acceptodds
Under review as a conference paper at ICLR 2027

Pathway-Selective Activation Steering: LLM Control Through Normalized Attention

Abstract

Activation steering can improve a language model’s target behavior, but increasing steering strength can severely degrade language modeling. We study how this trade-off depends on where the intervention enters the computation. In pre-normalized Transform- ers, residual-stream addition introduces a direct carry that grows with steering strength, even though the normalized branch outputs remain bounded. This distinction motivates Pathway-Selective Activation Steering (PSAS), which routes a normalized auxiliary steered state through the query and value pathways while preserving the original keys and identity path. We prove coefficient-independent bounds on its attention-output and full-layer perturbations and characterize its large-strength limit. Across three models, PSAS achieves the highest mean joint truthfulness–informativeness score over six steering methods. On two diagnostic models, removing CAA’s direct carry reduces its excess Wikipedia perplexity by 79–89%, while PSAS exhibits substantially lower sensitivity to coefficient overshoot. Pathway ablations show that Q/V steering offers a favorable trade-off between behavioral control and generation quality. These results show how intervention placement controls perturbation growth and motivate selecting pathway composition and steering strength jointly.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Activation Steering, Representation Engineering, Inference-Time Intervention, Model Alignment & Safety

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_MmzuOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how the placement of activation steering affects model behavior and language quality. It proposes PSAS, which steers the query and value inputs through an auxiliary normalized state while leaving the keys and residual connection unchanged. The analysis shows that this removes CAA’s unbounded residual addition. Experiments on three models report higher mean truthfulness–informativeness scores, while diagnostics on two models show lower perplexity sensitivity to steering strength. Ablations examine the effects of removing the residual addition and steering different pathways.

Critical strengths
  1. The paper gives a clear explanation of why strong residual steering can degrade model behavior. It separates the direct residual addition from changes inside normalized branches, then tests this distinction by removing the residual addition. This connects the analysis to a concrete source of language-modeling cost.

  2. The experiments examine where the gains come from. Carry-removal, pathway, and norm-constrained baselines test alternatives to the proposed design. These comparisons provide useful evidence that intervention placement matters beyond simply limiting update size.

Critical weaknesses
  1. The practical gain over existing Q/V steering is unclear. The main change is to add the direction before normalization. The boundedness analysis is useful, but the simplified DISCO-QV diagnostic does not show whether this change improves the control–quality trade-off at comparable behavioral control.

  2. The experiments offer limited guidance on when Q+V is preferable. The pathway ablations show that V-only or Q+K+V can achieve higher task scores under some settings, with different language-modeling costs. Without a fuller trade-off comparison, it remains unclear when the proposed Q+V configuration should be preferred.

Questions for authors
  • What practical benefit comes from adding the steering direction before normalization? At similar levels of behavioral control, does PSAS preserve generation quality better than DISCO with head selection?

  • Does PSAS offer a wider range of steering strengths that maintain both target behavior and informative responses? Bounded perturbations and slower perplexity growth alone do not show whether effective steering is easier to achieve.

2 comments
AuthorsOpenReview ↗

Reply to Reviewer Mmzu (1)

We appreciate the reviewer's questions about the practical benefits of pre-normalization steering and the choice of attention pathways.

What practical benefit comes from adding the steering direction before normalization? At similar levels of behavioral control, does PSAS preserve generation quality better than DISCO with head selection?

In Table 1 (Section 6.1), PSAS has higher mean joint truthfulness–informativeness and lower mean toxicity than head-selected DISCO-QV on all three models:

ModelPSAS T×IT\times IDISCO-QV T×IT\times IPSAS toxicityDISCO-QV toxicity
Qwen3-1.7B0.549 ± 0.0150.512 ± 0.0140.027 ± 0.0020.047 ± 0.005
Llama-3.1-8B0.666 ± 0.0110.617 ± 0.0100.115 ± 0.0100.166 ± 0.006
Qwen3-14B0.860 ± 0.0060.823 ± 0.0110.095 ± 0.0020.134 ± 0.008

The ± values are standard deviations across five decoding seeds. Detoxification generation perplexity is higher for PSAS on the two smaller models and lower on Qwen3-14B.

