Pith. sign in

REVIEW 3 major objections 5 minor 22 references

Activation steering for chain-of-thought faithfulness is transferable: on Gemma-3 12B, effect size is set by the evaluation setting, not by the training cue, dataset, or construction method, and steering reduces hidden cue use without incre

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 14:21 UTC pith:37H36A5T

load-bearing objection Useful, honest empirical map of faithfulness steering, but the headline transfer claim is confounded by per-scenario layer selection and needs a fixed-layer control. the 3 major comments →

arxiv 2607.29062 v1 pith:37H36A5T submitted 2026-07-31 cs.AI

On the Generalization of Steering Vectors for Chain-of-Thought Faithfulness

Matthew Nguyen , Kyle Cox , Austin Meek , Iv\'an Arcuschin This is my paper
classification cs.AI
keywords chain-of-thought faithfulnessactivation steeringsteering vector transfercue acknowledgmenthidden cue useLLM interpretabilitygeneralizationGemma-3 12B
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper asks whether activation steering for chain-of-thought faithfulness—making a model say out loud that it is relying on a prompt cue—works only in the exact setting where the steering vector was built, or whether it transfers. The authors test three models, four cue types, three datasets, and four ways of constructing the steering vector. They find that for the largest model, Gemma-3 12B, a vector trained on one cue or dataset increases cue acknowledgment about as much when applied to a different cue or dataset, and the size of the effect is set by the evaluation setting rather than the training setting. Construction method also barely matters: a synthetic vector built from templates and a cue-agnostic optimized direction perform as well as vectors built from judge-labeled rollouts. Steering does not appear to work by making the cue more salient—cue use stays roughly constant while hidden cue use (using the cue without mentioning it) falls, on MMLU from 0.31 to 0.05.

Core claim

For Gemma-3 12B, the steering effect transfers broadly: in cross-cue and cross-dataset matrices, off-diagonal cells (vector trained on setting A, applied to setting E) are close to the diagonal of the evaluation column E, not the training row A. The four construction methods produce similar Δack on every dataset, with the MMLU effects all between +0.24 and +0.28. On MMLU, steering raises cue acknowledgment from 0.62 to 0.89 while leaving cue use nearly unchanged (0.32→0.35), so hidden cue use drops from 0.31 to 0.05. The authors conclude that the effect is determined primarily by the evaluation setting, and that steering targets verbalization behavior rather than cue salience.

What carries the argument

The central object is the residual-stream steering vector: a unit-normalized direction added to the activations at every generated token at a selected layer, scaled by α=5. The paper compares four constructions—contrastive (difference of mean activations between judge-labeled faithful and unfaithful rollouts), synthetic (templated completions, no rollouts), and two one-shot optimized directions (cue-specific and cue-agnostic). The central metric is Δack, the matched change in acknowledgment rate between steered and unsteered traces, expressed as (converted minus regressed) fraction of paired traces. Transfer is measured by matrices with training settings as rows and evaluation settings as co

Load-bearing premise

The steering layer is chosen separately for each scenario using a probe that overfits its training set, so comparisons across datasets and cues entangle the effect of the layer with the effect of the scenario; if layer choice, not the vector, drives the transfer, the central generalization claim collapses.

What would settle it

For Gemma-3 12B, fix the steering layer to one layer (for example, the layer selected for MMLU) across all train/eval cells and recompute the cross-cue and cross-dataset transfer matrices. If off-diagonal cells no longer track the evaluation column's diagonal—if transfer effect no longer equals the evaluation setting's self-steering effect—the per-scenario layer choice was carrying the result.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A single steering vector trained on any cue or dataset can be applied elsewhere without losing effect, at least on a model where steering is effective.
  • Steering vectors can be built without model rollouts or judge labels: synthetic and cue-agnostic optimized vectors match contrastive vectors, so faithfulness interventions do not require expensive annotation.
  • The effect of steering on a new dataset can be approximated from that dataset's self-steering effect, potentially before training a vector.
  • Steering does not increase reliance on the cue; it reduces unacknowledged reliance, so it is a verbalization intervention rather than a salience nudge.
  • On Gemma-3 12B/MMLU, hidden cue use falls from 0.31 to 0.05, meaning the model nearly stops using cues it does not acknowledge.

