Pith. sign in

REVIEW 4 major objections 6 minor 15 references

Feature Extraction and Steering for Enhanced Chain-of-Thought Reasoning in Language Models

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a model's chain-of-thought reasoning can be strengthened at generation time by steering residual activations along a direction computed from the contrast between verbal and symbolic reasoning traces.

desk verdict The VS decomposition idea is genuinely useful, and the SAE-free recipe may work heuristically, but the derivation in Eq. 7 is mathematically wrong and the evaluation leaks the test distribution into feature and hyperparameter selection. read the letter →

arxiv 2505.15634 v4 pith:OJFRJBQU submitted 2025-05-21 cs.CL cs.LG

Zihao Li , Xu Wang , Yuzhe Yang , Ziyu Yao , Haoyi Xiong , Mengnan Du This is my paper · ORCID
classification cs.CLcs.LG MSC 68T5068T07
keywords chain-of-thoughtsparseautoencoderrepresentationsteeringverbal-symbolicdecompositionresidualactivationsmathematicalreasoningeigenvectorLLMinterpretability
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 claims that a language model's mathematical reasoning can be strengthened by steering its internal residual activations along directions extracted from ordinary chain-of-thought traces, without long-CoT data or fine-tuning. It proposes a verbal/symbolic decomposition and uses either sparse-autoencoder features or, when no SAE exists, a direct eigenvector computation from the mean residual difference between the two process types. Across four math benchmarks and four model families, the steering directions consistently improve accuracy over the base model and over the BoostStep and MathNeuro baselines. If correct, the method offers a parameter-free, training-free route to deeper reasoning behavior.

What carries the argument

The central object is the steering direction obtained from the verbal–symbolic contrast. In the SAE variant it is a decoder column of a sparse autoencoder; in the SAE-free variant it is the top left singular vector of the matrix of per-sample mean residual differences, obtained by solving a Rayleigh quotient problem for $AA^\top$. The argument works by assuming SAE features are sparse and approximately mutually orthogonal, so that the objective of maximizing summed absolute projections collapses to maximizing $\|A^\top z\|_2$.

What would settle it

Compute the true unit vector maximizing the sum of absolute projections for the paper's matrix $A$ (e.g., via a dense grid or a global-optimization routine that explicitly handles the piecewise-linear absolute value), and compare its steering accuracy on a held-out set against the top eigenvector of $AA^\top$; if the two directions yield different steering behavior, the paper's derivation is not the mechanism behind its results.

Watch

Extended reading notes

Core claim

At the core is the claim that ordinary chain-of-thought traces already contain the features that support deep reasoning, and that those features can be amplified in a targeted way. The paper's SAE-based method extracts feature activations from the residual stream, separates the trace into a verbal process and a symbolic process, and uses the absolute difference of their SAE features as a noise-suppressed importance score; the top scoring features then provide steering directions. The SAE-free method shows that the same direction can be obtained without any SAE, by steering along the leading eigenvector of the covariance of the mean residual differences between the verbal and symbolic processes. On GSM8K, MATH-L3&L4, MMLU-high, and MATHOAI, the paper reports consistent accuracy improvements for DeepSeek-Llama3-8B, DeepSeek-qwen-1.5B, Llama3.1-8B-it, and Gemma-2-9b-it.

Load-bearing premise

The derivation assumes that maximizing the sum of absolute projections of the verbal–symbolic differences onto a unit direction is equivalent to maximizing the Euclidean norm of the projected matrix, an equivalence that does not hold for general matrices and is not proved in the paper.

Editorial extensions

If this is right

  • Steering along the extracted direction consistently raises accuracy on the four evaluated math benchmarks over both the unsteered model and the BoostStep and MathNeuro baselines.
  • The SAE-free eigenvector method matches or beats the SAE-based method, so models without a pretrained sparse autoencoder can still be steered.
  • Steering increases the length and depth of generated reasoning traces, and the model's attention shifts toward mathematical tokens.
  • There is an optimal steering strength; too large a strength degrades generation quality, so the method requires tuning the strength parameter $\lambda$.

