{"id":"2cc2cf30-ca60-458f-a47e-2b0ddc283c81","arxiv_id":"2608.02957","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Highly discriminative steering vectors can reliably induce the opposite of the intended concept expression, and a representation-based spoof score can detect and correct this inversion.","lead":"This paper finds that some steering vectors in large language models, though highly accurate at detecting a concept like truthfulness, push the model toward the opposite behavior when used to steer. It introduces a representation-based score to detect such inverted vectors and uses sign flips to improve a standard steering method across most experiments.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ITI-RRF gains are confounded with a re-searched (k, alpha); the sign-flip mechanism is not isolated.","rationale":"The reader's weakest-assumption pinpoints spoof-score sign reliability outside the 90 mined heads, which is a real concern. I see an even more direct, load-bearing confound upstream of that: the ITI-RRF comparison changes both the steering signs and the (k, alpha) operating point, so the 27/30 headline cannot be attributed to sign flips without an ablation. The proposed control arm settles this directly. I do not object to the core existence claim for ISVs: the per-head steering curves on test sets and the geometric response patterns provide reasonable direct evidence, and the concentration bound is a genuine formal contribution. The weakness is concentrated in the applied ITI improvement claim, which is central to the paper's practical contribution and to the validation of the spoof-score sign. Since the missing ablation is a concrete, addressable condition rather than a contradiction of the phenomenon, the reader's CONDITIONAL verdict should stand; no stronger verdict change is warranted.","tokens_in":24153,"tokens_out":5100,"duration_ms":52800,"concrete_test":"Run three arms on the same model-concept pairs with the same degradation constraint and judge scoring: (A) ITI as reported; (B) ITI-RRF as reported; (C) ITI evaluated at the exact (k, alpha) selected for ITI-RRF in Table 4 but with canonical signs (no flips). If C is statistically indistinguishable from B, the reported gains come from hyperparameter re-search, not sign selection. A complementary arm is to take the ITI-selected (k, alpha) and flip only the signs flagged by s_spoof; if this does not beat A, the sign-flip signal is not driving the improvement. Report means and standard errors over at least 3 generation/judge runs so the comparisons are significance-testable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The applied claim that spoof-score sign flips improve ITI in 27/30 experiments is not established by the reported comparison. ITI and ITI-RRF differ in two ways: (i) which heads receive flipped signs, and (ii) the operating point (k, alpha), which is re-searched independently for each method (App. D.1). Table 4 shows these operating points diverge substantially, e.g., Gemma corrigibility promotion uses k=128, alpha=0.25 for ITI versus k=32, alpha=1.0 for ITI-RRF, and Olmo refusal promotion uses k=8, alpha=8.0 versus k=96, alpha=3.0. Since ITI is sensitive to k and alpha, a large share of the reported gains could come from moving to a better hyperparameter regime rather than from sign flips. No control evaluates ITI at the ITI-RRF-selected (k, alpha) without sign flips. The supporting validation of s_spoof is also in-sample: its 0.91 separation AUC is computed on the 90 extreme mined heads (Sec. 5, 'Representation Analysis'), and threshold-at-0 accuracy there is only 81%; Fig. 5 shows discriminability degrades as ISV/RSV criteria are relaxed, so extrapolating sign decisions to all top-k heads is untested. Both issues are addressable by one ablation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper documents an 'inverted detection-control' phenomenon in attention-head steering vectors: certain SVs that are highly discriminative for a concept and aligned with positive representations yet consistently promote the opposite behavior when applied with positive sign. The authors formalize inverted-steering vectors (ISVs) versus regular-steering vectors (RSVs) via discriminability, effect magnitude, and monotonicity criteria; mine 45 ISVs and 45 RSVs across 15 model-concept pairs; propose a forward-pass-only diagnostic, the representation response (sspoof), that measures how steering a candidate head shifts discriminative downstream heads' inner products with their own SVs; and show that sspoof separates the mined ISVs and RSVs with AUC 0.91. They then propose ITI-RRF, which flips the steering sign of heads with negative sspoof in the ITI pipeline, and report improved concept scores over ITI in 27/30 experiments, with gains from +0.9% to +138%, across Gemma 3 12B, Qwen 2.5 14B, and Olmo 3 7B. A finite-sample concentration bound for the inner-product response estimator is