At the same joint score of 0.468 on Qwen3-1.7B, PSAS has approximately 37% lower cost than simplified DISCO-QV without head selection and 54% lower cost than a control that adds steering after RMSNorm. The latter uses a direction estimated from the same training examples as PSAS and steers only Q/V. Here, cost is the increase in cross-entropy over the unsteered model on fixed Wikipedia text. The comparison interpolates between evaluated settings, and paired 95% confidence intervals support both reductions. Appendix J.1 reports how settings are chosen on validation data and evaluated on separate test data.

The controls that steer after normalization have lower costs at the evaluated moderate joint scores on Llama. We also tuned DISCO-QV with steering restricted to its selected heads. The confidence intervals do not resolve a Qwen cost difference over the joint scores PSAS achieves; DISCO-QV reaches a higher maximum score and has lower costs on Llama. Appendix K shows the curves relating behavioral scores to Wikipedia perplexity.

The experiments offer limited guidance on when Q+V is preferable. The pathway ablations show that V-only or Q+K+V can achieve higher task scores under some settings, with different language-modeling costs. Without a fuller trade-off comparison, it remains unclear when the proposed Q+V configuration should be preferred.

Our pathway ablations suggest that Q+V provides a favorable balance between behavioral control and language-modeling quality, although its advantages depend on the task and model.

On TruthfulQA, Q+V improves the joint truthfulness–informativeness score over V-only on Qwen (0.526 -> 0.548) at nearly identical Wikipedia excess cross-entropy (0.043 -> 0.044). On Llama, V-only achieves a higher score (0.707 vs. 0.663), but Q+V incurs substantially less Wikipedia cross-entropy degradation (0.063 vs. 0.117). Q+V also has lower degradation than Q+K+V on both models at their selected settings.

On detoxification, adding query steering consistently improves toxicity reduction on both models: from 0.093 to 0.026 on Qwen and from 0.305 to 0.105 on Llama, at the expense of higher generation perplexity.

These results suggest that Q+V is a useful choice when balancing target behavioral control against language-modeling degradation, particularly for detoxification. We do not claim it is universally optimal; the preferred pathway depends on the model, task, and operating point (Section 6.3 and Appendix G).Adding query steering to value steering improves detoxification on both tested models. With one decoding seed and steering strength held fixed, toxicity falls from 0.093 to 0.026 on Qwen and from 0.305 to 0.105 on Llama, with higher generation perplexity.

Table 22 compares settings with strength selected separately for each pathway and results averaged over five decoding seeds. On Qwen, Q+V raises the mean joint score from 0.526 to 0.548 while the increase in Wikipedia cross-entropy changes only from 0.043 to 0.044; adding key steering gives a slightly lower score at higher cost. On Llama, Q+V gives a lower-cost setting: a joint score of 0.663 at cost 0.063, compared with V-only's 0.707 at cost 0.117 and Q+K+V's 0.698 at cost 0.109. Section 6.3 and Appendix G discuss these selected settings.

AuthorsOpenReview ↗

Reply to Reviewer Mmzu (2)

Does PSAS offer a wider range of steering strengths that maintain both target behavior and informative responses? Bounded perturbations and slower perplexity growth alone do not show whether effective steering is easier to achieve.

To address this, we added Figure 10 in Appendix K, which compares behavioral control against Wikipedia perplexity across steering strengths for all three models on both TruthfulQA and detoxification.

The curves show that PSAS achieves favorable control–perplexity trade-offs across multiple evaluated settings, particularly for detoxification, where it remains competitive near the Pareto frontier on all three models. On TruthfulQA, PSAS also achieves strong joint truthfulness–informativeness scores with relatively low Wikipedia perplexity, although its advantage over other methods depends on the model and operating region.

Importantly, TruthfulQA is evaluated using the joint truthfulness–informativeness score, so improvements reflect responses that are both truthful and informative rather than simply more conservative. Appendix G.2 further examines the individual informativeness scores across PSAS steering strengths.