Reading between the lines

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

  • The verbal/symbolic contrast could be reused as a probe for isolating other reasoning faculties, such as planning or logical deduction, without retraining.
  • If the equivalence between the sum-of-absolute-projections objective and the L2 objective fails on some model families, the SAE-free recipe might still work for reasons unrelated to its derivation, which would point to a different, more robust explanation.
  • Because the steering direction comes from a small sampled set (N = 100 questions), the method's variance across different steering samples is not reported; a natural extension is measuring how much accuracy depends on the choice of the 100 questions.
  • The approach may offer a cheap alternative to long-CoT distillation by amplifying reasoning features that are already latent in ordinary CoT.
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 / 6 minor

Summary. The paper proposes two activation-steering methods intended to improve chain-of-thought mathematical reasoning. The SAE-based method decomposes vanilla CoT into a verbal stream and a symbolic stream, uses a sparse autoencoder to extract feature activations, ranks features by the absolute difference between the two streams, and adds a scaled SAE feature direction to residual activations. The SAE-free method replaces the SAE feature direction with top eigenvectors of AA^T, where the columns of A are mean residual-activation differences between verbal and symbolic streams over N=100 MATHOAI samples. Experiments report accuracy improvements on GSM8K, MATH-L3&L4, MMLU-high, and MATHOAI compared with original models and with BoostStep and MathNeuro baselines. The paper also provides case studies, attention analyses, and SAE feature visualizations.

Significance. If the reported improvements were robust and out-of-sample, the work would offer a lightweight alternative to long-CoT distillation, since it requires only vanilla CoT and ordinary activations rather than expensive long-CoT data or fine-tuning. The qualitative analyses, including the steering case studies and the eigenvector-to-feature similarity tables in Appendix J, are suggestive and could be useful for future interpretability work. However, the current evidentiary value is limited by the circular evaluation protocol and by a mathematical gap in the derivation of the SAE-free steering direction. The method might still work as a heuristic, but the manuscript as written does not establish either the theoretical or the empirical claim at the level claimed in the abstract.

major comments (4)
  1. [Section 4.2, Eq. (7) and Appendix E] The reduction from maximizing sum_p |<x_p - y_p, z>| to maximizing ||A^T z||_2 is mathematically invalid. The L1 objective and the L2 objective have different maximizers on the unit sphere. For example, with A = [e1, e2], ||A^T z||_2 = 1 for every unit vector z, while sum_p |<x_p - y_p, z>| = |z1| + |z2| is maximized only on the diagonal directions. The subsequent Rayleigh-quotient argument therefore does not establish that the top eigenvectors of AA^T solve the stated optimization problem. Because this derivation is the theoretical foundation for the steering direction in Eq. (8), it must be corrected, or the SAE-free method must be presented explicitly as a heuristic rather than as a derived solution.
  2. [Sections 2.3, 3.1 and Appendices A, C, D] The evaluation protocol is not out-of-sample. The steering statistics are computed from N=100 MATHOAI samples (Eq. (2) and the SAE-free A matrix), and MATHOAI is then an evaluation set in Tables 1 and 2. For GSM8K, Section 3.1 states that parameters and features are selected on a 300-sample subset, and Table 3 reports only that subset for the steering experiments, not the full set. Appendix C lists hand-picked 'suitable' features, Appendix D restricts the SAE-free method to top-10 eigenvectors, and Figure 4 tunes the steering strength lambda using accuracy on the evaluation sets. These choices are therefore fitted to the test distribution. With n=300 and no reported variance or multiple runs, the +3.0 point gain on GSM8K (82.67 to 85.67) may be within sampling variability, and no steered full-set GSM8K number is provided. The central empirical claim requires held-out evaluation and error bars or repeated-seed results.
  3. [Appendix B and Eq. (2)] The SAE-based feature ranking actually used for llama_scope_r1_distill is not Eq. (2) but the alternative alpha_t = (1/N) |sum_k (alpha^{x_k}_t - alpha^{y_k}_t)|, introduced because the Eq. (2) ranking 'contains many noise features.' This is a post-hoc modification of the core method, and the main text in Section 2.3 does not describe it. The choice is made based on the same evaluation datasets, and it undermines the reproducibility and consistency of the claimed VS-decomposition ranking procedure.
  4. [Section 4.2, Eq. (6)] The error-term cancellation assumes that the per-token reconstruction errors of the verbal and symbolic processes are equal and cancel in subtraction, but this is asserted without empirical support. In addition, Eq. (9) in Appendix E uses the sparsity and orthogonality assumptions to replace the absolute feature difference with the absolute projection, yet that replacement is already the quantity being optimized and does not repair the L1-versus-L2 problem in Eq. (7). The SAE-free derivation therefore rests on multiple unsupported equalities, and the paper should either supply a correct derivation or explicitly downgrade the theoretical status of the method.
minor comments (6)
  1. [Section 5.1 and Section 4.2] There are typos: 'BoosStep' should be 'BoostStep' in Section 5.1, and 'maximze' should be 'maximize' in Section 4.2.
  2. [Section 2.3, Eq. (2)] The notation uses ||cdot|| for the absolute value in Eq. (2) and surrounding text; using |cdot| would avoid confusion with vector norms.
  3. [Section 3.1] The text says the evaluation includes MATH-500, but Tables 1 and 2 report MATH-L3&L4; the relationship between MATH-500 and the reported MATH-L3&L4 subset should be stated explicitly.
  4. [Appendix B] The notation alpha^p_x is used in the replacement formula without being defined in that appendix; it should be defined consistently with Eq. (2).
  5. [Section 6, References] The reference 'Kharlapenko et al.' is incomplete: it lacks a year and publication venue, and it should be completed.
  6. [Figure 2] The caption says 'Top 10 activations' but the figure appears to show multiple panels; the caption should clarify which model and which activation ranking each panel corresponds to.

