{"id":"2ebe6a44-d420-48b1-a0ed-f68b9023748a","arxiv_id":"2505.15634","paper_version":4,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Activation steering guided by verbal-versus-symbolic CoT decomposition improves math accuracy by a few points, but the SAE-free derivation conflates L1 and L2 objectives.","lead":"This paper proposes two activation-steering methods to improve LLM math reasoning without long CoT data: one that uses sparse autoencoder features, and a simpler one that uses eigenvectors of contrastive activation differences. The methods show small accuracy gains, but the simpler method's derivation has a serious mathematical gap.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SAE-free derivation incorrectly equates maximizing sum of absolute projections with the top eigenvector of AA^T; the more decisive issue is that the reported gains are not reliable because steering directions are computed from the evaluation benchmarks and hyperparameters/features are tuned on…","rationale":"The reader's verdict is REJECT, and I agree the paper should not be accepted as is, but I would place the load-bearing weight on evaluation integrity rather than primarily on the Equation 7 derivation. The abstract's central claim is empirical: both steering methods 'significantly enhance the reasoning capabilities of LLMs.' For that claim to hold, the measured gains must be robust and indicative of generalization. The paper's protocol undermines this in several concrete ways. Section 2.3 samples N=100 problems from MATHOAI to compute the feature activations used for SAE-based steering, and the same dataset appears as an evaluation set in Tables 1-2. GSM8K results are reported on the same 300-sample subset used to search for features and hyperparameters, as stated in Section 3.1 and confirmed by Appendix A. Appendix C lists 'suitable' features chosen during this search, and Appendix D restricts the SAE-free method to top-10 eigenvectors; lambda is tuned using test-set curves in Figure 4. This post-hoc selection can inflate accuracy even without label leakage because the steering direction and strength are specialized to the exact test distribution and test sample. With n=300 and no error bars, the reported gains are within sampling noise: DeepSeek-Llama3-8B GSM8K 82.67 to 85.67 is about one standard error of the difference, and Llama3.1-8B-it 60.33 to 62.67 is similarly noisy. The derivation flaw in Equation 7 is also real: the objective sum_p |<x_p - y_p, z>| is an L1-type norm of A^T z, and its maximizer over the unit sphere need not be the leading eigenvector of AA^T, which maximizes the L2 norm. A simple counterexample with A = [e1, e2] in R^2 shows every unit vector maximizes the L2 objective, while only diagonal directions maximize the sum of absolute projections. Thus the SAE-free method currently lacks a valid theoretical justification. However, this error alone would not falsify the empirical claim; a heuristic could still work. The decisive question is whether the improvements survive a clean holdout evaluation. If they do, the paper could be revised to reframe the method as an empirical recipe and repair the derivation. If they do not, the central claim fails. Therefore I agree with REJECT and propose a concrete holdout re-evaluation as the settling test.","tokens_in":36472,"tokens_out":8908,"duration_ms":83192,"concrete_test":"Recompute Tables 1 and 2 with steering directions, feature indices, and hyperparameter lambda chosen only on a disjoint development set (e.g., 100 problems from the GSM8K training split), then evaluate on the full held-out GSM8K test set and on MATHOAI, MATH-L3&L4, and MMLU-high without any further tuning. Report mean accuracy with standard errors over at least three random seeds. If the average gains over the original models and baselines shrink to within noise or reverse, the claimed improvement is attributable to test-set-derived steering and post-hoc selection rather than general reasoning enhancement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim is empirical: both steering algorithms 'significantly enhance the reasoning capabilities of LLMs.' For that claim to hold, the measured gains must be robust and reflect generalization. The evaluation protocol does not establish this. In Section 2.3, N=100 samples from MATHOAI are used to construct the SAE feature ranking (Eq. 2) and, by the same setting, the SAE-free A matrix; MATHOAI is then an evaluation set in Tables 1-2. For GSM8K, Section 3.1 says parameters and features are found on a 300-sample subset, and Table 1 reports GSM8K results on that subset (Appendix A, Table 3), with no full-set steered number. Appendix C lists hand-picked 'suitable' features, Appendix D restricts to top-10 eigenvectors, and Figure 4 tunes lambda using test-set accuracy curves. These are post-hoc selections on the evaluation distribution. With n=300 and no error bars, a +3.0 point gain (82.67 to 85.67) is within roughly one standard error of the difference. Separately, Equation 7 is mathematically invalid: on the unit sphere, the sum of absolute projections is not proportional to the L2 norm of A^T z, so the top eigenvector of AA^T is not the claimed maximizer; a 2D counterexample with A = [e1, e2] shows every unit vector maximizes the L2 objective while only diagonal directions maximize the L1 objective. This invalidates the theoretical derivation of the SAE-free method, but the empirical claim could still survive if the recipe works heuristically. The decisive weakness is that the experiments do not separate genuine capability gains from test-set-derived steering and selection.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":36837,"tokens_out":6358,"duration_ms":56220,"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":[{"comment":"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.","section":"Section 4.2, Eq. (7) and Appendix E"},{"comment":"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.","section":"Sections 2.3, 3.1 and Appendices A, C, D"},{"comment":"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.","section":"Appendix B and Eq. (2)"},{"comment":"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.","section":"Section 4.2, Eq. (6)"}],"minor_comments":[{"comment":"There are typos: 'BoosStep' should be 'BoostStep' in Section 5.1, and 'maximze' should be 'maximize' in Section 4.2.","section":"Section 5.1 and Section 4.2"},{"comment":"The notation uses ||cdot|| for the absolute value in Eq. (2) and surrounding text; using |cdot| would avoid confusion with vector norms.","section":"Section 2.3, Eq. (2)"},{"comment":"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.","section":"Section 3.1"},{"comment":"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).","section":"Appendix B"},{"comment":"The reference 'Kharlapenko et al.' is incomplete: it lacks a year and publication venue, and it should be completed.","section":"Section 6, References"},{"comment":"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.