Together, these results provide evidence that PSAS can maintain useful behavioral control without substantial off-target language-modeling degradation across several operating settings. However, the curves do not establish that PSAS has a universally wider effective coefficient range than DISCO-QV. We have clarified this distinction in Section 6.4.

Reviewer_WCnLOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the effect of the steering coefficient in Large Language Model steering using mean difference vectors. They find that adding a steering vector to the residual stream can have unbounded effects in later outputs and show that intervening right before the normalization layer, rather than in the residual stream, can prevent this and propose to make interventions there.

Critical strengths

They identify a major drawback of steering the LLM residual stream and provide a simple and easy solution.

They provide theoretical results on the growth of the intervention vector based on the pathway and intervention coefficient.

Critical weaknesses

The novelty of the proposed method is small, as there is very little difference from the cited method, DISCO.

Many of the terms in the equations are undefined or redundant, making the understanding of the paper difficult at times.

Due to the previous point, I am uncertain, but it sounds like the boundedness results only apply to the output at that layer? If so, this calls into question the usefulness of the results, as the output of one layer is the input to the next and thus could still exhibit unbounded growth.

Questions for authors

Can the authors clarify the third point in the weaknesses section?

3 comments
AuthorsOpenReview ↗

Reply to Reviewer WCnL

We thank the reviewer for raising the question about downstream boundedness and for pointing out areas where the presentation could be clearer.

Due to the previous point, I am uncertain, but it sounds like the boundedness results only apply to the output at that layer? If so, this calls into question the usefulness of the results, as the output of one layer is the input to the next and thus could still exhibit unbounded growth.

The perturbation remains bounded through all later layers as steering strength increases. For a single PSAS intervention on fixed input tokens in a finite-depth model satisfying Assumption 1, each later block adds bounded changes in attention and in the MLP response to the altered attention output. Their sum gives a finite bound determined by depth and weights, independent of steering strength. Section 4.2 and Appendix B.3 state and prove the result. The bound need not be small, so Section 6.4 and Appendix G.2 evaluate behavioral quality separately.

The novelty of the proposed method is small, as there is very little difference from the cited method, DISCO.

PSAS was motivated by the direct residual addition that grows with steering strength. It normalizes a steered copy of the hidden state and uses it to compute queries and values, while retaining the original state on the residual connection. The resulting updates vary with the token and level off as strength increases. The contribution is this construction and the bounds on its effects; Q/V steering is shared with earlier work (Sections 1 and 3.3, Appendix I).

When each method's selected steering strength is multiplied by eight, the increase in Wikipedia cross-entropy over the unsteered model grows approximately 10–11-fold for PSAS, versus 52–166-fold for Q/V steering applied after projection without head selection (Appendix J.1). This measures sensitivity to increasing strength; comparisons at equal behavioral scores are reported separately.

Many of the terms in the equations are undefined or redundant, making the understanding of the paper difficult at times.

Section 3.2 now defines the original residual state, the normalized copy used for steering, the change in the full attention output, and the change in the residual state. The last of these includes both attention and MLP changes; CAA additionally includes the direct steering addition. We distinguish full attention outputs from individual head contributions and state when layer or token indices are omitted.

I would like to thank the authors for their response. While I do see the merit in pursuing boundedness and stability in steering, I still believe the contribution of the paper is marginal in its current form.

The given bound can indeed be applied to every layer to compose a bound on the entire model, but seeing as each bound relies on the norm of the weights, and the effects would be multiplied for each layer, I don't see a significant practical or theoretical benefit. Perhaps if, even using additional assumptions, a more constrained bound were to be derived that would provide additional insights, I could be convinced. However, as it stands, it seems like the bound is simply saying that normalization regulates the rate at which model hidden states and outputs grow, which is not a surprising result; rather, it is the intent.

The authors mention measuring the actual effects empirically and finding their method improves over existing baselines, which is great; however, considering the limited algorithmic novelty of the method, I believe there needs to be more of a contribution for this paper to be published. Thus, I will respectfully maintain my score.

Thank you for suggesting this refinement.

The given bound can indeed be applied to every layer to compose a bound on the entire model, but seeing as each bound relies on the norm of the weights, and the effects would be multiplied for each layer, I don't see a significant practical or theoretical benefit.