Circularity Check

3 steps flagged · score 6.0 of 10

Reported gains on MATHOAI and GSM8K are partly fitted: features and steering strength are derived from or tuned on the evaluation sets, while the SAE-free derivation's invalid L1-to-L2 reduction is a correctness issue rather than circularity.

  1. fitted input called prediction [Section 2.3, Eq. (2); Tables 1-2 (MATHOAI column)]
    "In practice, we randomly sample N = 100 samples:{qp}N p=1 from the MATHOAI dataset to generate the Verbal Reasoning Process {xp}N p=1 and the Symbolic Reasoning Process {yp}N p=1."

    The alpha_t ranking computed from these MATHOAI samples is exactly the input used to select the top-k steering features in Eq. (3), and Tables 1-2 list MATHOAI as an evaluation benchmark. The steering direction is thus constructed from the evaluation set; the reported MATHOAI gains are fitted, not predicted. The same N=100 MATHOAI pairs form the columns of A in Section 4.2, so the SAE-free MATHOAI result inherits the same overlap.

  2. fitted input called prediction [Section 3.1; Appendix A, Table 3]
    "For the GSM8K dataset, to find the best parameters and features more efficiently, we conduct our experiment on a subset that contains 300 randomly selected samples instead of on the entire dataset. The baseline results are detailedly shown in appendix Table 3. It means the results of subsets are roughly the same as the results on the entire dataset."

    Table 1's GSM8K numbers for each model equal the 'Subset' column of Appendix A Table 3 (e.g., DeepSeek-Llama3-8B 82.67), while no full-set steered accuracy is reported. Thus the +3.0 point gain is measured on the same 300-sample subset used to 'find the best parameters and features'; with n=300 the difference is within sampling noise and is not an out-of-sample validation.

1 more flagged steps
  1. fitted input called prediction [Figure 4; Appendix C; Appendix D]
    "Relationship between strength λ and accuracy in GSM8K and MATHOAI ... From Figure 4 we can see that a higher strength does not always improve the accuracy of the model's response. There is an optimal strength ... For SAE-free steering, we only conduct experiments on the eigenvectors corresponding to the top-10 eigenvalues, and the strength λ is still recommended to be less than 0.5."

