acceptodds
Under review as a conference paper at ICLR 2027

Refusal is non-stationary: Iterative steering for LLM uncensoring

Abstract

Despite recent progress in safety alignment, Large Language Models (LLMs) remain vulnerable to steering attacks that bypass safety mechanisms and produce uncensored models that no longer refuse harmful requests. Existing methods typically estimate fixed refusal-related directions and apply them through one-shot interventions, yielding limited success on strongly aligned models or substantially degrading performance on benign tasks. We formulate uncensoring as a constrained steering problem that minimizes refusal on harmful prompts while limiting behavioral deviation on benign inputs, thereby discouraging interventions that reduce refusal through broad corruption of model behavior. We further show that refusal is non-stationary under steering: each intervention changes how residual refusal is represented, causing previously estimated directions to become progressively less effective. Motivated by this observation, our method iteratively re-estimates refusal after each intervention, recomputes layer-specific steering directions, and selects successive interventions subject to the benign-behavior constraint. Across eleven LLMs, our approach consistently achieves higher attack success rates than existing steering methods, with particularly large gains on larger and more strongly aligned models where prior approaches largely fail. Finally, the ordered sequence of interventions produced by our method provides a tool for mechanistic interpretability, revealing how refusal is distributed across layers and how its representation evolves throughout the steering process.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large language models, model steering, safety alignment, robustness evaluation, mechanistic interpretability

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_4s3eOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
1
Soundness
2
Read review
Paper summary

The paper repeatedly updates its estimate of refusal behavior and selects interventions subject to a constraint on benign behavior. Its strongest contribution is the empirical demonstration that iterative adaptation exposes vulnerabilities missed by several existing steering methods.

Critical strengths
  • Table 1 reports attack success increasing from the strongest baseline’s 43% to 96% on Qwen2.5-32B, and from 65% to 85% on Qwen3.8-27B. The comparisons include approaches targeting multiple refusal directions.
  • idea is novel
Critical weaknesses
  • I feel the results make a much stronger case for iterative adaptation than they do for true mechanistic nonstationarity. From my perspective, seeing shifting directions and nearly orthogonal interventions from step to step could just as easily be explained by a static, multidimensional refusal representation, especially since earlier interventions can naturally warp the geometry we measure later on. I would find it much more convincing if the authors compared their re-estimation approach against a fixed multi-directional representation found at the very start, keeping the number of interventions, selection criteria, and compute budget identical.
  • I find the constraint meant to preserve benign behavior pretty narrow. Because it only restricts the KL divergence on the very first generated token, I'm not convinced that full responses or multi-step reasoning stay intact. I see this playing out in Table 1: on Mistral, MMLU drops from 57% down to 48%, and ARC-E falls from 69% to 53%. Rather than showing that capabilities are preserved across the board, I think the results really just show a favorable trade-off between attack success and utility on certain models.
  • In my opinion, the capability evaluations are too limited in scope. Running two 500-question multiple-choice benchmarks is a reasonable starting point, but it doesn't tell me whether coding, open-ended instruction following, long-form writing, or complex reasoning are holding up. Along the same lines, I don't think doing multiple generation runs is a substitute for testing across different optimization seeds and genuinely held-out harmful prompt distributions.
Questions for authors

see weakness

5 comments
AuthorsOpenReview ↗

Rebuttal 4s3e22 (PT1)

We thank the reviewer for recognizing the novelty of our approach and the substantial improvements in attack success over strong baselines, including methods targeting multiple refusal directions. We provide below our point-by-point responses to the concerns and questions raised by the reviewer.


I feel the results make a much stronger case for iterative adaptation than they do for true mechanistic nonstationarity.

Author's Response (UPDATED) We also wish to clarify that our evaluation already includes multidirectional approaches, namely RCO and SOM, and that our method achieves the highest ASR on ten of eleven models across the evaluated methods. Furthermore, Table 9 documents changes in refusal expression, while Tables 2 and 7 support the benefit of updating the refusal vocabulary and filtering the estimation data as steering progresses.

To examine whether our gains simply reflect a larger number of ablations, we have prepared a plot of ASR against the number of applied ablations across all evaluated models. We provide a summary table below and include the complete figure, showing the trajectory of our method alongside all evaluated baseline configurations, in Appendix A.3, Figure 5. Values come from a single greedy generation, so they can differ slightly from the five-run means in Table 1 and 3.

