acceptodds
Under review as a conference paper at ICLR 2027

Contrastive Covariance Subspaces For Activation Steering

23 views

Abstract

Activation steering controls large language model behavior by adding a direction to hidden activations during inference. For steering to work well, that direction must be a pure signal for the target concept: any confound entangled with it, such as an incidental topic or writing style, is injected into the residual stream alongside the concept itself and degrades the generated text. We introduce Contrastive Covariance Steering (CCS), a training-free method that eliminates this confound variation. CCS projects steering vectors to a low-dimensional subspace obtained in closed form by solving a contrastive variance objective, which retains the directions along which the target concept varies while suppressing directions shared with the other classes. Across three LLMs and four steering domains, CCS consistently beats all baseline steering methods in all 12 model–domain settings, with the largest gains on AxBench (– relative) and Emotion (–). A controlled diagnostic shows why: planting a topic nuisance into the training data reduces unprojected steering by , whereas CCS loses only and recovers most of the gap to an oracle that deletes the planted direction explicitly. A human study with 23 raters on generations reproduces the LLM judge's ranking. We also perform several ablation studies to understand important aspects of our method, like robustness to the number of nuisance classes and the importance of mean centering and contrastive projection. Finally, in 20-turn conversations, CCS sustains concept expression while fluency degrades markedly slower than other steering methods, matching system prompt performance as the latter drifts.

Primary area:
generative models
Keywords:
Activation Steering, Second-order Statistics, LLMs

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_p9agOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses a well-known failure mode of steering as steering vectors estimated from contrastive data (e.g., difference-of-means) entangle the target concept with incidental variation (topic, style, length), and injecting such vectors degrades generation quality. The authors propose a new method Contrastive Covariance Steering (CCS) which centers all class activations on the grand mean of class means and builds per-class scatter matrices about this common center. No need for training, the CCS beats every activation-steering baseline in 3 LLMs (Llama-3.1-8B, Qwen3-8B, Gemma-2-9B) × 4 domains (AxBench, BIG5/NPTI, Emotion, TruthfulQA).

Critical strengths
  1. The entanglement account of why steering vectors hurt fluency is intuitive, and the paper consistently ties design choices back to this story. This makes the method easy to follow and understand. Also, the human evaluation is more rigorous than most steering papers.
  2. Solid experiments and evaluation. The multi-turn evaluation is a genuinely useful contribution. The author conduct an important experiments on 20-turn steering with a joint concept-fluency metric and the finding that CCS overtakes system prompting by turn 20 (80% vs. 71%) is an interesting counterpoint to the "prompting is unbeatable" narrative from AxBench.
Critical weaknesses
  • The paper is well structured but the evaluation is confined to concept expression and text quality; capability and safety side effects are unmeasured. This makes the contribution limited. There is no measurement of what steering does to general capabilities (e.g., MMLU, reasoning) or safety-relevant behavior, yet recent multi-dimensional evaluations (e.g., SteerScope-style suites) show that steering efficacy and side effects are strongly coupled, and that method rankings can change once side effects are priced in.
  • Each component is a classical tool: grand-mean centering, the total-scatter decomposition (the standard LDA scatter identity), a contrastive variance objective in the spirit of contrastive PCA, and projection filtering via eigendecomposition. The paper's contribution is primarily a careful adaptation and thorough validation of these ideas for steering-vector construction. From my perspective, the paper is good but not that inspiring. I would like to raise my score if I can see more interesting and promising results for current reasoning and agentic era.
  • Some writing and presentation comments.
    • H~c\tilde{H}_c is used in Eq. 3 (l. 183) but only introduced in the following paragraph (l. 190). Moreover, the nc−1/2n_c^{-1/2} column scaling is part of H~c\tilde{H}_c's definition, yet Eq. 3 already relies on it to identify \tilde{H}_c\tilde{H}_c^\top. Please define \tilde{H} before Eq. 3.
    • WW in Eq. 4 (l. 192) is never formally introduced (e.g., "W∈Rd×kW \in \mathbb{R}^{d \times k} with orthonormal columns spanning the sought subspace"). WW is also an unfortunate choice of symbol, as it conventionally denotes model weight matrices; consider UU or QQ.
    • The target-class index tt, the contrastive weight rr, and the projector PtP_t all appear in Figure 2 before their formal definitions (l. 190–192, Eq. 6).
    • Please introduce every mathematical symbol near its first appearance, including inside figure panels.
    • Figure 1 is hard to parse and Figure 2 is clearer in structure but has consistency problems. It appears in the Introduction (p. 2), before any notation exists, yet uses μtarget\mu_{\text{target}}, μG\mu_G, vCCSv_{\text{CCS}}, 2σt2\sigma_t, 2σu2\sigma_u, 2σn2\sigma_n without any legend; the caption does not define them either. Also, the plane, the contrastive subspace, i.e., the central object of the paper, is unlabeled. The covariance-ellipse convention ("2σ2\sigma") is never explained.
    • In figure 2, Terminology drifts across the paper: panel 1 says "Competing", Figure 1 says "Nuisance", the text says "nuisance classes" and "non-target scatter". Please unify. Panel 3 states the intervention as h′=h+α vCCSh' = h + \alpha\, v_{\text{CCS}}, which matches neither Eq. 1 (unit-normalized v^\hat{v}) nor the actual intervention used (Eq. 7, per-token norm scaling). Please align the figure with the real steering rule. Panel 2 draws the nuisance ellipse centered on μG\mu_G and the target ellipse offset from it; it is unclear whether this asymmetry is intentional (scatter about the common center includes the displacement term δcδc⊤\delta_c\delta_c^\top, so both should in general be off-center).
Questions for authors
  1. Does CCS preserve general capabilities (e.g., MMLU or a comparable benchmark) at the steering strengths selected in Table 1, and how does this compare to CAA/Linear-AcT at matched concept-expression levels?
  2. Was the intervention layer (15 for Llama, 21 for Qwen/Gemma) re-validated for CCS, or inherited from prior work on other methods?
5 comments
AuthorsOpenReview ↗