The downstream bound in Section 4.2 and Appendix B.3 is additive across layers. Fix the input tokens and a finite-depth model satisfying Assumption 1, with a single PSAS intervention at layer ll. Write ok,io_{k,i} and mk,im_{k,i} for the complete attention and MLP outputs at token ii in layer kk. Both trajectories satisfy

\qquad \|m_{k,i}\|_2,\ \|\widetilde m_{k,i}\\|_2\leq M_k.$$ These constants depend on the layer's weights and normalization gains, but not on steering strength or the incoming residual state. RMSNorm bounds the branch inputs, attention mixes bounded values, and the continuous MLP is bounded on its normalized-input ball. Write $C_k=A_k+M_k$. Let $\Delta h_{k,i}=\widetilde h_{k,i}-h_{k,i}$. PSAS preserves the input identity path, so $\Delta h_{l,i}=0$. The residual updates give $$\begin{aligned} \Delta h_{k+1,i} &=\Delta h_{k,i}+\Delta o_{k,i}+\Delta m_{k,i},\\ \|\Delta h_{k+1,i}\|_2 &\leq\|\Delta h_{k,i}\|_2+2A_k+2M_k. \end{aligned}$$ Iterating this inequality yields $$\|\Delta h_{s,i}\|_2 \leq 2\sum_{k=l}^{s-1}C_k,\qquad s>l.$$ Weight norms enter the constants within each layer; this estimate does not multiply layerwise Lipschitz constants. It bounds the total change independently of steering strength, without excluding local amplification. > However, as it stands, it seems like the bound is simply saying that normalization regulates the rate at which model hidden states and outputs grow, which is not a surprising result; rather, it is the intent. Normalization bounds the attention and MLP computations in both PSAS and CAA, but does not bound an addition carried directly through the identity path. At a token receiving CAA's addition $\tau r$, the same argument gives $$\Delta h^{\mathrm{CAA}}_{s,i}=\tau r+E_{s,i}(\tau), \qquad \|E_{s,i}(\tau)\|_2\leq 2\sum_{k=l}^{s-1}C_k.$$ Consequently, $$\|\Delta h^{\mathrm{CAA}}_{s,i}\|_2 \geq |\tau|\,\|r\|_2-2\sum_{k=l}^{s-1}C_k.$$ For a fixed downstream layer, CAA's residual perturbation therefore grows without bound as $|\tau|$ increases, whereas PSAS's remains bounded. This is the role of preserving the original identity path: normalization bounds the branch computations but cannot remove the direct carry. > Perhaps if, even using additional assumptions, a more constrained bound were to be derived that would provide additional insights, I could be convinced. We have added a Lipschitz assumption on the relevant state regions to track the size of the initial perturbation as well as the absolute branch bounds. The refinement is described in Section 4.2 and proved in Appendix B.3. Let $H$ stack all token states and use the whole-context norm $\|H\|_{\infty,2}=\max_i\|h_i\|_2$, which includes attention's dependence on other tokens. At an unchanged downstream block, write $F_k(H)=H+R_k(H)$, where $R_k$ is the complete attention and MLP update and $\|R_k(H)\|_{\infty,2}\leq C_k$. For a coefficient range $\mathcal T$ containing zero, assume each region $\mathcal D_k$ lies within the block's domain and contains both trajectories for all $\tau\in\mathcal T$. The additional assumption is that finite constants $L_k\geq0$, uniform over this range, satisfy $$\|R_k(H)-R_k(H')\|_{\infty,2} \leq L_k\|H-H'\|_{\infty,2}, \qquad H,H'\in\mathcal D_k,\quad k\geq l+1.$$ Define $e_k(\tau)=\|\widetilde H_k-H_k\|_{\infty,2}$ and initialize with the actual intervention-layer difference, $$e_{l+1}(\tau)=\delta_l(\tau) =\|\widetilde H_{l+1}-H_{l+1}\|_{\infty,2} \leq 2C_l.$$ The residual recursion, Lipschitz assumption, and absolute branch bounds give $$\begin{aligned} e_{k+1}(\tau) &\leq e_k(\tau)+\|R_k(\widetilde H_k)-R_k(H_k)\|_{\infty,2}\\ &\leq e_k(\tau)+\min\{L_ke_k(\tau),\,2C_k\}, \qquad k\geq l+1. \end{aligned}$$ This uses the Lipschitz estimate for smaller perturbations and caps each additional change by the existing bound. Iterating yields \$ e\_s(\tau) \leq min \left\\{\delta\_l(\tau)\prod\_{k=l+1}\^{s-1}(1+L\_k),\delta\_l(\tau)+2\sum\_{k=l+1}\^{s-1}C\_k\right\\}, s>l. \$ Unlike the coarse sum alone, this refinement gives a two-regime bound. For small perturbations, the bound scales proportionally with the initial intervention $\delta_l(\tau)$ and tends to zero as $\tau\to0$. When the Lipschitz estimate becomes large, the additional change at each layer is capped by the normalization-induced ceiling $2C_k$. The refinement therefore captures sensitivity to small interventions while preserving bounded downstream perturbations as steering strength increases.
Reviewer_u1pPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that CAA's residual addition adds a "carry" term that grows linearly with steering strength, while normalized branch outputs stay bounded. It proposes PSAS: normalize h + τr on a side branch and feed it only to Q and V, keeping the original keys and identity path. It proves coefficient-independent perturbation bounds and evaluates on TruthfulQA and RealToxicityPrompts (Qwen3-1.7B, Llama-3.1-8B, Qwen3-14B) against six baselines.