MethodQwen3.8-27BGemma-2-9BQwen2.5-32BGemma-3-12BGpt-oss-20b
ASR#Int.ASR#Int.ASR#Int.ASR#Int.ASR#Int.
No intervention0.01900.01300.08200.20800.0190
SD0.6101920.3711260.4401920.2261440.44072
HER0.0501280.421840.1891280.220960.14548
RDO0.2391920.0631260.1451920.1951440.01372
RCO 2D0.0631920.3651260.1191920.2141440.00072
RCO 3D0.0311920.3211260.1641920.2581440.09472
SOM 5 directions0.7116450.7484250.1456450.2084850.914245
SOM 6 directions0.6737740.7615100.2337740.1575820.939294
SOM 7 directions0.5979030.7235950.2209030.0386790.862343
Ours0.906240.95070.950740.780570.91820

The results indicate that, on several models, our method achieves higher ASR despite applying fewer ablations. These results indicate that the number of ablations alone does not explain the observed advantage.

Regarding the proposed fixed-representation comparison, matching the number of orthogonal directions additionally requires specifying how those directions are extracted from the original model, selected, and assigned to layers. In our approach, their successive identification is itself part of the method. After each intervention, we estimate the refusal behaviour that remains and use it to identify the next layer-specific direction. Our existing comparisons evaluate the resulting method, while a matched fixed-representation control would further isolate the contribution of re-estimation.

The interpretation suggested by the reviewer also helps refine our mechanistic analysis. Near-zero cosine similarity is expected under directional projection, but the informative observation is that further interventions at the same layer remain useful after a single direction has been removed. This supports a multidimensional account of refusal and complements our analysis of the intervention sequence in Figure 2.

We revised the mechanistic interpretability analysis in our paper to incorporate this complementary interpretation and clarify the operational meaning of non-stationarity. We thank the reviewer for helping us strengthen this discussion and hope that these clarifications and additional results encourage a more positive reassessment of our contribution.

AuthorsOpenReview ↗

Rebuttal 4s3e22 (PT2)

I find the constraint meant to preserve benign behavior pretty narrow. Because it only restricts the KL divergence on the very first generated token, I'm not convinced that full responses or multi-step reasoning stay intact. I see this playing out in Table 1: on Mistral, MMLU drops from 57% down to 48%, and ARC-E falls from 69% to 53%. Rather than showing that capabilities are preserved across the board, I think the results really just show a favorable trade-off between attack success and utility on certain models.

Authors’ response. (UPDATED) We thank the reviewer for this fair reading. We wish to clarify that the first-token KL is a computationally efficient criterion for limiting behavioral deviation during intervention selection, rather than a guarantee on complete responses or multi-step reasoning. Our utility assessment therefore relies on independent evaluations: MMLU and ARC-E and, in the revised manuscript, IFEval, which grades complete generated responses with reasoning enabled for reasoning models (Appendix A.6, Table 6). The ASR–utility trade-offs are reported in Appendix A.2. On seven of the eleven models (Qwen2.5-3B, Gemma-2-9B, Gemma-3-12B, Llama-2-13B, Qwen2.5-32B, Qwen3.8-27B, and GPT-oss-20B), substantial ASR improvements coexist with decreases of no more than 0.04 on either MMLU or ARC-E relative to the original models. On IFEval, our method stays within 0.06 of the original model’s accuracy on ten of the eleven models.

As discussed in Section 5, Mistral-7B exhibits a more pronounced trade-off: ASR increases from at most 0.04 among the baselines to 0.42, while MMLU and ARC-E decrease by 0.09 and 0.16, respectively, and IFEval decreases from 0.46 to 0.23. We note that the evaluated model incorporates Representation Rerouting, the circuit-breaking approach of Zou et al. (2024), which directly modifies representations involved in harmful generation. This provides a possible explanation for why suppressing its safeguards may affect capabilities more substantially, although establishing this connection causally requires further investigation.

Regarding the reviewer’s concern about the ASR–utility trade-off, we agree that this may arise for some models. In this regard, Figure 4 in Appendix A.2 traces validation ASR and utility across iterations on Mistral-7B. Validation ASR remains at 0.00 for the first 17 iterations and then rises gradually, reaching its maximum of about 0.59 at iteration 95, the selected checkpoint. Over the same period, MMLU decreases from about 0.60 to 0.50 and ARC-E from about 0.67 to 0.53, with most of the ARC-E decline occurring within the first 45 iterations. Earlier checkpoints therefore offer different ASR–utility trade-offs. Selecting the checkpoint with the highest validation ASR is one possible choice; applications that prioritize utility could favor an earlier one. We have revised the manuscript to state that our method substantially improves attack effectiveness while retaining utility on several evaluated models, rather than claiming uniform capability preservation.

Lastly, please note that since our interventions are implemented through hooks without modifying the underlying weights, they can also be removed when needed to recover the original model behavior. We have clarified that the benign-behavior constraint is intended to discourage degenerate solutions in which refusal disappears because generation is broadly disrupted.

Thanks for the response. I would like to keep my score.