stated as Theorem 4.3 and proved in Appendix E.","tokens_in":24317,"tokens_out":18768,"duration_ms":153911,"significance":"If the existence claim holds, the paper identifies a genuine and surprising failure mode of a widely used assumption in activation steering, together with a diagnostic that requires no generation or LLM-judge scoring. Strengths: (i) the ISV phenomenon is substantiated by steering curves recomputed on test splits for 90 mined vectors across 15 model-concept pairs, not merely fitted on validation; (ii) Theorem 4.3 is a correct Hoeffding-style concentration bound with a complete proof in App. E; (iii) the sspoof diagnostic is training-free, and the paper is transparent about its search protocol (App. D.1) and reports standard deviations (Table 3). The main risk is that the applied claim, that sign flips themselves drive the reported ITI gains, is not isolated from the independently re-searched (k, alpha) operating points, and the sign rule is validated only on extreme mined heads. Both concerns are addressable with additional experiments; the central phenomenon appears defensible.","major_comments":[{"comment":"The reported comparison between ITI and ITI-RRF does not isolate the effect of the sign flips, because the two methods also operate at different (k, alpha) operating points that are re-searched independently on the validation set. Table 4 shows large divergences: for Gemma corrigibility promotion, ITI uses k=128, alpha=0.25 while ITI-RRF uses k=32, alpha=1.0; for Olmo refusal promotion, ITI uses k=8, alpha=8.0 while ITI-RRF uses k=96, alpha=3.0. Since ITI is known to be sensitive to both k and alpha, the headline '27/30 improvements' could be driven in substantial part by moving to a better hyperparameter regime rather than by correcting inverted signs. I recommend an ablation that evaluates ITI at the ITI-RRF-selected (k, alpha) without sign flips and, symmetrically, ITI-RRF at the ITI-selected (k, alpha) with flips; this would attribute the gain to the flip mechanism. As reported, the central applied claim of the paper is not yet established.","section":"Sec. 5 (Sign Selection for Detection-Based Steering); App. D.1; Table 4"},{"comment":"The sign-selection rule is validated on the 90 extreme mined ISV/RSV heads, and this validation is in-sample: the AUC of 0.91 and the threshold-at-0 accuracy of 81% are computed on exactly the vectors used to characterize the phenomenon (Fig. 4, right). Fig. 5 shows that the separation increases as the ISV/RSV criteria are tightened, implying it should be expected to decrease for heads that are less extreme; yet ITI-RRF applies the same threshold-at-0 rule to all top-k AUC-selected heads, including heads that are neither strong ISVs nor strong RSVs. No evaluation reports sign accuracy on that applied population. Given 81% accuracy even on the extreme heads, the net effect of the rule on the broader head population is untested. The paper should evaluate the spoof-score sign against behavioral ground truth on a random sample of top-AUC heads and report the distribution of sspoof among the heads actually flipped in the ITI-RRF experiments.","section":"Sec. 5 (Representation Analysis); App. D.2; Fig. 5"},{"comment":"Several of the 27 reported improvements are within the noise of the evaluation, so the abstract's claim of improvement in 27/30 experiments is stronger than the data support without significance testing. For example, Olmo myopia promotion changes from 2.26 to 2.28 (reported as +0.9%) with standard deviations of 0.57 and 0.59, and Olmo refusal suppression changes from 2.46 to 2.40 with standard deviations of about 0.90 and 0.98. The paper should report confidence intervals or paired tests on the per-experiment differences, or qualify the 27/30 count by the magnitude of the effect; the large improvements (e.g., Gemma corrigibility promotion, 1.50 to 3.57) would survive such a test, but the small ones would not.","section":"Table 1; Table 3"}],"minor_comments":[{"comment":"The TruthfulQA split listed in App. C sums to 790 (52+326+412) although the dataset is stated to contain 791 questions; please reconcile the counts.","section":"App. C"},{"comment":"Table 1 has a typesetting problem in the Qwen rows, where adjacent numbers are not separated (e.g., '1.962.772.072.072.07' and '1.422.211.99'); the table should be re-typeset with proper delimiters.","section":"Table 1"},{"comment":"In Definition 4.4 the random variable A over R is easily confused with the attention-head index a used throughout the paper; consider renaming the random variable (e.g., U) for readability.","section":"Sec. 4, Def. 4.4"},{"comment":"The sentence 'we leave the 3 cases to further investigation' should name the three experiments explicitly (per Table 1 these are Qwen truthfulness promotion, Qwen refusal promotion, and Qwen refusal suppression) and give at least a descriptive observation, since the refusal failures may reflect conflicting