Response to Reviewer p9ag (1/4)

W1 + Q1: Capabilities and safety side effects

We thank the reviewer for highlighting the importance of capability and safety side effects; we agree that steering effectiveness alone is insufficient. We ran additional evaluations of general capabilities and safety when steering towards AxBench concepts, for all three models, both at the original Table 1 strengths and at matched steering effectiveness. We also study a reasoning model with thinking enabled.

Evaluation setup.

  • General capabilities: MMLU (validation) (Hendrycks et al., 2021) and ARC-Challenge (test) (Clark et al., 2018), on a fixed 1,000-question subset stratified by MMLU subject and ARC (seed 0); we report the change in accuracy relative to the unsteered model.
  • Safety: HarmBench (200 harmful requests) (Mazeika et al., 2024) and XSTest (250 safe and 200 unsafe prompts) (Röttger et al., 2024), with Qwen2.5-32B-Instruct judging harmful compliance, over-refusal and unsafe compliance.

We additionally sweep each activation-based method over 10 steering strengths, evaluating capability on the same 1,000 questions and safety on all 650 prompts at every strength, which allows comparing methods at matched steering effectiveness. Due to time constraints, the sweeps use 10 of the 50 AxBench concepts, sampled at random with the same genre mix as the full set (7 text, 2 code, 1 math).

1. Capability at the original steering strengths. Table R1 reports the AxBench effectiveness from Table 1 together with the accuracy change at the same operating points, averaged over the three models and all 50 concepts.

Table R1.

UnsteeredCCSCAARepEPCALinear-AcTSEAAngularPrompting
Effectiveness0.160.830.400.360.490.490.240.271.02
ΔAcc–−18.6−19.2−12.2−16.9−17.4−3.2−4.5−2.6

At its Table 1 strengths, CCS reaches by far the highest effectiveness among activation-based methods (0.83 vs. at most 0.49), and its accuracy loss (−18.6 points) is in the same range as CAA, PCA and Linear-AcT (−16.9 to −19.2), which steer much less. Strong interventions therefore degrade capabilities for all additive methods, and the Table 1 operating points differ in strength across methods. This motivates comparing methods at matched effectiveness.

2. Capability at matched steering effectiveness. For each baseline, we take its Table 1 effectiveness and compare its capability change with that of CCS at the same effectiveness, using the strength sweeps.

Table R2. Accuracy change (points) at matched effectiveness, baseline / CCS, averaged over three models.

BaselineΔAcc (baseline / CCS)
CAA−18.9 / −4.4
RepE−17.4 / −2.7
PCA−13.0 / −8.9
Linear-AcT−18.8 / −7.5
SEA−1.5 / −0.7
Angular−3.3 / −1.3

At matched effectiveness, CCS loses less accuracy than every baseline, e.g. 4.4 vs. 18.9 points for CAA and 7.5 vs. 18.8 points for Linear-AcT. These values use the 10 sweep concepts and therefore differ from Table R1.

References

Clark, P., et al. Think you have solved question answering? Try ARC, the AI2 reasoning challenge. arXiv:1803.05457, 2018.

Hendrycks, D., et al. Measuring massive multitask language understanding. ICLR, 2021.

Mazeika, M., et al. HarmBench: A standardized evaluation framework for automated red teaming and robust refusal. ICML, 2024.

Röttger, P., et al. XSTest: A test suite for identifying exaggerated safety behaviours in large language models. NAACL, 2024.

AuthorsOpenReview ↗

Response to Reviewer p9ag (2/4)

3. Steering under capability constraints. To see how much can we steer with each method if we don't want to drop the model's general capabilities significantly, we evaluate the highest steering effectiveness achievable while limiting the average accuracy loss to 2 or 5 points. We choose the strength for each concept that maximize the average effectiveness, subject to the average accuracy loss staying within the budget, and average over the three models. We also report the best effectiveness without a budget.

Table R3. Highest average AxBench effectiveness within each accuracy budget (average accuracy change in parentheses); strengths in {0.2, 0.4, 0.6, 0.8, 1.0}.

BudgetUnsteeredCCSCAARepEPCALinear-AcTSEAAngular
≤ 2 points0.0780.366 (−1.9)0.216 (−1.8)0.210 (−1.6)0.303 (−2.0)0.180 (−1.9)0.118 (0.0)0.160 (−0.3)
≤ 5 points0.0780.474 (−4.5)0.255 (−4.7)0.223 (−3.1)0.362 (−3.9)0.261 (−4.9)0.123 (−0.9)0.160 (−0.3)
No budget0.0780.646 (−21.0)0.297 (−18.0)0.256 (−16.9)0.398 (−12.3)0.408 (−20.7)0.127 (−2.6)0.166 (−2.8)

The scores are computed on the 10 sweep concepts and therefore differ from Table 1. Under both budgets CCS achieves the highest effectiveness: with an average accuracy loss of at most 5 points it reaches 0.474, compared with 0.362 for PCA and 0.255 for CAA. CCS therefore retains its advantage when capability preservation is explicitly constrained.