Reviewer_9PHHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes an iterative model-steering attack to remove safety refusal mechanisms from LLMs, producing uncensored models. The authors observe that "refusal is non-stationary under steering" where each intervention changes how the remaining refusal is represented, rendering previously estimated directions ineffective. They design a greedy layer-wise algorithm that re-estimates refusal at each iteration, updates the refusal token set, adaptively filters training prompts, and selects the optimal layer and direction under a KL-divergence constraint. Experiments on 11 open-source models show significantly higher attack success rates compared to existing baselines (SD, RCO, SOM), while preserving benign-task performance. The ordered intervention sequence also provides a preliminary tool for mechanistic interpretability of refusal.

Critical strengths
  1. Practical empirical finding: The paper identifies and demonstrates that refusal directions drift under intervention, a phenomenon overlooked by prior work. Ablation studies show that ignoring this drift leads to catastrophic performance drops (e.g., Llama-2-7B ASR falls from 0.86 to 0 with a single iteration).
  2. Interpretability byproduct: The recorded intervention sequence offers a coarse-grained view of how refusal is distributed across layers, which could serve as a lightweight auditing tool.
Critical weaknesses
  1. Unclear positioning: The paper tries to serve two roles but excels at neither. As an academic work, it lacks depth since the core claim ("refusal drifts") is merely an empirical observation with no causal explanation or formal characterization, offering little conceptual insight. As a technical work, it fails to demonstrate sufficient power: experiments are limited to open-source models, with no validation against commercial systems like GPT-4 or Claude that employ system-level safety mechanisms (e.g., independent input/output classifiers). The result is a paper that is neither intellectually deep nor technically formidable.
  2. Shallow theory: The concept of "non-stationarity" is never formally defined or mechanistically explained. Why refusal drifts, and how to predict its evolution, remain unanswered.
  3. Limited technical novelty: The method is essentially a combination of iterative optimization, vocabulary updates, and KL regularization. All well-known techniques applied elsewhere. The contribution is incremental and feels more like a competent open source project than aconference publication.
  4. Narrow technical scope: The attack surface is confined to internal model representations. Against commercial models with system-level guardrails (input filtering, output monitoring), the method is likely ineffective, yet this limitation is not discussed.
Questions for authors
  1. How would the attack success rate change when applied to commercial models with independent input/output safety detectors (e.g., GPT-4o)? Was the experimental design intentionally limited to open-source models?
  2. Can you provide a formal definition or quantitative metric for "non-stationarity"? For instance, is there a measurable indicator that signals when re-estimation is needed?
  3. How was the KL threshold tau chosen? Is there a principled selection criterion, or was it tuned per model? How sensitive are the results to tau?
  4. Your interpretability analysis shows that the same layer is selected multiple times with near-zero cosine similarity between successive directions. Does this imply multiple orthogonal subspaces for refusal within a single layer? What semantic roles might these subspaces play?
4 comments
AuthorsOpenReview ↗

Rebuttal 9PHH (PT1)

We thank the reviewer for highlighting our finding that refusal directions drift under intervention and the ablation evidence showing why accounting for this drift matters. We also appreciate their recognition of the intervention sequence as a potential lightweight tool for auditing how refusal is distributed across layers.

We provide in the following a point-to-point response to their concern, wishing to clarify our position and contributions.


Unclear positioning: The paper tries to serve two roles but excels at neither. As an academic work, it lacks depth since the core claim ("refusal drifts") is merely an empirical observation with no causal explanation or formal characterization, offering little conceptual insight. As a technical work, it fails to demonstrate sufficient power: experiments are limited to open-source models, with no validation against commercial systems like GPT-4 or Claude that employ system-level safety mechanisms (e.g., independent input/output classifiers). The result is a paper that is neither intellectually deep nor technically formidable.

Authors’ response. (UPDATED) We thank the reviewer for raising this concern. We respectfully disagree with the overall assessment and wish to clarify the contribution and scope of our work. Our contribution goes beyond observing that refusal changes under intervention. We formulate uncensoring as a constrained optimization problem and develop an adaptive, layer-wise procedure whose components are evaluated through the ablations in Tables 2 and 7. Furthermore, the ordered intervention sequence provides a tool to investigate where steering acts and how residual refusal behaves across successive interventions. We acknowledge that a complete formal and causal characterization remains an open question. Nevertheless, we believe our analysis provides strong empirical evidence of how models distinguish harmful from benign requests at the representation level, consistent with prior work on refusal directions and multidimensional refusal structure (Arditi et al., 2024; Wollschläger et al., 2025; Piras et al., 2026). Regarding the experimental setting, Section 3 explicitly considers a white-box adversary with access to model weights and internal representations. This setting is shared by the steering approaches included in our comparisons (Arditi et al., NeurIPS 2024; Wollschläger et al., ICML 2025; Piras et al., AAAI 2026). Its practical relevance is also supported by Lin et al. (2025), who document the use of uncensored models in malicious services. Evaluating publicly available models is therefore consistent with the threat model, the established evaluation setting, and the related literature. Furthermore, the reviewer is correct that the method cannot be applied through a commercial API alone, as it does not provide white-box access to the adversary. However, developers with such access, including OpenAI and Anthropic, could in principle use this approach to audit their models’ alignment and investigate refusal mechanisms. Independent input/output classifiers instead provide an additional layer of deployment-level protection, which is outside the scope of the model intervention studied here. In particular, when the adversary controls the deployment of a modified model (white-box access), these external safeguards cannot be assumed to remain in place.