Where Pith is reading between the lines

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

  • If transfer is truly set by the evaluation setting, a small set of general faithfulness directions per model could replace per-cue and per-dataset vectors; the paper's per-scenario layer selection should be controlled with a fixed-layer experiment before relying on this.
  • The cue-agnostic optimized vector performing as well as cue-specific ones suggests the steering direction may correspond to a general 'verbalize your reasoning factors' feature; probing that shared direction across layers could locate a common circuit.
  • The sharp contrast between MMLU and GPQA/BBH effects on Gemma-3 12B suggests task difficulty or dataset trace length mediates steerability; sweeping more datasets and model scales would map where steering becomes usable.
  • Because all labels come from a single LLM judge, multi-judge or human validation of cue acknowledgment would be a direct robustness check on the transfer numbers.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This paper studies whether activation steering for chain-of-thought faithfulness generalizes across cue types, datasets, and vector-construction methods. Using Gemma-3 4B/12B and Qwen-3.5 9B, it injects four cues into GPQA, BBH, and MMLU questions, constructs contrastive, synthetic, cue-specific optimized, and cue-agnostic optimized steering vectors, and measures change in cue acknowledgment (Δack) and hidden cue use under steering. The headline finding is that for Gemma-3 12B, vectors trained on one cue or dataset transfer to others: off-diagonal transfer cells in Figures 2-3 are roughly constant within each evaluation column, and transfer magnitude correlates with the test setting's self-steering effect (Appendix B), suggesting effect size is set by the evaluation setting rather than the training setting. It also reports that steering reduces hidden cue use without increasing cue use, and that construction method has little influence. The paper explicitly lists limitations: per-scenario layer selection confounds layer with scenario, probe train AUROC is 1.00 everywhere, layer is selected on the test set, and all labels come from a single unvalidated judge.

Significance. The question is timely and the study is carefully designed in many respects: paired baseline/steered rollouts, matched within-setting comparisons, multiple construction methods including a cue-agnostic one, a split-half control for shared evaluation noise, and a released codebase. The paper is unusually candid about its own confounds. However, the central transfer claim—that effect size is determined primarily by the evaluation setting—is not yet established because the per-evaluation-scenario layer selection is entangled with the evaluation setting, and because all outcome labels come from a single unvalidated judge. If the transfer claim were confirmed with a fixed-layer control and judge validation, it would be a valuable result for practical steering: vectors would be interchangeable across cues/datasets and the main determinant would be the target setting. As it stands, the paper is a useful empirical contribution with an unresolved confound at its core.

major comments (3)
  1. [§3.3, §4, Figs. 2–3] The per-scenario steering layer confounds the central transfer claim. In Figs. 2–3 and Fig. 5, every entry in an evaluation column is obtained by adding the training vector at the layer selected for that evaluation scenario (argmax probe test AUROC; Table 5, Appendix D.4). Thus the column's shared layer/coefficient, rather than the vector's origin, could produce the observed column constancy. Section 4 explicitly acknowledges this ('the layers are confounded with the scenario parameters (dataset, cue)'), and Appendix B's split-half control only removes shared baseline noise; it does not disentangle layer from evaluation setting. The abstract's claim that effect size is determined primarily by the evaluation setting is therefore not supported as stated. Required: a control that fixes one steering layer across all train/eval combinations (or varies train and eval layers orthogonally) and r
  2. [§2 'Judge and metrics', Appendix F] All Δack values and the 'hidden cue use' reductions rest on labels from a single LLM judge (gpt-5-nano) using cue-specific acknowledgment criteria, with no validation against a second judge or human annotation. If the judge systematically over- or under-acknowledges under steering (e.g., due to response length or the cue-specific criterion), all transfer and faithfulness conclusions inherit that bias. Section 4 lists this as a limitation but it remains unresolved. Please validate on a sample (e.g., 100–200 traces per scenario) with a second judge/human and report agreement, or argue why judge bias cannot affect the relative comparisons.
  3. [§3.1, §4, Table 5, Fig. 7] The layer-selection procedure is too weak to support the layer-specific readout. Probe train AUROC is 1.00 everywhere because the activation dimension exceeds the training set (Appendix E); the reported test AUROC is the maximum over layers selected on the same test set (Section 3.1); and Figure 7 shows no relationship between probe AUROC and Δack. The selected layers vary widely (e.g., L3–L41 in Tables 5 and 8), making the per-scenario layer effectively another free parameter. A robustness check that fixes a common layer across all scenarios is needed to show that the headline +0.26 MMLU effect and the column constancy are not artifacts of choosing an unusually effective layer for MMLU.
minor comments (5)
  1. [§3.4] “not a prioriclear” should read “not a priori clear”.
  2. [Tables 5–8] The abbreviation “uni.” (e.g., “all-cues (uni.),” “all-data (uni.)”) is not defined; please spell out “unified” or “all-cue”.
  3. [Appendix C] The sentence “The values α∈{1,7,10} that appear in the raw data are single-cell pilot runs and are omitted” should be moved or flagged earlier so readers do not wonder about missing sweep points.
  4. [Appendix D.5] “Because the residual stream is additive, directions at different depths live in a shared space and are comparable” needs a brief justification: additivity alone does not make norms or scales comparable across layers.
  5. [Fig. 5 caption] “Red placements are uninformative” should be qualified as “uninformative with respect to the test setting” to avoid implying the train setting has no structure.