sign-flip decisions between promotion and suppression.","section":"Sec. 5"},{"comment":"The early cutoff of the alpha search upon >5% degradation, combined with the hierarchical tie-breaking rule, means that the two methods' selected operating points are determined by different portions of the validation grid; the paper should report whether the validation-selected (k, alpha) also performs best on the test set for each method.","section":"App. D.1"},{"comment":"The paper never reports the prevalence of ISVs among the screened discriminative heads (e.g., the fraction of heads with smono <= -0.5 per model-concept pair); reporting this statistic would contextualize both the practical impact of ITI-RRF and how often flips were actually applied in the 30 experiments.","section":"Sec. 5 / App. D"},{"comment":"The notation seffect is introduced only in the Fig. 5 caption; define it in the main text (as the appropriate of sisv and srsv) to avoid confusion.","section":"Fig. 5"},{"comment":"No code repository is provided, and dataset links are deferred to the camera-ready version; given the number of procedural details (mining thresholds, degradation pruning, tie-breaking in the alpha search), releasing code would substantially aid reproducibility.","section":"Availability"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest and well-documented, and the core phenomenon seems real; the main worry is that the quantitative headline (27/30) rests on a confounded comparison. I would like to see the ablation (ITI at ITI-RRF's (k, alpha) without flips; ITI-RRF at ITI's (k, alpha) with flips; and, ideally, sign accuracy of sspoof on a random sample of top-AUC heads) before publication. The three Qwen failures are currently swept under 'further investigation'; encouraging the authors to analyze them would strengthen the paper. If the ablation confirms the flip effect, this would be a solid contribution; if not, the ISV existence result may still be publishable, but with a substantially revised applied section."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the core empirical claim—that some highly discriminative steering vectors systematically invert their causal effect—is real and well-supported. The authors mine 90 extreme ISV/RSV heads across 15 model-concept pairs and recompute the steering curves on test sets; the effect holds there. That is the paper's strongest contribution, and I believe it. Second, the paper's applied claim—that spoof-score sign flips improve ITI in 27/30 experiments—is not established by the reported comparison, because ITI and ITI-RRF differ in two ways at once: which heads get sign-flipped, and the (k, alpha) operating point, which is re-searched independently for each method. Table 4 shows these diverge a lot (e.g., Gemma corrigibility promotion: ITI uses k=128, alpha=0.25; ITI-RRF uses k=32, alpha=1.0). No control runs ITI at ITI-RRF's (k, alpha) without sign flips. A single ablation would fix this; without it, the \"improvements\" could come largely from better hyperparameters.\n\nWhat is genuinely good: the geometric characterization is clever. The inner-product response and the spoof score are forward-pass-only, which is practically useful for sign selection. Theorem 4.3 is a correct Hoeffding-style concentration bound, carefully proven in the appendix. The related-work discussion is fair; the authors distinguish their aggregate-level finding from minority anti-steerable examples and low-discriminability explanations, and they cite the relevant prior work.\n\nSoft spots, in proportion: (1) The sign-flip ablation above is the main one. (2) The spoof score's separation AUC of 0.91 is computed in-sample on the 90 extreme mined heads, and the threshold-at-0 accuracy there is only 81%. Fig. 5 shows discriminability degrades as the ISV/RSV criteria are relaxed, so applying the sign to all top-k heads is untested. (3) The headline improvement percentages are not significance-tested; some \"large\" relative gains are measured against an ITI baseline that underperforms the unsteered model. (4) No code or data are released in this version, with links deferred to camera-ready. None of these issues contradict the existence of ISVs, which is the paper's best-supported claim.