Lastly, within this threat model established in the literature, our evaluation covers eleven models and multiple steering baselines, including multidirectional approaches. Our method achieves the highest ASR on ten of eleven models, improving from 0.43 to 0.96 on Qwen2.5-32B and from 0.65 to 0.85 on Qwen3.8-27B over the strongest evaluated baselines. These results demonstrate that existing steering methods can substantially underestimate vulnerability to direct model modification. We believe this represents a meaningful contribution to alignment robustness evaluation, even without claiming a complete causal theory of refusal.

We hope that these clarifications and the empirical evidence encourage the reviewer to reassess our contribution in light of its stated objectives.

AuthorsOpenReview ↗

Rebuttal 9PHH (PT2)

Shallow theory: The concept of "non-stationarity" is never formally defined or mechanistically explained... (Q2) Can you provide a formal definition or quantitative metric for "non-stationarity"?

Authors' response. (UPDATED) We thank the reviewer for motivating a more precise characterization of refusal under steering. At iteration ii, we describe refusal through the triplet F(i)=({r^l(i)}l=1L,Vr(i),R(i)(⋅))\mathcal{F}^{(i)} = \bigl(\lbrace\hat{\mathbf{r}}_{l}^{(i)}\rbrace_{l=1}^{L}, V_r^{(i)}, R^{(i)}(\cdot)\bigr). The directions r^l(i)\hat{\mathbf{r}}_{l}^{(i)} capture the representation-level contrast used to identify candidate interventions, the vocabulary Vr(i)V_r^{(i)} captures how refusal is expressed, and the score R(i)(x)=R(x,Vr(i);Mw~(i))R^{(i)}(x) = R(x, V_r^{(i)}; M_{\tilde{w}^{(i)}}) estimates the propensity to refuse prompt xx.

Non-stationarity, in our operational sense, means that this triplet cannot be assumed to remain fixed throughout steering. Each intervention modifies the representations from which the directional component is estimated. Meanwhile, changes in refusal expression affect the vocabulary component Vr(i)V_r^{(i)} and, consequently, the score R(i)(⋅)R^{(i)}(\cdot) used to select informative prompts. Changes in the selected prompts further affect the contrast used to estimate subsequent directions. Our method therefore updates these coupled components to describe the refusal behaviour that remains after each intervention. In particular, Table 9 documents changes in the vocabulary component, while Tables 2 and 7 demonstrate the practical benefit of updating the vocabulary and the prompt sets used to estimate the directions. These findings support tracking F(i)\mathcal{F}^{(i)} throughout steering, without assuming that steering creates new refusal components. They also explain why the estimated description can change even when refusal involves components already present in the original model. We will revise the paper to include this formal operational definition and make its connection to the empirical findings explicit.


Limited technical novelty: The method is essentially a combination of iterative optimization, vocabulary updates, and KL regularization. ...

Authors’ response. (UPDATED) We respectfully disagree that our contribution is merely a combination of established techniques. While iterative optimization and KL regularization are general tools, the steering methods considered in our comparison retain directions estimated from the original model. Our procedure instead recomputes layer-specific directions and updates the refusal vocabulary throughout steering, adapting the estimation process to the model’s current behavior. This adaptive treatment is absent from the compared approaches.

At a higher level, our contribution lies in three connected elements: (i) the formulation of uncensoring as a constrained optimization problem (Eq. 1), which discourages interventions that reduce refusal through broad behavioral degradation; (ii) the empirical evidence that updating refusal estimates during steering improves effectiveness; and (iii) the demonstration that this changes the assessment of model robustness. Our method achieves the highest ASR and JS on ten of eleven models, including an ASR gain of 0.53 over RCO on Gemma-3-12B and an increase from at most 0.04 among all baselines to 0.42 on Mistral-7B.

Table 2 further shows that the adaptive components materially contribute to these results, with their removal producing model-dependent ASR reductions of up to 0.85. The combination is therefore supported by targeted ablations. We believe that simplicity is a valuable property for an auditing tool, and that the significance of our contribution lies in exposing vulnerabilities that existing steering methods underestimate.


AuthorsOpenReview ↗

Rebuttal 9PHH (PT3)

Narrow technical scope: The attack surface is confined to internal model representations.

Authors’ response. As discussed in our response to Weakness 1, our work considers a white-box adversary with access to the weights and internal representations of publicly released models. This is an established threat model in the steering literature, including Arditi et al. (NeurIPS 2024), Wollschläger et al. (ICML 2025), and Piras et al. (AAAI 2026).