Critical strengths
  • Clear carry vs. bounded-branch decomposition; the carry-scaling ablation (Eq. 2, Table 2) isolates it cleanly.
  • Analysis covers RoPE/QK-Norm; theory analysis of bounds that arise thanks to their pre-norm intervention; extensive numerical checks of the theory.
  • Simple, cheap method.
Critical weaknesses

1. Limited novelty vs. DISCO-QV and Communication Steering (Torop et al., NeurIPS 2025).

  • Communication Steering is a baseline introduced and reported in the DISCO paper. It adds a vector after RMSNorm (to Q, K, V). DISCO-QV adds per-head vectors to the projected Q and V. PSAS adds τr before RMSNorm (to Q, V). Communication Steering is never evaluated, so the core design choice (pre- vs post-norm) is not isolated.
  • None of the three has CAA's carry, so carry removal is not specific to PSAS.
  • DISCO (Prop. 2) shows that Communication Steering is equivalent to DISCO-QV with tied strengths. Table 18 shows the PSAS shift is nearly collinear with g⊙r (cos ≈ 0.975), so PSAS ≈ Q/V-only Communication Steering plus a per-token 1/‖h‖ rescaling.
  • The genuinely new parts are token-dependent magnitude and saturation. Untouched keys are a minor point; no head selection and tied strengths are simplifications.

2. Small, inconsistent gains.

  • About +0.04 T×I over DISCO-QV. Within noise of CAA on Llama (0.666 vs 0.657) and of PID on Qwen3-1.7B (0.549 vs 0.537).
  • DISCO-QV has lower generation PPL on 2 of 3 models, and lower Wikipedia PPL at every strength multiple (Table 26).
  • On Llama, V-only beats Q+V (Tables 4, 22).

3. Statistical reliability.

  • From appendix C percentages reported, we should have about 408 held-out questions; the ± covers decoding seeds only. Question-level SE ≈ 0.025, i.e. 2–4× the reported spreads.
  • There are only 80 validation questions following these percentages; baseline tuning budgets are not reported.
  • Tables 28/29 duplicate each other, don't state the model, and the 0.18 constant-replacement score is below unsteered.