4. Safety side effects. We compare safety changes at matched steering effectiveness on AxBench, averaged over the three models (baseline / CCS; higher is worse).

Table R4.

BaselineHarmBench (harmful compliance)XSTest (over-refusal)XSTest (unsafe compliance)
CAA−1.4 / −0.9+3.2 / +4.6+0.1 / −1.5
RepE−1.5 / −0.2+2.9 / +2.6+0.1 / −0.7
PCA−1.6 / −1.6+4.8 / +5.1−0.9 / −1.4
Linear-AcT−1.4 / −1.4+5.0 / +5.3+10.6 / −1.2
SEA+0.1 / +0.1−0.1 / +0.4−0.1 / +0.1
Angular+0.1 / 0.0+0.4 / +0.9+2.2 / −0.3

Harmful compliance on HarmBench changes little for all methods (between −1.6 and +0.1 points). Over-refusal on XSTest increases for all additive methods, by similar amounts for CCS and its matched baselines (+2.6 to +5.3 vs. +2.9 to +5.0 points). Unsafe compliance decreases for CCS in almost all comparisons, whereas Linear-AcT increases it by 10.6 points on average. Overall, CCS does not show systematically larger safety side effects in these evaluations, although these results do not establish general safety guarantees.

AuthorsOpenReview ↗

Response to Reviewer p9ag (3/4)

5. Reasoning models: steering with thinking enabled. We evaluate Qwen3-8B with thinking enabled, applying steering either only to the final answer or to both the reasoning trace and the answer, and compare against the same model with thinking disabled. All settings use the same steering directions, 10 AxBench concepts, evaluation prompts and capability questions; accuracy changes are relative to each setting's unsteered baseline. Due to time constraints, the thinking results use a subset of about half of the 1,000 questions, the same for all three settings. As in Table R3, each concept uses its own strength, chosen to maximize average effectiveness within a 2- or 5-point accuracy budget, or without a budget.

Table R5. Qwen3-8B: highest average AxBench effectiveness within each budget (average accuracy change in parentheses).

SettingMethod≤ 2 points≤ 5 pointsNo budget
Thinking, answer onlyCAA0.41 (−1.7)0.41 (−3.7)0.41 (−3.7)
PCA0.39 (+1.2)0.39 (+1.2)0.39 (+1.2)
CCS0.66 (−1.3)0.66 (−1.3)0.66 (−1.3)
Thinking, reasoning + answerCAA0.33 (−2.0)0.37 (−4.7)0.39 (−14.6)
PCA0.40 (−2.0)0.43 (−3.3)0.43 (−3.3)
CCS0.59 (−1.7)0.74 (−4.8)0.74 (−4.8)
No thinkingCAA0.21 (−1.7)0.25 (−4.6)0.40 (−26.7)
PCA0.30 (−1.6)0.34 (−4.9)0.35 (−10.1)
CCS0.38 (−2.0)0.53 (−4.9)0.58 (−13.0)

Restricting steering to the final answer of a thinking-enabled model substantially improves the effectiveness–capability trade-off. Under a 2-point budget, CCS achieves 0.66 effectiveness with answer-only steering, compared with 0.38 without thinking. Without a budget, answer-only steering reaches 0.66 at a 1.3-point accuracy loss, whereas the model without thinking reaches 0.58 at a 13.0-point loss. Since the reasoning trace that solves the question is left unsteered, reasoning accuracy is largely preserved while the concept is still expressed in the answer. Steering both the reasoning trace and the answer reaches the highest unconstrained effectiveness (0.74), at a 4.8-point accuracy loss. Across all settings and budgets, CCS achieves the highest effectiveness. This table reports Qwen3-8B only, whereas Table R3 averages three models over separate generations. These preliminary findings suggest that applying interventions only during final-answer generation can substantially mitigate capability degradation while retaining effective control over the expressed behavior.

Summary and changes to the paper. CCS, like other additive methods, can substantially degrade general capabilities at strong intervention settings. However, at matched steering effectiveness and under explicit accuracy-loss budgets, it provides a better effectiveness–capability trade-off than competing activation-steering methods, without systematically worse safety outcomes. For reasoning models, steering only the final answer largely preserves reasoning accuracy while keeping concept effectiveness high. We have added these results to the revised paper: the capability and safety results in Appendix I (Tables 23–26) and the thinking-model results in Appendix J (Table 28).

Reviewer_q1RHOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper aims to provide a closed-form method that improves steering by projecting out variance that is shared between the target concept and irrelevant concepts, while keeping variance that is unique to the target concept. In doing so, the method removes directions from the steering vector that are often harmful to fluency or concept expression such as style, formatting etc. The method shows strong results across a comprehensive empirical evaluation.

Critical strengths
  • The authors provide a comprehensive empirical evaluation across three models, four benchmarks, multiple baselines, and a diverse range of concepts. The consistent results across these settings provide strong evidence of the method’s generalizability.

  • The reported results are strong, with CCS outperforming the activation steering baselines across various benchmarks. Given that the concerns are addressed, the paper’s contribution is meaningful.