Beyond its established use, we wish to emphasize its practical relevance both for providers releasing aligned open-weight models and for companies auditing their models internally. In the former case, providers seeking to limit misuse need to understand not only whether a model refuses harmful requests at release, but also how easily its safeguards can be weakened through subsequent modification. Public collections already distribute models described as uncensored variants produced using Heretic, illustrating that such modifications and redistribution occur in practice. Our method helps assess this vulnerability and can inform the development of alignment mechanisms that are more resistant to uncensoring. In the latter case, developers with internal access, including OpenAI and Anthropic, could in principle use this approach to audit model-level alignment and investigate refusal mechanisms.

We have revised the paper to expand the discussion of the practical relevance of our threat model. We hope that these clarifications address the concern and encourage the reviewer to reassess the contribution in light of these applications.


Your interpretability analysis shows that the same layer is selected multiple times with near-zero cosine similarity between successive directions. Does this imply multiple orthogonal subspaces for refusal within a single layer? What semantic roles might these subspaces play?

Authors’ response. The key finding is that, after removing one refusal-related direction, additional interventions at the same layer can further reduce refusal. For example, layer 16 is selected seven times on Llama-2-7B. This provides evidence that a single direction does not fully capture the refusal behaviour accessible through interventions at that layer, supporting a multidimensional account. The near-zero cosine similarities are expected under our projection procedure, since subsequent directions are estimated from activations whose component along the previously removed direction has been eliminated. The value of the analysis therefore lies in demonstrating the behavioural relevance of these additional directions, rather than their orthogonality alone. Therefore, their successive identification helps reveal refusal components that remain relevant after earlier interventions. We share the interest in understanding the semantic roles of these directions and reserve this analysis for future work. In this regard, we believe that a logit-lens analysis could be used to project each direction through the unembedding matrix to identify associated tokens, while tracking changes in refusal across prompt categories could help interpret its behavioural role.



Summary. We have posted responses to several points raised and are conducting further evaluations to complete our response. We will share these additional results as they become available.

In the meantime, we would be grateful if the reviewer could consider the clarifications and results provided so far. We hope that this evidence helps address the concerns raised and supports a more positive assessment of our contribution. Should any point remain unclear or require further discussion, we would welcome the opportunity to clarify it.

AuthorsOpenReview ↗

Rebuttal 9PHH (PT4)

Dear Reviewer 9PHH, we provide below a follow-up to our earlier responses addressing the remaining concerns.


How was the KL threshold tau chosen? Is there a principled selection criterion, or was it tuned per model? How sensitive are the results to tau?

Authors’ response. (UPDATED) We wish to clarify that (\tau=0.1) is fixed across all models in our main experiments, following the threshold used by Arditi et al. (2024). It was not tuned per model.

Appendix A.8, Table 8, already reports a sensitivity analysis on Qwen2.5-3B over thresholds ranging from 0.001 to 1.0. Very restrictive values, between 0.001 and 0.02, yield little or no improvement over the original model, with ASR between 0.13 and 0.15. At (\tau=0.05) and (0.10), ASR increases to 0.83 and 0.92, respectively, while MMLU and ARC-E remain close to their original values. These results support the adopted default as a useful balance between refusal reduction and behavioral preservation on this model.

To further assess sensitivity across models, we have extended this analysis to Llama-2-7B, Gemma-2-9B, and Mistral-7B-Instruct-RR, using the same threshold values and evaluation protocol. The tables below report the updated results for all four models as means and standard deviations over five evaluation seeds. A dash (–) indicates that no candidate layer satisfies the KL constraint, so no ablation is performed.