\n\nWho is this for? Anyone working on representation engineering or activation steering. The existence of inverted steering vectors should be known, and the sign-selection diagnostic is worth building on. It deserves a serious referee—I would send it out with a request for the ablation and for code/data. Reading group: yes.","headline":"A real phenomenon with a solid existence proof, but the applied ITI-RRF gains are confounded by a re-searched hyperparameter grid; worth refereeing after an ablation.","tokens_in":24974,"tokens_out":2630,"would_cite":true,"duration_ms":22570,"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":"Highly discriminative steering vectors can systematically promote the opposite of the concept they detect, and the paper introduces a generation-free representation statistic that identifies these inverted vectors and corrects their sign.","keywords":["steering vectors","inverted-steering vectors","representation response","inference-time intervention","linear representation hypothesis","sign selection","large language models","interpretability"],"falsifier":"Directly score a large sample of individual heads across a wide non-degraded range of alpha and compare each head's behavioral steering sign with its spoof-score sign; if a substantial share of heads with high AUC and monotonic inverted behavior have positive spoof scores, or monotonic regular heads have negative spoof scores, the representation-response characterization fails. A simpler version: run ITI-RRF with spoof-score sign flips replaced by signs taken from generated-text scoring; if those generated-text signs do not reproduce or beat the ITI-RRF gains, the sign signal is not the cause of the improvement.","tokens_in":23808,"feed_emoji":"🔄","tokens_out":6659,"duration_ms":59040,"temperature":0.7,"pith_summary":"This paper reports that a steering vector can be highly discriminative for a concept and still, when used to steer the model, reliably push behavior the other way. The authors call such directions inverted-steering vectors (ISVs), and document them in 15 model-concept pairs across three instruction-tuned language models. They explain the effect geometrically: steering any discriminative vector makes the steered head look concept-present, but an ISV sends downstream discriminative heads the opposite signal, as if the concept were absent. From that signature they build a generation-free spoof score, and show that flipping the steering sign of heads it flags improves a standard inference-time-intervention pipeline in 27 of 30 experiments.","feed_headline":"Steering vectors can detect a concept and still flip its effect","feed_subtitle":"A forward-pass-only score catches inverted heads and fixes inference-time steering in 27 of 30 experiments.","key_machinery":"The central object is the inner-product response (IPR): for a steered upstream head and a downstream head, it is the change in the downstream head's projection onto its own steering vector when the upstream head is steered, normalized by the natural variance of that projection. Averaged over steering factors and over all sufficiently discriminative downstream heads, the IPR becomes the representation response, or spoof score, whose sign separates regular steering vectors (positive) from inverted steering vectors (negative). This quantity carries the argument because it can be estimated from forward passes alone, without generation or judge-based scoring, and it supplies the sign-flip rule for ITI-RRF. Theorem 4.3 provides a concentration bound showing that the IPR can be estimated reliably from finitely many samples.","core_discovery":"The paper's central claim is that the causal direction of a steering vector is not determined by how well it detects the concept, even for vectors that are strongly aligned with concept-positive representations. In attention-head output spaces, mean-difference vectors with high AUC can have the opposite behavioral effect: positive steering suppresses the concept on average and negative steering promotes it, across the whole input distribution rather than on a minority of examples. The authors characterize this inversion by measuring how steering one head changes the inner products of downstream discriminative heads with their own concept directions; regular steering vectors make downstream heads look more concept-present, while ISVs make them look concept-absent. These measurements, taken at the final question token before generation, define the representation response, which the paper uses to identify ISVs and to flip steering signs in an ITI pipeline, yielding improvements in 27 of 30 experiments with gains from 0.9% to 138%.","pith_inferences":["Extending beyond the paper, the spoof-score sign could be adapted to layer-level steering or to affine and nonlinear interventions; if the geometric signature transfers, it would give a general diagnostic for causal direction without decoding.","A testable consequence of the characterization is that in an ISV, the sign inversion should be traceable to a specific downstream circuit, such as an OV matrix or MLP pathway; ablating that pathway should change the spoof score's sign.","Because the spoof score is computed on a conservative alpha range from -10 to 10, applying it to heads outside the extreme-validation population may need a calibration step; measuring its accuracy on moderate-AUC heads would show how far the current rule extends."],"forward_implications":["Detection-only selection of top-k heads by AUC is insufficient: among the heads it chooses, some may be pressing against the intended direction, so sign selection is a necessary