4. Narrow evaluation.

  • Behaviors. Only truthfulness and toxicity are tested, although the claim concerns steering in general and gains are already behavior-dependent (Q+V helps toxicity more than TruthfulQA; V-only beats Q+V on Llama). Suggested additions:
    • refusal on harmful prompts, e.g. HarmBench [1], jointly with over-refusal on benign prompts, e.g. XSTest [2];
    • persona/alignment behaviors from Model-Written Evaluations [3], as used by CAA [4] and DISCO [5], enabling direct comparison with DISCO's reported results;
    • out-of-distribution transfer of the TruthfulQA direction, as done in ITI [6], e.g. to TriviaQA [7] or HaluEval [8];
    • AxBench [9], which scores concept steering against fluency and instruction following across steering factors;
    • on RealToxicityPrompts, also evaluate non-toxic prompts, where any change is an unintended effect.
  • Trade-off curves. The control-vs-cost curve of Fig. 3 is given only for Qwen3-1.7B on TruthfulQA; please extend it to Llama-3.1-8B, Qwen3-14B, and toxicity, since comparisons at "k× each method's operating coefficient" mix different levels of control.
  • Architectures. All models are pre-norm RMSNorm Transformers. Since the analysis relies on this structure, a model with additional post-sublayer normalization, e.g. Gemma 2 [10] or OLMo 2 [11], would test the scope of the claim.

References

[1] Mazeika et al., HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal, ICML, 2024

[2] Röttger et al., XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models, NAACL, 2024

[3] Perez et al., Discovering Language Model Behaviors with Model-Written Evaluations, Findings of ACL, 2023

[4] Rimsky et al., Steering Llama 2 via Contrastive Activation Addition, ACL, 2024

[5] Torop et al., DISCO: Disentangled Communication Steering for Large Language Models, NeurIPS, 2025

[6] Li et al., Inference-Time Intervention: Eliciting Truthful Answers from a Language Model, NeurIPS, 2023

[7] Joshi et al., TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension, ACL, 2017

[8] Li et al., HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models, EMNLP, 2023

[9] Wu et al., AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders, ICML, 2025

[10] Gemma Team et al., Gemma 2: Improving Open Language Models at a Practical Size, arXiv, 2024

[11] OLMo Team et al., 2 OLMo 2 Furious, arXiv, 2025

Questions for authors

Appendix G states that Table 1 uses last-position CAA, whereas the pathway sweeps use all-token CAA, and PSAS is defined as full-context steering (Sec. 3.2). Please clarify which positions each method (PSAS and all baselines) steers in Table 1. If coverage differs, pathway and position coverage are confounded; an all-token CAA row in Table 1 would resolve this.

3 comments
AuthorsOpenReview ↗

Reply to Reviewer u1pP (1)

We appreciate the reviewer's detailed feedback, particularly regarding the novelty and empirical evaluation of PSAS.

1. Novelty and normalization placement

Limited novelty vs. DISCO-QV and Communication Steering (Torop et al., NeurIPS 2025).

PSAS follows from our analysis of the direct steering addition to the residual stream. It applies steering to a separate copy of the hidden state before RMSNorm, then uses the normalized copy to compute queries and values. The original state supplies the keys and residual connection. Adding steering before normalization makes the update depend on each token's state and remain bounded as strength increases. This is the contribution clarified in Sections 1 and 3.3; Q/V steering itself is shared with prior work.

Table 18 shows nearly parallel update directions. Even with similar directions, PSAS's update varies in size across tokens and levels off as steering strength increases (Appendix I).

Communication Steering is never evaluated, so the core design choice (pre- vs post-norm) is not isolated.

We added a control whose direction is estimated from the same training examples and added after RMSNorm, before computing queries and values. It preserves the original keys and residual connection (Appendix J.1). At the same Qwen3-1.7B joint truthfulness–informativeness score of 0.468, PSAS has approximately 54% lower cost than this control and 37% lower cost than simplified DISCO-QV, which applies Q/V shifts without head selection. Here, cost is the increase in cross-entropy over the unsteered model on fixed Wikipedia text. The comparison uses interpolation between evaluated settings, and paired 95% confidence intervals support both reductions. The controls that steer after normalization have lower costs at the evaluated moderate joint scores on Llama.

2. Empirical gains and pathway choice

Small, inconsistent gains.

About +0.04 T×I over DISCO-QV. Within noise of CAA on Llama (0.666 vs 0.657) and of PID on Qwen3-1.7B (0.549 vs 0.537). DISCO-QV has lower generation PPL on 2 of 3 models, and lower Wikipedia PPL at every strength multiple (Table 26). On Llama, V-only beats Q+V (Tables 4, 22).