τQwen2.5-3B ASRQwen2.5-3B MMLUQwen2.5-3B ARC-EQwen2.5-3B Iter.Llama-2-7B ASRLlama-2-7B MMLULlama-2-7B ARC-ELlama-2-7B Iter.
0.0010.13 ± 0.010.67 ± 0.020.69 ± 0.02–0.00 ± 0.000.39 ± 0.010.74 ± 0.02–
0.0020.13 ± 0.010.67 ± 0.020.69 ± 0.02–0.00 ± 0.000.39 ± 0.010.74 ± 0.02–
0.0050.13 ± 0.010.67 ± 0.020.69 ± 0.02–0.00 ± 0.000.39 ± 0.010.74 ± 0.02–
0.010.13 ± 0.000.67 ± 0.020.69 ± 0.011/40.00 ± 0.000.39 ± 0.010.74 ± 0.018/8
0.020.15 ± 0.010.66 ± 0.020.70 ± 0.014/40.53 ± 0.010.37 ± 0.010.73 ± 0.0235/66
0.050.83 ± 0.020.56 ± 0.010.62 ± 0.0182/990.40 ± 0.020.40 ± 0.020.73 ± 0.0129/60
0.1 (default)0.92 ± 0.010.63 ± 0.010.68 ± 0.0123/960.86 ± 0.030.32 ± 0.020.68 ± 0.0162/77
0.20.90 ± 0.010.65 ± 0.010.68 ± 0.014/350.78 ± 0.020.37 ± 0.010.73 ± 0.0130/61
0.50.93 ± 0.020.65 ± 0.010.67 ± 0.014/350.70 ± 0.040.29 ± 0.020.68 ± 0.0268/75
1.00.92 ± 0.010.61 ± 0.010.61 ± 0.0117/480.84 ± 0.030.37 ± 0.030.70 ± 0.0151/57
τGemma-2-9B ASRGemma-2-9B MMLUGemma-2-9B ARC-EGemma-2-9B Iter.Mistral-RR ASRMistral-RR MMLUMistral-RR ARC-EMistral-RR Iter.
0.0010.01 ± 0.000.34 ± 0.030.78 ± 0.0114/140.00 ± 0.000.57 ± 0.020.69 ± 0.01–
0.0020.02 ± 0.000.47 ± 0.020.80 ± 0.0098/990.00 ± 0.000.57 ± 0.020.69 ± 0.01–
0.0050.51 ± 0.000.50 ± 0.020.80 ± 0.0197/990.00 ± 0.000.57 ± 0.020.70 ± 0.0190/99
0.010.79 ± 0.000.58 ± 0.010.81 ± 0.0172/990.00 ± 0.000.58 ± 0.020.70 ± 0.0114/14
0.020.86 ± 0.000.50 ± 0.010.81 ± 0.0119/550.00 ± 0.000.57 ± 0.020.68 ± 0.0298/99
0.050.92 ± 0.000.59 ± 0.010.82 ± 0.0127/580.00 ± 0.000.51 ± 0.030.65 ± 0.0271/99
0.1 (default)0.96 ± 0.000.29 ± 0.030.79 ± 0.006/400.42 ± 0.000.48 ± 0.020.53 ± 0.0195/99
0.20.91 ± 0.000.44 ± 0.030.80 ± 0.014/450.09 ± 0.000.56 ± 0.020.69 ± 0.0122/56
0.50.97 ± 0.000.56 ± 0.000.80 ± 0.013/320.19 ± 0.000.49 ± 0.030.58 ± 0.0125/56
1.00.97 ± 0.000.56 ± 0.000.80 ± 0.013/150.20 ± 0.000.50 ± 0.030.54 ± 0.0137/63

These results show that the threshold yielding the highest mean ASR varies across models. In our main experiments, we fix τ=0.1\tau=0.1 across all models, following Arditi et al.~(2024), to ensure a consistent evaluation setting without model-specific tuning. The additional analysis shows that tuning τ\tau can yield further improvements over this shared default. For example, setting τ=0.5\tau=0.5 increases mean ASR from 0.920.92 to 0.930.93 on Qwen2.5-3B and from 0.960.96 to 0.970.97 on Gemma-2-9B. Our main results therefore demonstrate the effectiveness of the method under a common threshold, while leaving room for further gains through model-specific tuning.


Conclusion of rebuttal. We would be grateful if you could consider these additions, which we hope clarify the contribution of our work and support a more positive reassessment. We remain available to discuss any points that may require further clarification.

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

This paper studies the uncensoring problem of LLMs by ablating refusal directions. The authors cast the problem as minimizing a first-token refusal score on harmful prompts, subject to a KL bound on benign prompts relative to the original model. Their key contribution is that they defines refusal as non-stationary under steering: each ablation changes the tokens, prompts, and directions after each ablation. Therefore, they propose a greedy iterative method that, at each step, re-estimates the refusal tokens, filters the prompts, recomputes per-layer directions, and ablates the layer that most reduces refusal under the KL constraint. On eleven LLMs, the method achieves the best ASR/JS on ten against five steering baselines. Besides, ablations and layer-selection/PCA analyses are offered as supporting evidence.

Critical strengths
  1. The paper argues that the apparent robustness of larger, aligned open-weight models to refusal-direction attacks may reflect the weakness of fixed-direction attacks rather than genuine robustness. Hence, they offer a sound formulation relevant to white-box safety auditing.
  2. The method is compared with five steering baselines on eleven LLMs, using two judges and component ablations, and the gains are outstanding on several models.
Critical weaknesses
  1. The central claim that refusal is non-stationary is not isolated from simpler explanations, and part of the supporting evidence is confounded.
  • Figure 1 is without the construction of the method. If the ablation is applied to the activation h_l as in Eq. (5), the hidden states of the modified model have no component along the removed direction at that layer. However, the near-zero cosine similarity in Figure 1 does not show that the refusal representation has changed.
  • The gains are equally consistent with a static alternative. Refusal always occupy a multi-dimensional or redundant subspace (as suggested by the RCO and SOM lines of work).
  • The ablation evidence is statistically thin and not fully consistent with the narrative. Table 2 reports a single deterministic run on a test set of 159 prompts without confidence intervals.
  1. Utility preservation and the fairness of the baseline comparison are not convincingly established.
  • The benign-behavior constraint is weak. The KL in Eq. (3) is computed only on the first-token distribution, which does not control degradation over multi-token generation.
Questions for authors

Q1. Please add a non-iterative baseline that removes the same number of orthogonal directions from the original model, and plot ASR against the number of interventions for all methods. This would separate the effect of re-estimation from the effect of removing directions.