    The steering strength lambda is selected from accuracy-versus-lambda curves computed on the same evaluation sets whose accuracy is then reported (GSM8K subset and MATHOAI). The set of 'suitable' SAE feature indices in Appendix C and the top-10 eigenvector restriction in Appendix D are likewise chosen by looking at benchmark performance. The consistent-improvement claim therefore incorporates test-set information into feature and hyperparameter selection.

full rationale

The paper's central SAE-free derivation in Eq. (7) is not circular: replacing the sum of absolute projections with the L2 norm is mathematically unjustified, but invalidity is a correctness defect, not self-reference. The self-citations to He et al. 2025, Shu et al. 2025, and Li et al. 2024 are descriptive and not load-bearing, so they do not raise the circularity score. The circularity that is present is in the evaluation protocol. The SAE feature ranking (Eq. 2) and the SAE-free matrix A are built from N=100 MATHOAI samples, and MATHOAI is one of the four benchmarks in Tables 1-2; the reported MATHOAI improvements are therefore in-sample fits. GSM8K results in Table 1 are reported on the same 300-sample subset used for the 'best parameters and features' search, with no full-set steered accuracy; the +3.0 point gain is within sampling error for n=300 and is not an independent prediction. Steering strength lambda is chosen from accuracy-versus-lambda curves on these same evaluation sets (Figure 4), and the feature indices and top-10 eigenvector restriction in Appendices C-D are post-hoc selections on the evaluation distribution. For MMLU-high and MATH-L3&L4 the features are transferred from MATHOAI, so those results retain some predictive content; hence the paper is partially but not wholly circular.

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

The free parameters show that the method depends on several hand-picked choices (lambda, layer, features, N, k) that are fitted to the test benchmarks. The axioms include two empirically-motivated assumptions (sparsity, orthogonality) that are only demonstrated for a single SAE, plus two ad hoc assumptions (error cancellation and L1/L2 equivalence) that lack support. The L1/L2 equivalence is the most serious because it is false in general.

free parameters (5)
  • Steering strength lambda = between 0.1 and 0.5, per dataset
    Tuned to maximize accuracy on the test benchmarks, as shown in Figure 4; no held-out validation.
  • Layer l = 15 for DeepSeek/Llama models, 9 for Gemma-2-9b
    Chosen for the experiments without a systematic ablation, and the SAE-free method also uses layer 15.
  • Number of samples N = 100
    Randomly sampled from MATHOAI to compute features; no sensitivity analysis is provided.
  • Selected SAE feature indices (SAE-based steering) = 24715, 20737, 20236, 14276, 15593, 17831 for Llama-Scope; 13419, 12085, 9540 for Gemma-Scope
    Manually identified in Appendix C as 'suitable features', which is a hand-picked selection based on empirical performance on the test sets.
  • Top-k for SAE-free steering = 10
    The paper reports only using eigenvectors corresponding to the top-10 eigenvalues (Appendix D), with no ablation across k.
assumptions (5)
  • domain assumption SAE reconstruction: h = sum_t alpha_t z_t + b + epsilon (Eq 4)
    Assumes the hidden activation can be decomposed into a weighted sum of SAE feature vectors plus bias and error, which is only approximate since SAEs are trained with a reconstruction loss.
  • domain assumption Sparsity assumption: top-k features capture about 90% of the norm
    Empirically shown for one SAE (Gemma-2-9b-it) in Figure 6, but assumed to hold for all SAEs and all layers without verification.
  • domain assumption Orthogonality assumption: SAE feature vectors are approximately pairwise orthogonal
    Empirically shown for top-150 features of one SAE in Figure 7, but assumed to hold for the top-k features used in the derivation.
  • ad hoc to paper Error cancellation: the mean residual errors of verbal and symbolic processes are equal and cancel in subtraction
    In Eq (6) the paper drops the error terms by 'approximately assum[ing] that the error of each token is equal'. No justification is given for this assumption.
  • ad hoc to paper L1/L2 equivalence: max sum_p |<x_p - y_p, z>| is proportional to max ||A^T z||_2
    This is mathematically incorrect in general. The maximizer of the sum of absolute projections is not generally the top eigenvector of AA^T. This assumption is load-bearing for the SAE-free derivation and is not supported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Feature Extraction and Steering for Enhanced Chain-of-Thought Reasoning in Language Models." pith.science (2026). https://pith.science/paper/OJFRJBQU