Table 1 in Section 6.1 shows higher mean joint truthfulness–informativeness and lower mean toxicity for PSAS than DISCO-QV on all three models, with joint-score gains of 3.7–4.9 percentage points. Detoxification generation perplexity is higher on the two smaller models and lower on Qwen3-14B. Table 26 varies each method's strength relative to its own selected setting; the added comparison above instead evaluates cost at the same behavioral score.

Overall, Q+V provides a favorable balance between behavioral control and language-modeling quality across the evaluated settings. It improves detoxification on both models and offers competitive TruthfulQA performance, although the optimal pathway configuration remains model- and task-dependent (Section 6.3).For pathway choice, Q+V has lower Wikipedia cost on Llama, while V-only reaches a higher TruthfulQA score at its selected setting (Section 6.3).

Appendix G states that Table 1 uses last-position CAA, whereas the pathway sweeps use all-token CAA, and PSAS is defined as full-context steering (Sec. 3.2). Please clarify which positions each method (PSAS and all baselines) steers in Table 1. If coverage differs, pathway and position coverage are confounded; an all-token CAA row in Table 1 would resolve this.

The added comparison at matched behavioral scores applies every method at all prompt and generated-token positions (Appendix J.1). Appendix C clarifies the distinction between last-position CAA in Table 1 and all-position CAA in the pathway study in Appendix G. Appendix K also identifies the all-position CAA variant in the Qwen3-14B panel.

AuthorsOpenReview ↗

Reply to Reviewer u1pP (2)

3. Statistical reliability and reporting

From appendix C percentages reported, we should have about 408 held-out questions; the ±\pm covers decoding seeds only. Question-level SE ≈\approx 0.025, i.e. 2–4× the reported spreads. There are only 80 validation questions following these percentages; baseline tuning budgets are not reported.

The added study evaluates all 817 TruthfulQA questions across two test sets, with each question tested once using five decoding seeds. It also evaluates 500 held-out Wikipedia sequences. Separate training, validation, and test questions are used to estimate directions, select settings, and measure performance. Confidence intervals are computed by resampling questions and Wikipedia sequences, with methods compared on the same resampled items. Appendix J.1 reports the tuning budgets and procedure. Table 1's ±\pm values describe variation across decoding seeds; the added study accounts for variation across questions.

Tables 28/29 duplicate each other, don't state the model, and the 0.18 constant-replacement score is below unsteered.

We removed the duplicate table and retained the complete experiment as Table 28 in Appendix I.3. On Llama-3.1-8B, we replaced PSAS's token-specific steering adjustments to the attention value vectors with either one average adjustment for all tokens or randomly reassigned adjustments. Under Q+V steering, using the average lowers the joint score from 0.865 to 0.180. This shows that the original adjustments matter in this setting. Averaging also changes their sizes, and reassignment changes how they align with each token's state, so the experiment does not isolate token dependence alone.

AuthorsOpenReview ↗

Reply to Reviewer u1pP (3)

4. Evaluation scope

Behaviors. Only truthfulness and toxicity are tested, although the claim concerns steering in general and gains are already behavior-dependent (Q+V helps toxicity more than TruthfulQA; V-only beats Q+V on Llama).

PSAS improves refusal of harmful requests while answering most safe prompts on Qwen3-1.7B. Requiring at least 90% of XSTest safe prompts to be answered gives 44% refusal of harmful prompts for the best evaluated PSAS setting, compared with 30% unsteered, 39% for ITI, and 37% for Linear-AcT. Appendix J.3 reports the HarmBench/XSTest evaluation. Refusals are identified by matching refusal phrases, and the tested Llama settings show limited improvement.

On 1,199 non-toxic prompts, PSAS's continuations are most similar in content to the unsteered continuations among tested settings that reduce toxicity below 0.05 on toxic prompts. With detoxification settings unchanged, similarity is 0.370 for PSAS versus 0.312 for DISCO-QV. For comparison, unsteered continuations generated for the same prompts with different seeds have similarity 0.408. Mean toxicity on non-toxic prompts does not increase. DISCO-QV has a smaller generation-perplexity increase (+2.6 versus +4.4). Appendix J.4 reports both content similarity and generation perplexity.