Q2. Why are the base MMLU scores so low (0.22 for Llama-2-7B and 0.28 for GPT-oss-20B)? Please describe the evaluation protocol (prompt format, chat template, answer extraction), including how the reasoning models were evaluated.

Q3. Were the baselines on Gemma-4-12B run correctly? Please show MMLU and ARC-E near-chance scores of SD, RDO, and RCO. Besides, the gains without the HarmBench classifier in the loop need to clarify.

5 comments
AuthorsOpenReview ↗

Rebuttal nRTP (PT1)

We thank the reviewers for their constructive feedback and for recognizing the relevance of our findings to auditing, the breadth of our evaluation, and the substantial gains over existing steering methods. Below, we provide a point-by-point response to the main concerns raised by the reviewers.


W1: The central claim that refusal is non-stationary is not isolated from simpler explanations, and part of the supporting evidence is confounded.

Authors’ response (UPDATED). We thank the reviewer for raising this concern. We address each point separately below. In summary, we wish to clarify that the evidence we rely on for non-stationarity is behavioural rather than geometric. Specifically, Table 9 shows that the dominant token beginning refusal responses on Gemma-2-9B changes from I to Please, That, and You. Furthermore, the ablations in Table 2 show the importance of updating the refusal vocabulary and adaptively selecting the prompts used to estimate refusal directions. Keeping both components fixed (see the w/o update and filtering variant) reduces ASR from 0.86 to 0.04 on Llama-2-7B and from 0.96 to 0.28 on Gemma-2-9B. Lastly, the set of prompts that still elicit refusal shrinks and shifts, which is why adaptive filtering matters (Table 2), and Appendix B already shows that neighbouring-layer directions become less correlated as steering progresses. We revised the paper to clarify the distinction between these behavioural observations and the geometric analysis in the manuscript. In the following we provide point-to-point responses to the more specific arguments raised by the reviewer.


W2: Figure 1 is without the construction of the method. If the ablation is applied to the activation h_l as in Eq. (5), the hidden states of the modified model have no component along the removed direction at that layer. However, the near-zero cosine similarity in Figure 1 does not show that the refusal representation has changed.

Authors’ response. We agree that, after the ablation in Eq. (5), the activations at the intervened layer have no component along the removed direction. Consequently, the near-zero cosine similarities should not, by themselves, be interpreted as evidence of non-stationarity. Instead, these observations are consistent with a multidimensional account of refusal: removing one direction does not eliminate refusal, and subsequent interventions along additional, mutually orthogonal directions at the same layer can further reduce it. For example, layer 16 is selected seven times on Llama-2-7B. The relevant finding is therefore the contribution of multiple directions to refusal reduction, rather than their orthogonality alone. We therefore view this observation as consistent with the rationale of our approach: refusal can persist after removing a single direction, and targeting additional directions at the same layer can further reduce it. In light of this comment from the reviewer, we revised the Mechanistic Interpretability analysis in the experimental result section to make this distinction explicit and clarify its relationship with a multidimensional refusal representation.


W3: The gains are equally consistent with a static alternative. Refusal always occupy a multi-dimensional or redundant subspace (as suggested by the RCO and SOM lines of work).

Authors’ response (UPDATED). This is a fair alternative explanation. We agree that refusal may involve multiple or redundant components already present in the original model. However, this does not imply that a fixed estimate of those components remains adequate throughout steering. Removing one component may change which prompts remain refused and which residual components are most informative. The existence of a multidimensional refusal structure is therefore compatible with our motivation for adaptive estimation. They are not mutually exclusive. We wish, however, to present that our evaluation already includes methods that explicitly model multidimensional refusal. RCO with up to three dimensions and SOM with five, six, and seven directions. Our method achieves the highest ASR on ten of eleven models, including gains of 0.53 over RCO on Gemma-3-12B and 0.20 over SOM on Qwen3.8-27B. These comparisons show that the improvement extends beyond static-direction baselines. Tables 2 and 7 further demonstrate the benefit of updating the refusal vocabulary and filtering the estimation data as steering progresses. The fundamental question that arises thanks to the reviewer is whether an initial estimate remains adequate after successive interventions. In this sense, our results show the value of tracking residual refusal as steering progresses, whether its components are reorganized or already present in the original model. We clarified that our finding concerns the importance of adaptive estimation and does not require assuming that steering creates new refusal components.