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"I agree with the stress-test note that the load-bearing concerns land: the evaluation is circular and Eq. (7) is invalid. I recommend major revision rather than rejection because the empirical recipe could in principle be rehabilitated with a proper held-out evaluation, full GSM8K results, variance estimates, and a corrected derivation or an explicit heuristic framing. If the authors cannot provide out-of-sample results or fix the derivation, the claims should be substantially downgraded."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing worth knowing: this paper's real contribution is treating CoT as two streams—verbal and symbolic—and using their contrast to find steering directions. That decomposition is new in this context, and the SAE-free method (PCA on paired activation differences) is a sensible cheap approximation that deserves a fair test. The SAE-based steering results are also plausible; the case studies in the appendix show real qualitative changes in generation, and the attention analysis is a nice touch.\n\nThat said, the paper has two load-bearing soft spots. First, the derivation of the SAE-free direction in Section 4.2 and Appendix E is not valid. Maximizing the sum of absolute projections, sum_p |<x_p - y_p, z>|, is not the same as maximizing ||A^T z||_2. On the unit sphere these objectives have different maximizers; a 2D example with A = [e1, e2] shows every unit vector maximizes the L2 norm while only diagonal directions maximize the L1 sum. The paper needs a repair here—either a real argument for why the L1 maximizer can be replaced by the top eigenvector, or a weaker claim that the eigenvector is a good heuristic. As written, the central theoretical claim collapses.\n\nSecond, the evaluation protocol does not separate genuine gains from test-set fitting. Steering directions are computed from N=100 MATHOAI samples, and then accuracy is reported on MATHOAI. For GSM8K, only a 300-sample subset is used, with no full-set steered number. Steering strength lambda is tuned using accuracy curves on the same benchmarks, and the selected SAE features are hand-picked based on those results. With no error bars, a +3-point gain on a 300-sample subset is within roughly one standard error. The abstract's claim of \"significant enhancement\" is not supported by the evidence as presented.\n\nI would not cite this in its current form, but I would send it to a serious referee. The core idea—using verbal/symbolic decomposition to guide steering—is worth engaging with, and the SAE-free recipe is testable even if the derivation is flawed. The paper needs a major revision: fix or drop the theoretical derivation, add held-out validation for feature selection and lambda, and report full-set numbers with error bars. If that happens, it could become a solid contribution to the steering literature. For now, treat the empirical claims with caution.","headline":"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.","tokens_in":37383,"tokens_out":1359,"would_cite":false,"duration_ms":15141,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T50","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["chain-of-thought","sparse autoencoder","representation steering","verbal-symbolic decomposition","residual activations","mathematical reasoning","eigenvector steering","LLM interpretability"],"falsifier":"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.","tokens_in":36248,"feed_emoji":"🧮","tokens_out":5726,"duration_ms":41739,"temperature":0.7,"pith_summary":"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.","feed_headline":"Steering LLM activations boosts math reasoning without fine-tuning","feed_subtitle":"Verbal-vs-symbolic contrast yields a steering direction that lifts accuracy on four benchmarks, no SAE needed.","key_machinery":"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 $\\|A^\\top z\\|_2$.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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 $\\lambda$."],"supporting_citations":[{"why":"Supplies the DeepSeek-R1 deep-thinking paradigm that motivates amplifying latent reasoning, and the DeepSeek-Llama3-8B evaluation model.","marker":"Guo et al., 2025"},{"why":"Provides the GSM8K dataset used as a benchmark and for sampling steering examples.","marker":"Cobbe et al., 2021"},{"why":"Supplies the MMLU-high subset used as an evaluation benchmark.","marker":"Hendrycks et al., 2021"},{"why":"Supplies the MATH-500 (MATH-L3&L4) benchmark used for evaluation.","marker":"Lightman et al., 2023"},{"why":"Introduces sparse autoencoders as the feature-extraction tool used in the SAE steering variant.","marker":"Cunningham et al., 2023"},{"why":"Supplies the activation-steering formula the paper adopts for SAE-based steering.","marker":"Templeton et al., 2024"},{"why":"Represents the contrasting-representations steering approach the paper improves upon by using verbal/symbolic contrast instead of long-vs-short CoT.","marker":"Tang et al., 2025"},{"why":"Provides the BoostStep baseline that the SAE-free method is compared against.","marker":"Zhang et al., 2025"},{"why":"Provides the MathNeuro neuron-scaling baseline that the SAE-free method is compared against.","marker":"Christ et al., 2024"},{"why":"Supplies the Gemma Scope SAEs used to extract features for Gemma-2-9b-it.","marker":"Lieberum et al., 2024"}],"fun_headline_variants":["No fine-tuning: steering LLM activations lifts math reasoning","Steer LLM chain-of-thought features to boost math accuracy","Activation steering, no SAE, improves LLM math without extra data","How to make LLMs reason deeper without any fine-tuning data","LLM math skills up via steering, not training, per new study"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["No fine-tuning: steering LLM activations lifts math reasoning","Steer LLM chain-of-thought features to boost math accuracy","Activation steering, no SAE, improves LLM math without extra data","How to make LLMs reason deeper without any fine-tuning data","LLM math skills up via steering, not training, per new study"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000195,"raw_usage":{"total_tokens":1348,"prompt_tokens":929,"completion_tokens":419,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":545,"completion_tokens_details":{"reasoning_tokens":327}},"tokens_in":545,"tokens_out":419,"duration_ms":4118,"temperature":1.0,"reasoning_tokens":327,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:12:58.578942+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[],"review_version":1}