Critical weaknesses
  • While the paper shows strong results against multiple baselines, I believe comparisons with stronger activation steering methods, such as ReFT-r1 (among others), are important. Such results would strengthen the contribution by showing that the method is competitive with stronger activation steering methods.

  • The GMC ablation on Axbench helps isolate the contribution of the projection beyond using the nuisance classes for mean centering. However, for the remaining methods and benchmarks, it is unclear how the nuisance dataset was constructed and whether the baselines had access to the same data. If CCS used data unavailable to the other baselines, additional comparisons in the same exact data setting are warranted, for example adding the nuisance classes to their negative or reference distributions for other methods. This would help establish the method’s contribution beyond differences in dataset construction. This is related to the first question in “questions for authors”.

Given the paper's strengths, I'd be happy to reconsider my assessment if these concerns are addressed.

Questions for authors
  • The construction of CCS depends on the nuisance classes, but the paper does not clearly specify which datasets and classes are used as nuisance for each benchmark. It would be appreciated if the authors could clarify how the nuisance data was constructed and whether it is a different dataset per benchmark? Additionally, do the baselines receive access to the same nuisance examples, for example as part of their negative or reference distributions? This distinction is important for determining how much CCS improves steering beyond what can be achieved by the baselines using the same data.

  • Could the authors provide additional statistics for the main results such as variance or confidence intervals?

4 comments
AuthorsOpenReview ↗

Response to Reviewer q1RH (1/2)

W1: Stronger baselines

We welcome the reviewer's suggestion, as it strengthens our paper's results. We added three gradient-trained steering methods: ReFT-r1 and LoReFT from AxBench, and the preference-based BiPO. ReFT-r1 (Wu et al., 2025) learns a rank-1 steering direction jointly as a concept detector and steering vector, with a sparsity penalty on its latent. LoReFT (Wu et al., 2024) learns a low-rank intervention that edits hidden representations within a learned subspace. BiPO (Cao et al., 2024) optimises a steering vector so that it raises the likelihood of preferred over dispreferred responses.

Each method is run as published (its own loss and steering operator), trained on the benchmark's own data at the same layer as CCS, evaluated with the same prompts, decoding and judges as Table 1, and tuned over its published hyperparameter space with the same selection protocol as CCS. Training details are in Appendix A of the revised paper.

Table R1.

ModelReFT-r1LoReFTBiPOCCS (training-free)
TruthfulQALlama53.260.474.480.3
Qwen48.750.152.676.8
Gemma78.464.694.492.2
EmotionLlama26.526.916.977.9
Qwen26.339.822.480.1
Gemma36.457.537.487.3
BIG5Llama75.181.971.883.1
Qwen69.873.270.786.3
Gemma60.780.976.082.4
AxBenchLlama0.980.800.490.78
Qwen0.710.840.260.86
Gemma0.730.870.310.84

CCS, which is closed-form and requires no training, outperforms all three trained methods in 9 of 12 settings: on every model for Emotion (by 30–51 points over the strongest, LoReFT) and BIG5, on TruthfulQA for Llama and Qwen, and on AxBench for Qwen. On Gemma/TruthfulQA, BiPO, which is trained directly on true-vs-false answers, is higher (94.4 vs. 92.2). On AxBench, ReFT-r1 is higher on Llama (0.98 vs. 0.78) and LoReFT on Gemma (0.87 vs. 0.84). We have added these results to Table 1.

References

Cao, Y., et al. Personalized steering of large language models: Versatile steering vectors through bi-directional preference optimization. NeurIPS, 2024.

Wu, Z., et al. ReFT: Representation finetuning for language models. NeurIPS, 2024.

Wu, Z., et al. AxBench: Steering LLMs? Even simple baselines outperform sparse autoencoders. ICML, 2025.

AuthorsOpenReview ↗

Response to Reviewer q1RH (2/2)

W2 + Q1: Nuisance data and data parity with baselines

As described in the Method section, the nuisance classes for each target are all other classes of the same benchmark. For example, when steering happiness in the Emotion benchmark, all remaining emotion classes, including the neutral class, are nuisance classes. The nuisance data is therefore benchmark-specific, and no examples from other datasets or benchmarks are used.

The baselines have access to the same benchmark data as CCS. However, their standard formulations construct steering directions from the target class and a single negative or neutral reference class, rather than exploiting the structure of all non-target classes. One way to give them the additional data is to pool all non-target examples into their negative/reference distribution, as the reviewer suggests. For CAA, Linear-AcT, Angular and SEA we replace the negative/reference set by all other classes pooled (AxBench: the other 49 concepts plus the negative class; Emotion: the other six emotions plus neutral; BIG5: the other traits' poles), with everything else unchanged (layer, prompts, strength grid, judge and the selection rule of Table 1). TruthfulQA has only two classes, so we don't include it in the comparison. CAA is run on all three models, the remaining baselines on Qwen3-8B.

Table R2. Baselines with their published reference set and with all other classes pooled.