Circularity Check

0 steps flagged

No circularity: the transfer results are empirical measurements on held-out rollouts; the acknowledged per-scenario layer confound is a validity limitation, not a circular reduction.

full rationale

This is an empirical generalization study rather than a derivation chain, and I find no step where a reported quantity is equivalent to a fitted input by construction. Steering vectors are constructed on train splits and evaluated on held-out test items; Δack is computed from paired fresh rollouts on external benchmarks (BBH, GPQA, MMLU). The central claim—that transfer Δack is determined mainly by the evaluation setting—is supported by transfer matrices and a split-half control for shared evaluation noise (Appendix B, Table 3); the test-setting diagonal is not used to construct the transfer vectors, so the correlation between them is an empirical regularity rather than a definitional identity. The self-citations (Meek et al. 2025 for cues and the cue-acknowledgment operationalization; Arcuschin et al. for unfaithful CoT; Venhoff et al. for steering-related features) are context and measurement choices, not the target result. The cue-acknowledgment operationalization is explicitly acknowledged as a proxy for faithfulness (Section 2: 'cue acknowledgment does not establish faithfulness in the conventional sense'), and the present paper's transfer conclusions are measured on fresh rollouts rather than imported from those citations. The strongest limitation is the paper's own acknowledged per-scenario steering-layer selection confound (Section 4: 'because we select steering layers per-scenario, the layers are confounded with the scenario parameters (dataset, cue). So, comparisons between, e.g., GPQA and MMLU, are not like-for-like; the layer effect and dataset effect are entangled.'). This means cross-column comparisons in Figures 2–3 may reflect the eval layer as much as the eval dataset/cue. However, the within-column row comparisons still vary only the train setting while holding the eval layer fixed, so the finding that train setting has little effect is not circular. The confound limits how the 'evaluation setting' effect should be decomposed, but it does not reduce the transfer claim to an input of the analysis by construction. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work, and no ansatz is smuggled in via self-citation. Score 0.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

Central claims are empirical and rest on measurement assumptions: acknowledgment-as-faithfulness, cued-option-as-use, one judge, and test-set-based layer selection. No invented entities or novel theoretical constructs are introduced.

free parameters (2)
  • Steering coefficient α = 5 (swept 2.5, 5, 7.5)
    Scaling of the unit-normalized steering direction; chosen by hand as the default. Effects are stable over the narrow swept range, but no wide-range or saturation curve is provided.
  • Per-scenario steering layer = Varies by scenario (e.g., Gemma-3 12B: L15 GPQA, L3 BBH, L10 MMLU; Table 8)
    Layer selected as argmax probe test AUROC. Selection is on the test set, train AUROC is 1.00 everywhere, and the layer is confounded with dataset/cue. All steering and transfer results depend on this choice.
axioms (3)
  • domain assumption Cue acknowledgment is a valid operationalization of faithfulness
    The paper states acknowledgment does not establish faithfulness in the conventional sense (Jacovi & Goldberg, 2020). All central metrics measure acknowledgment, not full faithfulness.
  • domain assumption Selecting the cued option equals cue use
    Supported by the errors-cued concentration (76-87%) but acknowledged to fail in two ways (selecting cued option without reliance; influence without selection). Hidden-cue-use inherits this assumption.
  • domain assumption Single LLM judge gpt-5-nano provides reliable labels
    All acknowledgment, degeneracy, and answer labels come from one unvalidated judge; judge error is not bounded. This affects every reported rate and difference.

