaviors like sycophancy and refusal. These results demonstrate that multi-layer circuit steering, enabled by enforcing geometric alignment among selected features, yields strictly more robust and effective behavioral control than static single-point interventions. Code is available at https://github.com/mehrshad-sdtn/CircuitSteer.

Figures

Figures reproduced from arXiv: 2608.05732 by the authors.

Figure 1
Figure 1. CircuitSteer achieves the best trade-off between behavioral reduction and fluency preservation. Each marker shows the mean behavioral reduction (∆ ↑) and normalized perplexity (PPL ↓) over 8 conditions. CircuitSteer shows highest ∆ at PPL < 1.0, outperforming other baseline methods. Aligning LLMs requires balancing behavioral control, stability, and computational efficiency. Methods such as Reinforcement Learning fr… view at source ↗
Figure 2
Figure 2. Overview of CircuitSteer this with an overcomplete dictionary: the encoder produces a sparse activation vector z ∈ RN (N ≫ dmodel) via z = ReLU(Wenc hl + benc), and the decoder reconstructs the residual stream as hˆ l = Wdec z + bdec. Each nonzero entry zi corresponds to a decoder direction Wdec[:, i] ∈ Rdmodel encoding a single interpretable concept (Cunningham et al., 2023). Feature Flow Circuits. The Transformer … view at source ↗
Figure 3
Figure 3. Coefficient sweep on RTP and Sycophancy. Behavioral reduction (∆ ↑, top) and normalized perplexity (PPL ↓, bottom) vs. λ for GEMMA (top row) and LLAMA (bottom row). Dashed red line marks ppl=1.5; values ≥3.0 clipped. CIRCUITSTEER matches or exceeds baselines in ∆ on RTP while maintaining stable ppl, and is the only method to achieve meaningful reduction on Sycophancy for GEMMA. During inference, we intervene simulta… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Judge Based Fluency-Toxicity Trade off Across the Evaluated Coefficient Range. [PITH_FULL_IMAGE:figures/full_fig_p021_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 13 canonical work pages

  1. [1]

    TimingLLM: A Two-Stage Retrieval-Augmented Framework for Pre-Synthesis Timing Prediction from Verilog

    Armin Abdollahi, Negin Ashrafi, Mehdi Kamal, and Massoud Pedram. Timingllm: A two- stage retrieval-augmented framework for pre-synthesis timing prediction from verilog. arXiv preprint arXiv:2604.23602, 2026a. Armin Abdollahi, Mehdi Kamal, and Massoud Pedram. Unified llm model for power, performance, and area prediction from hardware code. Proceedings of M...

    work page Pith review arXiv

  2. [3]

    From filters to vlms: Benchmarking defogging methods through object detection and segmentation performance

    Ardalan Aryashad, Parsa Razmara, Amin Mahjoub, Seyedarmin Azizi, Mahdi Salmani, and Arad Firouzkouhi. From filters to vlms: Benchmarking defogging methods through object detection and segmentation performance. arXiv preprint arXiv:2510.03906,

    arXiv

  3. [5]

    Daniel Borkan, Lucas Dixon, Jeffrey Sorensen, Nithum Thain, and Lucy Vasserman

    URL https://github.com/jbloomAus/SAELens. Daniel Borkan, Lucas Dixon, Jeffrey Sorensen, Nithum Thain, and Lucy Vasserman. Nu- anced metrics for measuring unintended bias with real data for text classification. In Companion Proceedings of The 2019 World Wide Web Conference, pp. 491–500,

    work page 2019

  4. [9]

    11 Preprint

    https://transformer- circuits.pub/2021/framework/index.html. 11 Preprint. Under review. Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, Roger Grosse, Sam McCandlish, Jared Kaplan, Dario Amodei, Martin Wattenberg, and Christopher Olah. Toy models of s...

    work page 2021

  5. [11]

    Yeganeh Gharedaghi, Gene Cheung, and Xianming Liu

    URLhttps://www.kaggle.com/m/3301. Yeganeh Gharedaghi, Gene Cheung, and Xianming Liu. Retinex-based image denois- ing/contrast enhancement using gradient graph laplacian regularizer. In 2023 IEEE International Conference on Image Processing (ICIP), pp. 2710–2714. IEEE,

    work page 2023

  6. [13]

    Can MLLMs Critique Like Humans? Evaluating Open-Ended Aesthetic Reasoning in Multimodal Large Language Models

    Sajjad Ghiasvand, Maryam Amirizaniani, Haniyeh Ehsani Oskouie, Mahnoosh Alizadeh, and Ramtin Pedarsani. Can mllms critique like humans? evaluating open-ended aesthetic reasoning in multimodal large language models. arXiv preprint arXiv:2606.29689, 2026a. Sajjad Ghiasvand, Haniyeh Ehsani Oskouie, Mahnoosh Alizadeh, and Ramtin Pedarsani. MMLoP: Multi-modal ...

    work page Pith review arXiv 2026

  7. [15]

    Under review

    12 Preprint. Under review. Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. arXiv preprint arXiv:2212.04089,

    arXiv

  8. [17]

    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. arXiv preprint arXiv:2403.19647,

    arXiv

Show all 30 references
  1. [18]

    https://distill.pub/2020/circuits/zoom-in

    doi: 10.23915/distill.00024.001. https://distill.pub/2020/circuits/zoom-in. Catherine Olsson, Nelson Elhage, Neel Nanda, Nicholas Joseph, Nova DasSarma, Tom Henighan, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, Dawn Drain, Deep Ganguli, Zac Hatfield-Dodds, Dan...