MethodModelAxB publishedAxB all classesEmo publishedEmo all classesBIG5 publishedBIG5 all classes
CAALlama0.310.5659.273.874.480.5
CAAQwen0.410.6462.976.976.582.3
CAAGemma0.470.6378.385.677.079.8
Linear-AcTQwen0.560.4950.641.780.879.6
AngularQwen0.280.2526.036.068.368.0
SEAQwen0.220.2220.620.767.967.8
CCSLlama0.7877.983.1
CCSQwen0.8680.186.3
CCSGemma0.8487.382.4

Pooling the other classes into CAA's reference helps CAA substantially (AxBench +0.16 to +0.25, Emotion +7 to +15, BIG5 +3 to +6), but CCS, with its subspace (k,r)(k, r) tuned as in Table 1, remains the best method on every benchmark and model (by 0.21–0.22 on AxBench, 1.7–4.1 points on Emotion and 2.6–4.0 on BIG5). For Linear-AcT, Angular and SEA, pooling does not help: it lowers Linear-AcT on Emotion and AxBench and Angular on AxBench, leaves SEA unchanged, and only raises Angular on Emotion; all three remain far below CCS. The extra classes are informative and also help CAA, but the way CCS uses them, a per-target contrastive subspace, is what keeps it ahead of the same data pooled into a single reference. We have included this ablation in Appendix H of the revised paper (Table 22).

Q2: Variance / confidence intervals

We report 95% bootstrap confidence intervals for every cell of Table 1 and paired tests of CCS against each baseline (Appendix F of the revised paper). CCS has a higher score than every baseline in all 72 comparisons (12 model–benchmark settings × 6 baselines), and the difference is significant (95% CI of the paired difference excluding 0) in 64 of them, including every comparison on AxBench, Emotion and TruthfulQA for all three models.

Thank you for the comprehensive rebuttal. The additional baselines and experiments substantially address my original concerns and strengthen the paper. I have updated the score accordingly.

One remaining point concerns the held-out evaluation. The main results select the steering strength (alpha) based on performance on the evaluation prompts. While Appendix E reports nested cross-validation results for the original baselines, it does not include the newly added ReFT-r1, LoReFT, and BiPO methods. Could the authors also report held-out results for these three methods under the same protocol?

Thank you for the positive feedback and for updating your score. We have run the three trained methods under the same nested cross-validation protocol as Table 7 with all hyperparameters, including the inference-time steering factor selected with the same protocol. CCS remains the best method in 9 of 12 settings against the trained baselines even in the nested CV comparison.

  • AxBench. The trained methods lose less of their score than CCS, and a trained method now leads on all three models (ReFT-r1 on Llama, LoReFT on Qwen and Gemma).
  • T×I, BIG5 and Emotion. CCS is ahead in all 9 settings.

Table R3. Table 1 → fully held-out (nested cross-validation) scores.

ModelMethodT×IAxBBIG5Emo
LlamaReFT-r153.2 → 52.90.98 → 0.7275.1 → 74.826.5 → 25.6
LoReFT60.4 → 59.70.80 → 0.6881.9 → 81.426.9 → 25.2
BiPO74.4 → 74.40.49 → 0.3271.8 → 70.716.9 → 16.7
CCS80.3 → 78.20.78 → 0.5383.1 → 81.577.9 → 73.6
QwenReFT-r148.7 → 47.90.71 → 0.5069.8 → 68.826.3 → 22.3
LoReFT50.1 → 49.50.85 → 0.7373.2 → 77.439.8 → 36.8
BiPO52.6 → 51.30.26 → 0.1670.7 → 69.822.4 → 21.5
CCS76.8 → 75.30.86 → 0.5886.3 → 85.380.1 → 78.7
GemmaReFT-r178.4 → 77.20.73 → 0.5860.7 → 59.436.4 → 35.2
LoReFT64.6 → 63.30.87 → 0.7680.9 → 79.857.5 → 56.9
BiPO94.4 → 92.90.31 → 0.1976.0 → 74.837.4 → 35.9
CCS92.2 → 93.00.84 → 0.5582.4 → 80.887.3 → 86.6

Overall, the held-out results support the same conclusion as Table 1: CCS, which is closed-form and training-free, is competitive with gradient-trained methods and ahead of them outside AxBench. We have added these results to Appendix E of the revised paper, next to Table 7.

We hope this resolves the remaining point, and we thank the reviewer again for the constructive suggestions.

Reviewer_oDQiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work introduces a method that constructs an additive steering vector by projecting target-grand-mean contrast onto a subspace derived from target and non-target scatter matrices. The reported results suggest that CCS improves the achievable balance between target expression and generation quality relative to the evaluated activation-steering baselines.

Critical strengths
  1. The central idea makes sense, a steering vector can contain both the desired behaviour and incidental features of the examples used to construct it. CCS filters that vector through a subspace that emphasizes target-class structure relative to other class.
  2. I particularly appreciate the extension to multi-turn conversations, not many steering papers do this. Evaluating steering over 20 turns addresses an important practical challenge, and CCS’s stronger retention of concept expression and fluency is encouraging evidence of sustained effectiveness.
Critical weaknesses
  1. Although CCS explicitly uses non-target classes to suppress unwanted variation, the experiments primarily measure target expression, fluency, and instruction relevance. They do not systematically establish whether steering one behavior leaves other behaviors or nuisance attributes unchanged. Suppressing non-target activation scatter does not necessarily imply reduced behavioral interference. The authors should evaluate how each single-target intervention changes other behaviors, comparing CCS with strong baselines at matched target effectiveness. Expected invariances should be specified, since some behaviors naturally overlap or conflict. This would directly test whether CCS achieves more selective control, rather than simply stronger target expression. The planted-topic diagnostic is useful but insufficient to support general claims of eliminating confounds. A nuisance concentrated in the target demonstrations may itself appear as target-specific structure and survive the projection. Experiments that systematically vary target-nuisance correlations would strengthen the proposed explanation. Without such evidence, the results support improved steering performance more directly than deconfounding, and the claims should be narrowed accordingly.