pith-pipeline@v1.3.0-daily-deepseek · 19412 in / 12449 out tokens · 133241 ms · 2026-08-03T14:21:39.682912+00:00 · methodology

0 comments
read the original abstract

Model capabilities have improved in large part due to scaling chain of thought. This has been a promising development for AI safety--where models verbalize their reasoning, it is possible to monitor it. However, in some cases, models do not verbalize important steps in their reasoning process. For example, models prompted with a cue suggesting the incorrect answer may fail to acknowledge that cue, even when it appears instrumental to their conclusion. When chain of thought (CoT) fails to disclose instrumental reasoning steps, we describe it as unfaithful. Prior work has shown that activation steering can be a useful method to improve faithfulness in CoT. We extend this line of work by studying how well steering for faithfulness generalizes across cue types, datasets, and methods of constructing the steering vector for three models (Gemma-3 4B, Qwen-3.5 9B, Gemma-3 12B) in a cued question-answering setting. While steering reliably increases cue acknowledgment for only the largest model (Gemma-3 12B), we find that when steering is effective, its effect generalizes broadly across cue types and datasets--in cross-cue and cross-dataset analyses, effect size is determined primarily by the evaluation setting, rather than the vector's train setting. How the vector is built also matters little--four construction methods, including one whose optimization target mentions no specific cue, yield similar effect sizes. Finally, we consider the possibility that steering promotes the salience of the cue and causes greater cue use, rather than targeting verbalization behaviors. However, we find no evidence for this--steering leaves the rate of cue use roughly unchanged while reducing hidden cue use, i.e., cue use that is not acknowledged.

Figures

Figures reproduced from arXiv: 2607.29062 by Austin Meek, Iv\'an Arcuschin, Kyle Cox, Matthew Nguyen.

Figure 1
Figure 1. Figure 1: Steering efficacy depends on the model and the evaluation dataset, and not clearly [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Steering transfer across cue types. ∆ack on GPQA by training cue (rows) and evaluation cue (columns) using contrastive vectors with α=5. Appendix B attempts to quantify (1): off-diagonal effects correlate with the test setting’s own (diagonal) effect but not the train setting’s, and the relationship survives a split-half control for shared evaluation noise. Conclusion (2) inspires the brief investigation i… view at source ↗
Figure 3
Figure 3. Figure 3: Steering transfer across datasets. ∆ack under the Stanford cue by training dataset (rows) and evaluation dataset (columns) using contrastive vectors with α=5. 3.4 Cue acknowledgment and cue use While our work nominally studies faithfulness, it is not a priori clear that increases in acknowledgment indicate increases in faithfulness. Because the evaluation setting is not conditioned on the model following t… view at source ↗
Figure 4
Figure 4. Figure 4: ∆ack on GPQA by cue, with the pooled estimate at right (α=5; 90% normal￾approximation CIs for the paired change; the pooled n is larger, hence tighter intervals). Companion to [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: restates the column-constancy of the transfer matrices (Figures 2 and 3) as a single relationship. Each off-diagonal cell (a vector trained on setting A, applied to setting E) is plotted twice at the same height, its transfer ∆ack: once at the self-steering ∆ack of the train setting (A on A; red) and once at that of the test setting (E on E; blue), with a connector joining the two placements. Blue placemen… view at source ↗
Figure 6
Figure 6. Figure 6: Robustness to the steering coefficient. Matched [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Probe test AUROC (selected layer) versus [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Cross-cue alignment of synthetic difference-of-means vectors, reconstructed at a [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Selected linear-probe test AUROC for cue acknowledgment, by dataset and cue [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