    2020 doi

  2. [19]

    Nina Panickssery, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexan- der Matt Turner

    https://transformer-circuits.pub/2022/in- context-learning-and-induction-heads/index.html. Nina Panickssery, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexan- der Matt Turner. Steering llama 2 via contrastive activation addition. arXiv preprint arXiv:2312.06681,

    2022 arXiv

  3. [20]

    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,

    arXiv

  4. [21]

    Carer: Contextualized affect representations for emotion recognition

    Elvis Saravia, Hsien-Chi Toby Liu, Yen-Hao Huang, Junlin Wu, and Yi-Shin Chen. Carer: Contextualized affect representations for emotion recognition. In Proceedings of the 2018 conference on empirical methods in natural language processing, pp. 3687–3697,

    2018

  5. [23]

    13 Preprint

    URLhttps://arxiv.org/abs/2407.12404. 13 Preprint. Under review. Yasaman Torabi, Parsa Razmara, Hamed Ajorlou, and Bardia Baraeinejad. Neuromamballm: Dynamic graph learning of fmri functional connectivity in autistic brains using mamba and language model reasoning. arXiv prepri...

    arXiv

  6. [24]

    Activation addition: Steering language models without optimization

    Alex Turner, Lisa Thiergart, David Udell, Gavin Leech, David Mini, and Monte MacDiarmid. Activation addition: Steering language models without optimization. arXiv preprint arXiv:2308.10248,

    arXiv

  7. [25]

    Improving llm reasoning through interpretable role-playing steering

    Anyi Wang, Dong Shu, Yifan Wang, Yunpu Ma, and Mengnan Du. Improving llm reasoning through interpretable role-playing steering. arXiv preprint arXiv:2506.07335,

  8. [26]

    Exploring representations and interventions in time series foundation models

    Michał Wili ´nski, Mononito Goswami, Willa Potosnak, Nina ˙Zukowska, and Artur Dubrawski. Exploring representations and interventions in time series foundation models. arXiv preprint arXiv:2409.12915,

    arXiv

  9. [27]

    Small graph is all you need: Deepstategnn for scalable traffic forecasting

    Yannick W¨olker, Arash Hajisafi, Cyrus Shahabi, and Matthias Renz. Small graph is all you need: Deepstategnn for scalable traffic forecasting. arXiv preprint arXiv:2502.14525,

    arXiv

  10. [28]

    Axbench: Steering llms? even simple baselines 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 baselines outperform sparse autoencoders. arXiv preprint arXiv:2501.17148,

    arXiv

  11. [29]

    idiots” what “our

    Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, et al. Representation engineering: A top-down approach to ai transparency. arXiv preprint arXiv:2310.01405, 2023a. Andy Zou, Zifan Wang...

    arXiv

  12. [30]

    The circuit discovery hyperparameters are maintained: sparsity threshold τact = 1.5, contrastive specificity threshold τdiff = 0.05, geometric alignment thresholdτ sim =0.1, and the top-kedges are truncated atk=50. E.2 Results Table 10 summarizes the refusal reduction and flue...

  13. [2016]

    Morfi: Mutimodal zero-shot reasoning for financial time-series inference

    Tina Khezresmaeilzadeh, Parsa Razmara, Mohammad Erfan Sadeghi, Seyedarmin Azizi, and Erfan Baghaei Potraghloo. Morfi: Mutimodal zero-shot reasoning for financial time-series inference. In 2025 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pp. 4295–4304. IEEE,

    2025

  14. [2017]

    Training verifiers to solve math word problems

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168,

    arXiv

  15. [2018]

    Towards understanding sycophancy in language models

    Mrinank Sharma, Meg Tong, Tomasz Korbak, David Duvenaud, Arthur Conmy, Esin Durmus, Lauro Larson, et al. Towards understanding sycophancy in language models. arXiv preprint arXiv:2310.13548,

    arXiv

  16. [2020]

    Zeqing He, Zhibo Wang, Huiyu Xu, Hejun Lin, Wenhui Zhang, and Zhixuan Chu

    URL https://github.com/unitaryai/ detoxify. Zeqing He, Zhibo Wang, Huiyu Xu, Hejun Lin, Wenhui Zhang, and Zhixuan Chu. Interpretable llm guardrails via sparse representation steering. arXiv preprint arXiv:2503.16851, 2025a. Zirui He, Mingyu Jin, Bo Shen, Ali Payani, Yongfeng Z...

    2025

  17. [2021]

    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,

    arXiv

  18. [2022]

    Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A

    https://transformer- circuits.pub/2022/toy model/index.html. Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A. Smith. RealTox- icityPrompts: Evaluating neural toxic degeneration in language models. In Findings of the Association for Computational Linguisti...

    2022

  19. [2023]

    Decentralized low-rank fine-tuning of large language models

    Sajjad Ghiasvand, Mahnoosh Alizadeh, and Ramtin Pedarsani. Decentralized low-rank fine-tuning of large language models. In Proceedings of the 1st Workshop for Research on Agent Language Models (REALM 2025), pp. 334–345,

    2025

  20. [2024]

    Dana Arad, Aaron Mueller, and Yonatan Belinkov

    URL https://github.com/meta-llama/llama3/blob/ main/MODEL CARD.md. Dana Arad, Aaron Mueller, and Yonatan Belinkov. Saes are good for steering–if you select the right features. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp. 10252–10270,

    2025

  21. [2025]

    Mechanistic permutability: Match features across layers

    Nikita Balagansky, Ian Maksimov, and Daniil Gavrilov. Mechanistic permutability: Match features across layers. arXiv preprint arXiv:2410.07656,

    arXiv

  22. [2026]

    URL https: //arxiv.org/abs/2605.11887. Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, Nova DasSarma, Dawn Drain, Deep Ganguli, Zac Hatfield-Dodds, Danny Hernandez, Andy Jones, Jackson Ker...

    arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.