@misc{pith2026250515634,
  author       = {Pith},
  title        = {Pith review of: Feature Extraction and Steering for Enhanced Chain-of-Thought Reasoning in Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OJFRJBQU}},
  note         = {Machine review of arXiv:2505.15634}
}
read the original abstract

Large Language Models (LLMs) demonstrate the ability to solve reasoning and mathematical problems using the Chain-of-Thought (CoT) technique. Expanding CoT length, as seen in models such as DeepSeek-R1, significantly enhances this reasoning for complex problems, but requires costly and high-quality long CoT data and fine-tuning. This work, inspired by the deep thinking paradigm of DeepSeek-R1, utilizes a steering technique to enhance the reasoning ability of an LLM without external datasets. Our method first employs Sparse Autoencoders (SAEs) to extract interpretable features from vanilla CoT. These features are then used to steer the LLM's internal states during generation. Recognizing that many LLMs do not have corresponding pre-trained SAEs, we further introduce a novel SAE-free steering algorithm, which directly computes steering directions from the residual activations of an LLM, obviating the need for an explicit SAE. Experimental results demonstrate that both our SAE-based and subsequent SAE-free steering algorithms significantly enhance the reasoning capabilities of LLMs.

Figures

Figures reproduced from arXiv: 2505.15634 by the authors.

Figure 1
Figure 1. Main framework of SAE-based steering and SAE-free algorithm. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Top 10 activations for DeepSeek-Llama3-8B and Gemma-2-9b-it. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The generation length variation of the questions with the top 25% shortest answers generated by the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Relationship between strength λ and accuracy in GSM8K and MATHOAI. manual reviewing, we found that for non-reasoning LLM Llama3.1-8B-it, too long generation length will lead the LLM to generate some irrelevant in￾formation, steering may improve this phenomenon by gener…
Figure 5
Figure 5. Figure 5: Attention mapping of DeepSeek-Llama3-8B and DeepSeek-qwen-1.5B. Here α x t = 1 m Pm i=1 α r x i t and α y t = 1 n Pn j=1 α r y j t represent the mean activation of all tokens in x and y for feature αt at layer L. Back to our SAE-based steering, to obtain the activation…
Figure 6
Figure 6. Figure 6: Relationship between reconstruction error and [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 8
Figure 8. Figure 8: The prompt for generating Verbal Process. [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 7
Figure 7. Figure 7: Cosine Similarity of Top-k feature vectors [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 9
Figure 9. Figure 9: The prompt for generating Symbolic Process. [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 11
Figure 11. Figure 11: Eigenvector steering visualization for feature [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 10
Figure 10. Figure 10: Eigenvector steering visualization for feature idx 20903(structures and elements related to mathematical expressions and equations) for DeepSeek-Llama3-8B. J Explanation of Eigenvector In order to verify that the obtained eigenvectors have good interpretability, we ta…
Figure 12
Figure 12. Figure 12: Eigenvector steering visualization for feature [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]
Figure 13
Figure 13. Figure 13: Eigenvector steering visualization for feature [PITH_FULL_IMAGE:figures/full_fig_p016_13.png]
Figure 14
Figure 14. Figure 14: Spearman Correlation of Subtraction and Addition 16k/canonical) as the corresponding SAE. The result shows that compared with subtraction, addi￾tion will make the correlation lower, even less than 0, which means that the activation result obtained by subtracting can b…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 8 canonical work pages

  1. [2]