22 extracted references · 15 linked inside Pith

  1. [1]

    Biases in the blind spot: Detecting what LLMs fail to mention

    Iv \'a n Arcuschin, David Chanin, Adri \`a Garriga-Alonso, and Oana-Maria Camburu. Biases in the blind spot: Detecting what LLMs fail to mention. In 43rd International Conference on Machine Learning (ICML), 2026. URL https://arxiv.org/abs/2602.10117

    Pith/arXiv arXiv 2026

  2. [2]

    Chain-of- Thought Reasoning In The Wild Is Not Always Faithful , March 2025

    Iván Arcuschin, Jett Janiak, Robert Krzyzanowski, Senthooran Rajamanoharan, Neel Nanda, and Arthur Conmy. Chain-of- Thought Reasoning In The Wild Is Not Always Faithful , March 2025. URL http://arxiv.org/abs/2503.08679. arXiv:2503.08679 [cs]

    Pith/arXiv arXiv 2025

  3. [3]

    Guan, Aleksander Madry, Wojciech Zaremba, Jakub Pachocki, and David Farhi

    Bowen Baker, Joost Huizinga, Leo Gao, Zehao Dou, Melody Y. Guan, Aleksander Madry, Wojciech Zaremba, Jakub Pachocki, and David Farhi. Monitoring reasoning models for misbehavior and the risks of promoting obfuscation, 2025. URL https://arxiv.org/abs/2503.11926

    Pith/arXiv arXiv 2025

  4. [4]

    Reasoning theater: Disentangling model beliefs from chain-of-thought, 2026

    Siddharth Boppana, Annabel Ma, Max Loeffler, Raphael Sarfati, Eric Bigelow, Atticus Geiger, Owen Lewis, and Jack Merullo. Reasoning theater: Disentangling model beliefs from chain-of-thought, 2026. URL https://arxiv.org/abs/2603.05488

    Pith/arXiv arXiv 2026

  5. [5]

    Bowman, Jan Leike, Jared Kaplan, and Ethan Perez

    Yanda Chen, Joe Benton, Ansh Radhakrishnan, Jonathan Uesato, Carson Denison, John Schulman, Arushi Somani, Peter Hase, Misha Wagner, Fabien Roger, Vlad Mikulik, Samuel R. Bowman, Jan Leike, Jared Kaplan, and Ethan Perez. Reasoning Models Don 't Always Say What They Think , May 2025. URL http://arxiv.org/abs/2505.05410. arXiv:2505.05410 [cs.CL]

    Pith/arXiv arXiv 2025

  6. [6]

    Output supervision can obfuscate the chain of thought, 2025

    Jacob Drori, Luke Marks, Bryce Woodworth, Alex Cloud, and Alexander Matt Turner. Output supervision can obfuscate the chain of thought, 2025. URL https://arxiv.org/abs/2511.11584

    arXiv 2025

  7. [7]

    One-shot Optimized Steering Vectors Mediate Safety -relevant Behaviors in LLMs , August 2025

    Jacob Dunefsky and Arman Cohan. One-shot Optimized Steering Vectors Mediate Safety -relevant Behaviors in LLMs , August 2025. URL http://arxiv.org/abs/2502.18862. arXiv:2502.18862 [cs]

    Pith/arXiv arXiv 2025

  8. [8]

    Measuring Massive Multitask Language Understanding , January 2021

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring Massive Multitask Language Understanding , January 2021. URL http://arxiv.org/abs/2009.03300. arXiv:2009.03300 [cs]

    Pith/arXiv arXiv 2021

  9. [9]

    Alon Jacovi and Yoav Goldberg. Towards faithfully interpretable NLP systems: How should we define and evaluate faithfulness? In Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel Tetreault (eds.), Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp.\ 4198--4205, Online, July 2020. Association for Computational Lin...

    doi:10.18653/v1/2020.acl-main.386 2020

  10. [10]

    Chain of Thought Monitorability : A New and Fragile Opportunity for AI Safety , July 2025

    Tomek Korbak, Mikita Balesni, Elizabeth Barnes, Yoshua Bengio, Joe Benton, Joseph Bloom, Mark Chen, Alan Cooney, Allan Dafoe, Anca Dragan, Scott Emmons, Owain Evans, David Farhi, Ryan Greenblatt, Dan Hendrycks, Marius Hobbhahn, Evan Hubinger, Geoffrey Irving, Erik Jenner, Daniel Kokotajlo, Victoria Krakovna, Shane Legg, David Lindner, David Luan, Aleksand...