  2. Appendix E states that steering strength is selected by an argmax and is not cross-validated, although the subspace parameters are. The authors should clarify the separation between parameter selection and final evaluation and report fully held-out results. Confidence intervals would also help assess the universal-win claim, since some differences are small.

Questions for authors
  1. Do you have existing measurements showing whether steering one behavior changes other behaviors or nuisance attributes, ideally comparing methods at matched target effectiveness? If not, what evidence supports the general deconfounding claim beyond the specific planted-topic diagnostic?

  2. Appendix E states that steering strength α\alpha is selected by an argmax rather than cross-validation. Are the examples used to select α,k\alpha, k, and rr separate from those used to report final performance? Please clarify the selection and reporting splits for each benchmark.

6 comments
AuthorsOpenReview ↗

Response to Reviewer oDQi (1/4)

W1 + Q1: Support for the deconfounding claim

We thank the reviewer for raising this point. We agree that reducing variation in non-target activations does not necessarily mean that steering will interfere less with other behaviors. To test this directly, we added an experiment measuring off-target interference, i.e. how much each steering method affects behaviors other than the intended target.

We also agree that the higher steering quality in Table 1 (effectiveness with little loss in fluency and prompt adherence) is consistent with, but does not by itself demonstrate, confound suppression. The controlled experiment in Section 3.2 was designed to test this link directly: when a nuisance topic is planted in the target's data, the unprojected direction absorbs it and loses steering quality, while the projection keeps it out. To strengthen this evidence, we include an expanded version of this experiment with all 50 AxBench concepts, controlling both how strongly the target's data is skewed toward the nuisance topic and how much of it the other classes contain (see part 2/4).

(1) Off-target interference. For each method, model and target behavior, we score the generated responses not only for the target, but also for the other classes of the benchmark: the remaining 6 emotions in Emotion, the other 4 personality traits in BIG5, and 10 genre-stratified non-target concepts in AxBench. We exclude TruthfulQA because it contains only two classes. We define off-target bleed as the mean absolute change in these scores relative to the unsteered model.

Since stronger steering naturally produces more off-target changes, we evaluate each method across steering strengths and report the normalised area under the Pareto frontier of target gain against bleed; a higher value means a method achieves more target improvement with less interference. The full curves are in Figure 9 (Appendix G of the revised paper).

Table R1. Normalised area under the bleed Pareto frontier (0–1, higher is better), averaged over the three models.

MethodEmotionBIG5AxBenchMean
CCS0.5550.7270.3180.533
PCA0.5150.5790.1720.422
Linear-AcT0.2870.6160.1600.354
Prompting0.1850.5310.2840.333
CAA0.3610.4900.0980.316
RepE0.3670.2740.0890.243
Angular0.0460.0770.0570.060
SEA0.0220.0300.0240.025

CCS achieves the best trade-off on all three benchmarks, providing direct behavioral evidence that CCS steers the intended target more selectively.

Expected changes in related behaviors. We also agree that not every change in a non-target score is undesirable interference: some behaviors are naturally related, and steering one may reasonably affect another. For example, in the Emotion dataset, steering towards happiness is expected to lower the negative emotions. In BIG5, following the higher-order factors of Digman (1997) and DeYoung et al. (2002), when we steer toward high neuroticism, the opposite pole of emotional stability, we expect it to move in the opposite direction to agreeableness and conscientiousness. For AxBench, we distinguish concepts of the same genre from those of different genres. In Table 19 (Appendix G) we report how related, opposing and unrelated classes move under each method. These relations are strongest in Emotion and BIG5, where related classes move with the target and opposing classes against it; on AxBench, same-genre concepts do not always move together, as concepts are more varied. Despite this semantic entanglement, CCS has the best selectivity of all steering methods, as shown above.

References

DeYoung, C. G., Peterson, J. B., and Higgins, D. M. Higher-order factors of the Big Five predict conformity: Are there neuroses of health? Personality and Individual Differences, 33(4):533–552, 2002.

Digman, J. M. Higher-order factors of the Big Five. Journal of Personality and Social Psychology, 73(6):1246–1256, 1997.

AuthorsOpenReview ↗

Response to Reviewer oDQi (2/4)

(2) Systematically varied target–nuisance correlation. The planted-topic experiment in Section 3.2 considered ten AxBench concepts under a fixed nuisance correlation. We now extend it to all 50 concepts and vary two factors: (i) the proportion ρ\rho of target examples drawn from a nuisance topic gg, and (ii) how much of that topic is also present in the non-target classes. We compare CCS with GMC (the same mean shift without the contrastive projection) and an oracle that explicitly removes the known topic direction ugu_g. We measure steering effectiveness (AxBench score) and the nuisance retained in the steering vector, ∣cos⁡(v,ug)∣|\cos(v, u_g)| (lower is better).

How important is the presence of the nuisance in other classes? We fix ρ=0.9\rho = 0.9 and keep all, half or none of the non-target classes' examples from topic gg, leaving their other examples unchanged.

Table R2.