The TruthfulQA settings transfer to HaluEval on Qwen3-1.7B without retuning. Across 10,000 QA items, PSAS improves accuracy in choosing the correct answer over the hallucinated answer from 84.6% to 86.4%. Accuracy improves for directions learned from both TruthfulQA training splits. On the 241 items where the correct answer is longer, the gains are 4.1 and 2.1 percentage points for the two directions, so the benefit is not confined to shorter correct answers. CAA, ODESteer, and DISCO-QV score higher overall. We score the provided answers by their summed conditional token log-probabilities, with the knowledge passage withheld, and average accuracy over both directions on the same items (Appendix J.5). Model-Written Evaluations and AxBench remain outside the current evaluation.

Trade-off curves. The control-vs-cost curve of Fig. 3 is given only for Qwen3-1.7B on TruthfulQA; please extend it to Llama-3.1-8B, Qwen3-14B, and toxicity, since comparisons at "k× each method's operating coefficient" mix different levels of control.

Appendix K now contains six panels covering all three models on TruthfulQA and detoxification. They plot the joint truthfulness–informativeness score or toxicity against Wikipedia perplexity across steering strengths, with the benchmark settings marked. This allows comparison at similar behavioral scores.

Architectures. All models are pre-norm RMSNorm Transformers. Since the analysis relies on this structure, a model with additional post-sublayer normalization, e.g. Gemma 2 [10] or OLMo 2 [11], would test the scope of the claim.

We added Gemma-2-9B-IT, which includes additional normalization after its sublayers. PSAS has the highest mean joint TruthfulQA score among the evaluated methods, 0.856 versus 0.843 for DISCO-QV. Mean toxicity is 0.096 versus 0.154, approximately 38% lower, and PSAS has lower toxicity than every evaluated baseline in all five decoding seeds. The toxicity ranking is unchanged after excluding seed 42, which was used to select settings.

Appendix J.2 reports the selection procedure and five-seed results, including slightly higher generation perplexity for PSAS than DISCO-QV. The theoretical claims retain the architectural assumptions stated in Section 4.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to Reviewers

We thank the reviewers for their comments on the contribution, comparisons, and analysis. We have revised the manuscript to address these concerns, with changes marked in blue.

Our motivation comes from a practical limitation of activation steering. Increasing the steering coefficient can strengthen the desired behavior, but it can also substantially degrade the model's original capabilities, including language modeling quality and the ability to produce informative responses. We investigated this trade-off and found that conventional additive steering introduces a direct perturbation into the residual stream that grows with the steering coefficient. This observation motivated us to explore whether behavioral control could be achieved without directly modifying the residual state.

PSAS addresses this issue by applying steering before RMSNorm to a separate copy of the hidden state. The normalized copy is used to compute attention queries and values, while the original state supplies the keys and residual connection. This produces token-dependent, bounded attention updates without a direct steering addition to the residual stream. We do not claim that Q/V steering itself is new; our contribution is this construction, its motivation, and the analysis of its effects (Sections 1 and 3.3, Appendix I).

We expanded the experiments to examine both behavioral control and its costs. Table 1 reports higher mean joint truthfulness–informativeness and lower mean toxicity for PSAS than the evaluated DISCO-QV baseline across all three models. Appendix J.1 provides additional matched-control comparisons, while Appendix K presents control–quality curves across three models and two tasks. We also added evaluations of safety steering, non-toxic prompts, out-of-distribution transfer, and Gemma-2-9B.

Finally, we clarify what the theoretical guarantees mean in practice. Under the stated assumptions, the residual perturbation remains bounded through subsequent layers as the steering coefficient increases (Section 4.2, Appendix B.3). This does not guarantee preservation of model behavior: excessive steering can still reduce informativeness. We therefore examine behavioral performance across steering strengths in Section 6.4 and Appendix G.2, rather than relying on boundedness alone.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.