step in pipelines like ITI.","Because the spoof score is computed without generation, sign correction can be added to steering pipelines at negligible extra inference cost.","ISVs occur systematically across many model-concept pairs, so the assumed link between a direction's discriminative orientation and its behavioral effect should not be taken for granted.","Sign choices made on validation generalize to test-time open-ended generations, improving both promotion and suppression in most settings.","Variability in ITI performance across models and concepts may in part reflect incorrectly oriented vectors, not just weak vectors."],"supporting_citations":[{"why":"Supplies the linear representation hypothesis, the assumption that a direction aligned with positive examples also controls the concept, which ISVs challenge.","marker":"[27]"},{"why":"Provides ITI, the foundational detection-based steering pipeline that the paper modifies and evaluates against.","marker":"[21]"},{"why":"Supplies the mean-difference steering vector estimation method used throughout the paper.","marker":"[17]"},{"why":"Argues that steering vectors should be viewed as correlational hypotheses requiring causal validation, the viewpoint this paper extends.","marker":"[33]"},{"why":"Documents anti-steerable examples where steering fails on a minority of inputs, a prior phenomenon the paper distinguishes from systematic ISVs.","marker":"[35]"},{"why":"Attributes anti-steerable behavior to low-discriminability vectors, in contrast to the high-AUC ISVs studied here.","marker":"[4]"},{"why":"Supplies the MWE datasets for corrigibility, wealth-seeking, and myopia used to mine and evaluate ISVs.","marker":"[28]"},{"why":"Provides the refusal dataset and the contrastive activation addition methodology used for open-ended steering.","marker":"[30]"},{"why":"Supplies TruthfulQA, the dataset used for the truthfulness concept and its evaluation.","marker":"[22]"}],"fun_headline_variants":["Steering vectors detect a concept, yet flip its effect","Detection doesn't equal direction in steering vectors","When a steering vector detects but inverts behavior","Sign flip fix boosts steering in 27 of 30 tests","Forward-pass score spots inverted steering vectors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method relies on the assumption that the sign of the spoof score, measured with cheap forward passes over a small range of steering strengths, reliably gives the correct steering sign for every top-k head during open-ended generation, even though this correspondence was validated on a set of 90 extreme mined heads and then applied to all candidate heads.","fun_headline_variants_meta":{"raw":{"variants":["Steering vectors detect a concept, yet flip its effect","Detection doesn't equal direction in steering vectors","When a steering vector detects but inverts behavior","Sign flip fix boosts steering in 27 of 30 tests","Forward-pass score spots inverted steering vectors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000163,"raw_usage":{"total_tokens":1253,"prompt_tokens":967,"completion_tokens":286,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":583,"completion_tokens_details":{"reasoning_tokens":213}},"tokens_in":583,"tokens_out":286,"duration_ms":5905,"temperature":1.0,"reasoning_tokens":213,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:55:10.072508+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Directly score a large sample of individual heads across a wide non-degraded range of alpha and compare each head's behavioral steering sign with its spoof-score sign; if a substantial share of heads with high AUC and monotonic inverted behavior have positive spoof scores, or monotonic regular heads have negative spoof scores, the representation-response characterization fails. A simpler version: run ITI-RRF with spoof-score sign flips replaced by signs taken from generated-text scoring; if those generated-text signs do not reproduce or beat the ITI-RRF gains, the sign signal is not the cause of the improvement.","supporting_citations":[{"cited_title":"The linear representation hypothesis and the geometry of large language models","cited_arxiv_id":null,"evidence_quote":"Supplies the linear representation hypothesis, the assumption that a direction aligned with positive examples also controls the concept, which ISVs challenge."},{"cited_title":"Open problems in mechanistic interpretability.Transactions on Machine Learning Research, 2025","cited_arxiv_id":null,"evidence_quote":"Argues that steering vectors should be viewed as correlational hypotheses requiring causal validation, the viewpoint this paper extends."},{"cited_title":"Analysing the generalisation and reliability of steering vectors","cited_arxiv_id":null,"evidence_quote":"Documents anti-steerable examples where steering fails on a minority of inputs, a prior phenomenon the paper distinguishes from systematic ISVs."}],"review_version":1}