    Pith/arXiv arXiv 2025

  11. [11]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention, 2023. URL https://arxiv.org/abs/2309.06180

    Pith/arXiv arXiv 2023

  12. [12]

    Brockmeier, and Steven Basart

    Austin Meek, Eitan Sprejer, Iván Arcuschin, Austin J. Brockmeier, and Steven Basart. Measuring Chain -of- Thought Monitorability Through Faithfulness and Verbosity , 2025. arXiv:2510.27378 [cs]

    arXiv 2025

  13. [13]

    Probing and Steering Chain -of- Thought Unfaithfulness in Language Models

    Giovanni Maria Occhipinti, Alessandro Abate, and Nandi Schoots. Probing and Steering Chain -of- Thought Unfaithfulness in Language Models . March 2026. URL https://openreview.net/forum?id=JL8sNbnSWK

    2026

  14. [14]

    Openai gpt-5 system card, 2026

    OpenAI . Openai gpt-5 system card, 2026. URL https://arxiv.org/abs/2601.03267

    Pith/arXiv arXiv 2026

  15. [15]

    David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R. Bowman. GPQA : A Graduate - Level Google - Proof Q & A Benchmark , November 2023. URL http://arxiv.org/abs/2311.12022. arXiv:2311.12022 [cs]

    Pith/arXiv arXiv 2023

  16. [16]

    Steering llama 2 via contrastive activation addition

    Nina Rimsky, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Turner. Steering llama 2 via contrastive activation addition. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar (eds.), Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.\ 15504--15522, Bangkok, Thailand, Augus...

    doi:10.18653/v1/2024.acl-long.828 2024

  17. [17]

    Le, Ed H

    Mirac Suzgun, Nathan Scales, Nathanael Schärli, Sebastian Gehrmann, Yi Tay, Hyung Won Chung, Aakanksha Chowdhery, Quoc V. Le, Ed H. Chi, Denny Zhou, and Jason Wei. Challenging BIG - Bench Tasks and Whether Chain -of- Thought Can Solve Them , October 2022. URL http://arxiv.org/abs/2210.09261. arXiv:2210.09261 [cs]

    Pith/arXiv arXiv 2022

  18. [18]

    Gemma Team. Gemma 3. 2025. URL https://arxiv.org/abs/2503.19786

    Pith/arXiv arXiv 2025

  19. [19]

    Qwen3.5: Accelerating productivity with native multimodal agents, February 2026

    Qwen Team. Qwen3.5: Accelerating productivity with native multimodal agents, February 2026. URL https://qwen.ai/blog?id=qwen3.5

    2026

  20. [20]

    Vazquez, Ulisse Mini, and Monte MacDiarmid

    Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J. Vazquez, Ulisse Mini, and Monte MacDiarmid. Steering language models with activation engineering, 2024. URL https://arxiv.org/abs/2308.10248

    Pith/arXiv arXiv 2024

  21. [21]

    Base Models Know How to Reason , Thinking Models Learn When , October 2025 a

    Constantin Venhoff, Iván Arcuschin, Philip Torr, Arthur Conmy, and Neel Nanda. Base Models Know How to Reason , Thinking Models Learn When , October 2025 a . URL http://arxiv.org/abs/2510.07364. arXiv:2510.07364 [cs]

    Pith/arXiv arXiv 2025

  22. [22]

    Understanding Reasoning in Thinking Language Models via Steering Vectors

    Constantin Venhoff, Iván Arcuschin, Philip Torr, Arthur Conmy, and Neel Nanda. Understanding Reasoning in Thinking Language Models via Steering Vectors . In Workshop on Reasoning and Planning for Large Language Models , 2025 b . URL https://openreview.net/forum?id=OwhVWNOBcz

    2025