    **Substitute into the functional equation:** ax2 + bx + c + ay2 + by + c = a(x + y)2 + b(x + y) +c − xy − 1

    work page

  2. [3]

    arXiv preprint arXiv:2412.16720

    Openai o1 system card. arXiv preprint arXiv:2412.16720. Adam Karvonen, Benjamin Wright, Can Rager, Rico Angell, Jannik Brinkmann, Logan Smith, Claudio Mayrink Verdun, David Bau, and Samuel Marks

    arXiv

  3. [4]

    Advances in Neural Information Processing Systems, 37:83091–83118

    Measuring progress in dictionary learning for language model interpretability with board game models. Advances in Neural Information Processing Systems, 37:83091–83118. Dmitrii Kharlapenko, Stepan Shabalin, Fazl Barez, Neel Nanda, and Arthur Conmy. Scaling sparse feature circuits for studying in-context learning. Kenneth Li, Oam Patel, Fernanda Viégas, Ha...

    arXiv 2023

  4. [5]

    **Collect like terms:** (2a − 1)xy + (c − 2c + 1) = 0

    work page

  5. [6]

    **Set coefficients to zero:** 2a − 1 = 0 ⇒ a = 1 2 −c + 1 = 0 ⇒ c = 1

    work page

  6. [7]

    final answer

    Humanity’s last exam. arXiv preprint arXiv:2501.14249. Xin Quan, Marco Valentino, Danilo S Carvalho, Dhairya Dalal, and André Freitas. 2025. Peirce: Unifying material and formal reasoning via llm- driven neuro-symbolic refinement. arXiv preprint arXiv:2504.04110. Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Tom Lieberum, Vikrant Varma, János Kramá...

    arXiv 2025

  7. [8]

    **Final form of f (x):** f (x) = 1 2 x2 − 1 2 x + 1

    work page

  8. [9]

    **Expand and simplify:** ax2 + ay2 + bx + by + 2c = ax2 + 2axy + ay2 + bx + by + c − xy − 1

    work page

Show all 15 references
  1. [10]

    **Cancel common terms:** 2c = 2axy − xy + c − 1

  2. [13]

    **Use f (1) = 1:** 1 2 (1)2 + b(1) + 1 = 1 ⇒ 1 2 + b + 1 = 1 b = − 1 2

  3. [15]

    mathematical parameters related to dimensions and positioning

    **Solve f (n) = n:** 1 2 n2 − 1 2 n + 1 =n n2 − 3n + 2 = 0 (n − 1)(n − 2) = 0 n = 1, 2 **Final Answer:** 1, 2 User: {question} Figure 9: The prompt for generating Symbolic Process. I Eigenvector Steering Visualization To explore in more detail how SAE-free works and how its ro...

    2023

  4. [2013]

    Rout- ledge

    Research design and statistical analysis. Rout- ledge. Andrew Ng et al. 2011. Sparse autoencoder. CS294A Lecture notes, 72(2011):1–19. nostalgebraist. 2020. Logit lens. Accessed: 2024-06- 06. Liangming Pan, Alon Albalak, Xinyi Wang, and William Yang Wang. 2023. Logic-lm: Empow...

    2011 arXiv

  5. [2023]

    In The Twelfth Inter- national Conference on Learning Representations

    Let’s verify step by step. In The Twelfth Inter- national Conference on Learning Representations. Samuel Marks, Can Rager, Eric J Michaud, Yonatan Be- linkov, David Bau, and Aaron Mueller. 2024. Sparse feature circuits: Discovering and editing interpretable causal graphs in la...

    2024 arXiv

  6. [2024]

    arXiv preprint arXiv:2410.19278

    Applying sparse autoencoders to unlearn knowledge in language models. arXiv preprint arXiv:2410.19278. Andrey Galichin, Alexey Dontsov, Polina Druzhinina, Anton Razzhigaev, Oleg Y Rogov, Elena Tutubalina, and Ivan Oseledets. 2025. I have covered all the bases here: Interpretin...

    arXiv

  7. [2025]

    arXiv preprint arXiv:2504.19483

    Improving reasoning performance in large lan- guage models via representation engineering. arXiv preprint arXiv:2504.19483. Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richard- son, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.