{"id":"4154fed3-64ed-4be3-85d7-a6e8d9da7e38","arxiv_id":"2607.29062","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Steering vectors that make LLMs disclose injected hints transfer across cue types, datasets, and construction methods, but only reliably improve disclosure in the largest tested model and mainly on MMLU.","lead":"This paper tests whether steering an AI model's internal activity to make it mention hidden clues in its reasoning works beyond the exact setting it was built for. It finds the method only reliably works on the largest model, and that when it works, vectors built on other clues or datasets work about as well.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Per-scenario layer selection confounds the transfer claim: columns in Figs 2–3 may reflect eval-layer effects, not vector interchangeability; a fixed-layer control is required.","rationale":"The reader's weakest assumption—per-scenario layer selection confounding the central transfer result—is exactly the load-bearing concern. The paper's own conclusion flags it, which is honest but does not resolve it. A fixed-layer control is a concrete, feasible check that would settle whether the column-constant transfer pattern is a property of the vectors or of the eval-selected layers. No other concern is as directly tied to the headline claim: the single-judge labels could shift absolute numbers but are less likely to invert the train/eval asymmetry; the cue-use proxy has an explicit validation (Appendix D.8). Since the paper acknowledges this limitation and presents substantial supporting evidence (e.g., stability across construction methods, split-half control), the appropriate verdict remains CONDITIONAL, matching the reader. The recommended action is unchanged: require the fixed-layer control before accepting the abstract's strong generalization claim.","tokens_in":19724,"tokens_out":3194,"duration_ms":36042,"concrete_test":"For Gemma-3 12B, recompute the 4×4 cross-cue and 3×3 cross-dataset transfer matrices using a single fixed steering layer for all train/eval cells (e.g., L15, the layer selected for the GPQA-Stanford scenario, or a common mid-layer like L24), holding α=5. Then check whether off-diagonal Δack still clusters by eval setting (column-constant) rather than by train setting. If the pattern disappears or becomes train-dependent, per-scenario layer selection—not vector generality—was the driver; if the pattern persists, the transfer conclusion is robust to this confound.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Δack is determined primarily by the evaluation setting, not the train setting, supported by column constancy in the transfer matrices (Figures 2 and 3). But steering layers are selected per evaluation scenario as the argmax probe test AUROC, and every vector—regardless of where it was trained—is applied at that eval-specific layer. Thus each column shares not only the eval dataset/cue but also the same readout layer and steering coefficient. A column-constant pattern could arise simply because that fixed eval-layer produces a characteristic effect size for any input vector, with the vector's origin playing no causal role. The paper's own Section 4 acknowledges this: '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.' The Appendix D.5 cosine analyses use a common layer only for synthetic vectors and do not measure behavioral transfer, so they do not resolve this. Consequently, the abstract's 'effect size is determined primarily by the evaluation setting' conflates the evaluation setting with its selected layer; the generalization claim is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":20045,"tokens_out":6271,"duration_ms":69074,"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":[{"comment":"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","section":"§3.3, §4, Figs. 2–3"},{"comment":"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.","section":"§2 'Judge and metrics', Appendix F"},{"comment":"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.","section":"§3.1, §4, Table 5, Fig. 7"}],"minor_comments":[{"comment":"“not a prioriclear” should read “not a priori clear”.","section":"§3.4"},{"comment":"The abbreviation “uni.” (e.g., “all-cues (uni.),” “all-data (uni.)”) is not defined; please spell out “unified” or “all-cue”.","section":"Tables 5–8"},{"comment":"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.","section":"Appendix C"},{"comment":"“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.","section":"Appendix D.5"},{"comment":"“Red placements are uninformative” should be qualified as “uninformative with respect to the test setting” to avoid implying the train setting has no structure.","section":"Fig. 5 caption"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the manuscript is honest, technically competent, and accompanied by code and a good-faith split-half control. My major_revision recommendation rests on two fixable issues: a fixed-layer control for the transfer claim, and validation of the judge. Neither requires new theory, and the paper would be a solid workshop contribution after these additions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—\n\nYou should know this about arXiv:2607.29062: it is a genuinely careful, honest empirical study of whether faithfulness steering vectors generalize, but its headline claim is not yet established. The paper says that when steering works, the effect size is determined by the evaluation setting, not the training setting. The authors themselves flag the reason for caution: the steering layer is selected per scenario as the argmax probe test AUROC. So every vector applied to a given evaluation column shares the same layer and the same coefficient. Column constancy in their transfer matrices could just be that layer's characteristic effect, not vector interchangeability. A fixed-layer control is the missing experiment.\n\nWhat is actually new: this is the first systematic cross-cue, cross-dataset, cross-construction generalization study for faithfulness steering. It includes a cue-agnostic optimized vector, a hidden-cue-use decomposition, and a split-half control for shared evaluation noise. The four construction methods comparison is useful, and the synthetic vectors matching the contrastive ones on the largest model is a real, reproducible result. The paper is also unusually candid: train AUROC is 1.00 everywhere, layer selection uses test AUROC, and all labels come from a single unvalidated judge. It releases code.\n\nThe soft spots are, in order of importance. First, the layer confound is load-bearing. The transfer matrices in Figures 2 and 3 show columns consistent within an eval setting, but that consistency is exactly what you would expect if the eval layer dominates. The paper admits in Section 4 that layer and dataset effects are entangled, but it does not offer a control. So the central generalization claim remains unproven. Second, the single judge could be systematically wrong for some cues; with one unvalidated judge, judge error is unbounded. That is a real concern but more minor because the relative comparisons may still hold. Third, the experimental coverage is partial: cues vary only on GPQA, datasets only on the Stanford cue. The authors acknowledge this.\n\nWho gets value: anyone working on activation steering or CoT monitoring. The paper is a solid extension of Occhipinti et al., and the negative results on smaller models plus the hidden-cue-use finding are worth having. It deserves a serious referee—the empirical core is worth engaging—but the referee should ask for a fixed-layer control or a revised, more modest conclusion.\n\nRecommendation: accept for review with expectation of revision.","headline":"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.","tokens_in":20471,"tokens_out":4215,"would_cite":false,"duration_ms":43590,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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","keywords":["chain-of-thought faithfulness","activation steering","steering vector transfer","cue acknowledgment","hidden cue use","LLM interpretability","generalization","Gemma-3 12B"],"falsifier":"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.","tokens_in":19666,"feed_emoji":"🧭","tokens_out":6736,"duration_ms":66279,"temperature":0.7,"pith_summary":"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.","feed_headline":"Evaluation setting, not training data, decides steering effect","feed_subtitle":"On Gemma-3 12B, one faithfulness vector works across cues and datasets, cutting hidden cue use on MMLU from 31% to 5%.","key_machinery":"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","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"fun_headline_variants":["Steering vector effect transfers across cue types and datasets","For faithfulness, evaluation setting beats training setting","One steering method works across cues, but only on big models","Hidden reasoning drops from 31% to 5% with one vector","Steering targets verbalization, not cue salience"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Steering vector effect transfers across cue types and datasets","For faithfulness, evaluation setting beats training setting","One steering method works across cues, but only on big models","Hidden reasoning drops from 31% to 5% with one vector","Steering targets verbalization, not cue salience"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000572,"raw_usage":{"total_tokens":2613,"prompt_tokens":886,"completion_tokens":1727,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":630,"completion_tokens_details":{"reasoning_tokens":1647}},"tokens_in":630,"tokens_out":1727,"duration_ms":11560,"temperature":1.0,"reasoning_tokens":1647,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T14:21:39.682912+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[],"review_version":1}