Other concepts' gg examples keptScore GMCScore OracleScore CCS ↑Topic GMCTopic OracleTopic CCS ↓
No skew (natural data)0.450.500.500.410.000.02
all0.210.400.430.800.000.17
half0.220.400.360.820.000.27
none0.210.390.270.840.000.42

When the nuisance topic is also present in the non-target classes, CCS performs comparably to the topic-removal oracle (0.43 vs. 0.40) while greatly reducing the topic component of the steering vector (0.17 vs. 0.80 for GMC). As fewer non-target examples contain the nuisance, CCS becomes less effective and retains more of the topic direction. When the nuisance appears only in the target demonstrations, CCS can no longer reliably distinguish it from target-specific variation; it still remains less contaminated than GMC (0.42 vs. 0.84) and scores higher (0.27 vs. 0.21), but falls below the oracle (0.39).

How robust is CCS to stronger target–nuisance correlations? We vary ρ\rho from 0.3 to 1.0 while keeping all nuisance-topic examples in the non-target classes.

Table R3.

naturalρ=0.3\rho=0.30.50.70.91.0
Score: GMC0.450.350.280.230.210.19
Score: Oracle0.500.430.440.440.400.42
Score: CCS ↑0.500.490.470.440.430.41
Topic in vector: GMC0.410.500.620.740.800.82
Topic in vector: Oracle0.000.000.000.000.000.00
Topic in vector: CCS ↓0.020.020.070.120.170.19

As the correlation increases, GMC increasingly captures the nuisance direction: its topic component rises from 0.41 to 0.82 and its score drops from 0.45 to 0.19. CCS remains close to the oracle at every level. Even at ρ=1.0\rho = 1.0, where all target examples come from the nuisance topic, CCS scores 0.41 (oracle 0.42, GMC 0.19) while keeping its topic component below 0.2. For every ρ≥0.3\rho \geq 0.3, CCS outperforms GMC by 0.14–0.22 (95% confidence intervals exclude zero).

These experiments clarify both the effectiveness and the limitations of CCS: the contrastive projection substantially suppresses nuisance variation when that variation is represented in the competing classes, even under strong target–nuisance correlation, but becomes less effective when the nuisance is exclusive to the target class. As the reviewer requested, we have narrowed our claims in the abstract, introduction and conclusion accordingly: they now state that CCS suppresses variation that the target shares with the other classes, rather than that it eliminates confounds. We have included this expanded experiment in Section G.1 of the revised paper (Tables 20–21).

AuthorsOpenReview ↗

Response to Reviewer oDQi (3/4)

W2 + Q2: Hyperparameter selection and confidence intervals

To answer Q2 directly: in Table 1 they are not separate. The cross-validation folds are drawn from the evaluation prompts, so we additionally report fully held-out results with nested cross-validation (part 4/4), where CCS remains the best method in 11 of 12 settings.

All method hyperparameters in the paper (kk and rr for CCS, and the rank kk for PCA) are selected per target by 3-fold cross-validation, a standard way to choose hyperparameters, with a majority vote over the three folds so that each experiment reports a single hyperparameter setting (full table in Appendix E). The other training-free baselines have no hyperparameter besides the steering strength. This keeps our results reproducible and the method easy to use. Being able to tune its subspace makes CCS more flexible than the baselines, which can only change the steering strength for a given dataset. Even with a fixed (k,r)=(32,1)(k, r) = (32, 1) and no tuning, CCS is best in 10 of 12 settings (Table 1).

We do not treat the steering strength α\alpha as a hyperparameter of the method: in additive steering it is the knob that sets how much of the concept is added, and a user varies it at inference time. A good steering direction should therefore degrade quality little over a wide range of strengths, which is why Appendix K of the revised paper reports the full quality–effectiveness fronts of every method on every benchmark: CCS is ahead across strengths, not only at the selected point. For Table 1, α\alpha is the best strength on the evaluation prompts, chosen by the same rule and from the same grid for every method, so the comparison is like for like. Reporting the best strength of a sweep on the evaluation data is common practice in the steering literature, e.g. CAA (Rimsky et al., 2024), Linear-AcT (Rodriguez et al., 2025), ActAdd (Turner et al., 2023), mean-centring (Jorgensen et al., 2023), conceptor steering (Postmus & Abreu, 2024), latent guidance (von Rütte et al., 2024), SAE-targeted steering (Chalnev et al., 2024) and neuron-based personality induction (Deng et al., 2025).

Including α\alpha in the cross-validation does not change our conclusions. We recomputed all Table 1 scores with the strength chosen in the same 3-fold majority vote as the other hyperparameters (fluency and adherence gate applied as in Table 1). CCS remains the best method in all 12 settings, with margins similar to Table 1. Every method loses slightly, since the strength is no longer chosen on all prompts: AxBench drops by 0.02–0.09 and the other benchmarks by at most about 3 points. This table is now in Appendix E of the revised paper (Table 6).

Table R4. Table 1 with α\alpha in the cross-validation vote (L = Llama-3.1-8B, Q = Qwen3-8B, G = Gemma-2-9B; T*I = TruthfulQA, AxB = AxBench, Emo = Emotion).

MethodL T*IL AxBL BIG5L EmoQ T*IQ AxBQ BIG5Q EmoG T*IG AxBG BIG5G Emo
CAA72.40.2775.258.663.40.3777.158.784.80.3976.878.3
RepE56.10.2568.756.451.10.2971.759.679.40.3675.876.9
PCA54.70.4379.464.751.60.4481.262.585.70.4772.777.4
SEA53.10.1967.017.051.90.1968.020.179.60.2061.524.0
Linear-AcT58.50.3182.949.857.20.5281.650.881.60.4376.451.6
Angular56.30.2067.517.454.60.2170.021.381.60.1962.626.5
CCS79.10.7382.976.676.80.8386.380.192.20.8082.486.9

References

Chalnev, S., Siu, M., and Conmy, A. Improving steering vectors by targeting sparse autoencoder features. arXiv:2411.02193, 2024.

Deng, J., et al. Neuron-based personality trait induction in large language models. ICLR, 2025.

Jorgensen, O., et al. Improving activation steering in language models with mean-centring. arXiv:2312.03813, 2023.

Postmus, J. and Abreu, S. Steering large language models using conceptors: Improving addition-based activation engineering. NeurIPS Workshop on Foundation Model Interventions, 2024.

Rimsky, N., et al. Steering Llama 2 via contrastive activation addition. ACL, 2024.

Rodriguez, P., et al. Controlling language and diffusion models by transporting activations. ICLR, 2025.

Turner, A. M., et al. Steering language models with activation engineering. arXiv:2308.10248, 2023.

von Rütte, D., et al. A language model's guide through latent space. ICML, 2024.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Rebuttal Summary

We thank the reviewers for their insightful comments, which helped us significantly improve the paper.

Strengths pointed out by the reviewers

  • Well-motivated method. oDQi: "The central idea makes sense"; p9ag: "The entanglement account of why steering vectors hurt fluency is intuitive, and the paper consistently ties design choices back to this story."
  • Strong results. q1RH: "The reported results are strong, with CCS outperforming the activation steering baselines across various benchmarks."; oDQi: "CCS improves the achievable balance between target expression and generation quality".
  • Broad evaluation. q1RH: "a comprehensive empirical evaluation across three models, four benchmarks, multiple baselines, and a diverse range of concepts"; p9ag: "Solid experiments and evaluation", "the human evaluation is more rigorous than most steering papers".
  • Informative diagnostics. oDQi: "The planted-topic diagnostic is useful"; q1RH: "The GMC ablation on Axbench helps isolate the contribution of the projection".
  • Multi-turn steering. oDQi: "addresses an important practical challenge"; p9ag: "a genuinely useful contribution".
  • Clear writing. oDQi and q1RH rated the presentation "Clear and easy to follow"; p9ag: "The paper is well structured".

New experiments

  • Off-target interference (oDQi). We measure how steering one target changes every other class across full strength sweeps, with expected invariances specified in advance. CCS has the best selectivity trade-off on all three benchmarks.
  • Varied target–nuisance correlation (oDQi). We extend the planted-topic diagnostic to all 50 AxBench concepts, varying the topic skew and the topic's presence in the other classes. When the topic is shared with the other classes, CCS suppresses it and stays close to the oracle at every skew level; when it appears only in the target's data, CCS degrades and falls below the oracle. We narrowed the paper's claims accordingly and added this as a limitation.
  • Held-out evaluation (oDQi). We add fully nested cross-validation: CCS is best in 11 of 12 settings and significantly better than the baseline in 65 of 72 comparisons.
  • Confidence intervals (q1RH, oDQi). 95% bootstrap intervals for every cell of Table 1, including the variance of the selection, with paired tests: CCS is higher in all 72 comparisons, significantly in 64.
  • Gradient-trained baselines (q1RH). We add ReFT-r1, LoReFT and BiPO on all models and benchmarks. CCS, which is closed-form and training-free, outperforms all three in 9 of 12 settings.
  • Capability and safety (p9ag). We evaluate MMLU+ARC, HarmBench and XSTest when steering towards AxBench concepts. Strong steering reduces accuracy for all additive methods (18.6 points for CCS at its Table 1 strengths), but at matched concept expression CCS loses less than every baseline (e.g. 4.4 vs. 18.9 points for CAA) and is the most effective method under 2- and 5-point accuracy budgets. CCS shows no systematically larger safety side effects.
  • Reasoning models (p9ag). We steer Qwen3-8B with thinking enabled. CCS remains the strongest method in every setting, and steering only the final answer largely preserves capabilities: 0.66 effectiveness at a 1.3-point accuracy loss, vs. 0.58 at a 13.0-point loss without thinking.

Clarifications and presentation. We clarified the nuisance classes and the data available to each method (q1RH), the selection protocol (oDQi) and the choice of layer (p9ag). Following p9ag, we define H~c\tilde{H}_c and UU before use, unified the nuisance terminology, aligned Figure 2 with the steering rule we use, and added legends to Figures 1 and 2.

As the review process is coming to an end, we would like to thank all reviewers again for their effort and insightful comments, which helped us improve our paper. During the rebuttal we significantly expanded the experiments in the paper and addressed all of the reviewers’ concerns.

Two reviewers raised their scores to weak acceptance, and the third stated that the rebuttal strengthens the paper and wouldn't mind if the paper got accepted. Specifically:

  • Reviewer q1RH wrote that “the additional baselines and experiments substantially address my original concerns and strengthen the paper” and raised their score from 2 to 3.
  • Reviewer p9ag wrote that “the additional results have addressed my concerns” and raised their score from 2 to 3.
  • Reviewer oDQi wrote that “the held-out evaluation substantially addresses my concern about parameter selection, and the expanded nuisance study clarifies the method’s limitations”, that “the rebuttal strengthens the paper”, and that they “would not mind if this work get accepted”. They raised two remaining points, both of which we have addressed in our follow-up reply; the reviewer has not yet responded to it.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.