AuthorsOpenReview ↗

Rebuttal nRTP (PT2)

W4: Utility preservation and the fairness of the baseline comparison are not convincingly established.

Authors’ response (UPDATED). Regarding comparison fairness, all methods use identical training, validation, and test splits. Baselines retain their default hyperparameters, except for the RCO dimensionality and SOM direction count, for which we evaluate multiple configurations and select the best by validation ASR. Our method likewise uses validation ASR to select its final iteration. Appendix A.1 reports all baseline configurations. To accommodate the reviewer, we revised the experimental setup to make this fairness point clearer. Our benign-behavior constraint follows the same principle implemented in Arditi et al. (SD), Weidmann et al. (HER), and Wollschläger et al. (RCO and RDO). The remaining baseline SOM use different intervention objectives and do not incorporate this constraint in their original formulations. Imposing it would therefore modify it rather than evaluate them as proposed. We instead assess all resulting models using the same utility benchmarks and report the ASR–utility trade-offs in Appendix A.2. We hope these clarifications address the reviewer’s concern about comparison fairness and would welcome any remaining concerns or suggestions.


Q1:. Please add a non-iterative baseline that removes the same number of orthogonal directions from the original model, and plot ASR against the number of interventions for all methods. This would separate the effect of re-estimation from the effect of removing directions.

Authors’ response. (UPDATED) We thank the reviewer for this suggestion. Our evaluation already includes approaches that estimate multiple refusal directions from the original model, including multiple variants of SOM. However, the number of estimated directions is not directly equivalent to the number of interventions: these baselines apply their directions across multiple layers, whereas each iteration of our method targets a single layer. Furthermore, SOM identifies closely related, non-orthogonal directions, rather than the orthogonal directions specified in the proposed comparison. Consequently, the number of directions used by these baselines is not equivalent to the number of interventions in our procedure. Constructing an iterative trajectory for them would change their original formulations.

To examine whether our gains simply reflect a larger number of ablations, we have prepared a plot of ASR against the number of applied ablations across all evaluated models. We provide a summary table below and include the complete figure, showing the trajectory of our method alongside all evaluated baseline configurations, in Appendix A.3, Figure 5. Values come from a single greedy generation, so they can differ slightly from the five-run means in Table 1 and 3.

MethodQwen3.8-27BGemma-2-9BQwen2.5-32BGemma-3-12BGpt-oss-20b
ASR#Int.ASR#Int.ASR#Int.ASR#Int.ASR#Int.
No intervention0.01900.01300.08200.20800.0190
SD0.6101920.3711260.4401920.2261440.44072
HER0.0501280.421840.1891280.220960.14548
RDO0.2391920.0631260.1451920.1951440.01372
RCO 2D0.0631920.3651260.1191920.2141440.00072
RCO 3D0.0311920.3211260.1641920.2581440.09472
SOM 5 directions0.7116450.7484250.1456450.2084850.914245
SOM 6 directions0.6737740.7615100.2337740.1575820.939294
SOM 7 directions0.5979030.7235950.2209030.0386790.862343
Ours0.906240.95070.950740.780570.91820

The results show that, on several models, our method achieves higher ASR despite applying fewer ablations. These results indicate that the number of ablations alone does not explain the observed advantage.

Lastly, we wish to clarify that matching the number of orthogonal directions does not, by itself, specify a non-iterative counterpart to our method. Such a comparison also requires a procedure for extracting and selecting those directions from the original model and determining where to apply them. In our approach, identifying the subsequent directions is itself part of the method: after each intervention, we estimate the refusal behaviour that remains and use it to identify the next layer-specific direction.

AuthorsOpenReview ↗

Rebuttal nRTP (PT3)

W5: The benign-behavior constraint is weak. The KL in Eq. (3) is computed only on the first-token distribution, which does not control degradation over multi-token generation.

Authors’ response (UPDATED). We thank the reviewer for raising this point. We wish to clarify that the first-token KL is intended as a computationally efficient criterion for limiting behavioral deviation, rather than a guarantee that every subsequent token remains unaffected. The beginning of a response is relevant because it conditions subsequent generation, an observation that also motivates the response-prefix objective of Zou et al. (2023). In our setting, evaluating one candidate intervention per layer at every iteration makes a criterion that avoids generating full responses particularly valuable. However, beyond this deliberate approximation to reduce computational cost, we wish to emphasize that our test-time evaluation examines complete generated responses and assesses model utility on MMLU and ARC-E. We report these measures in Table 1 and Appendix A.1.

To further validate this approximation, we examined the relationship between first-token KL and multi-token divergence on held-out benign responses across intermediate checkpoints by computing their Spearman correlation.

ModelPearsonSpearman
Qwen2.5-3B-Instruct0.6290.614
Qwen2.5-32B-Instruct0.7430.732
Qwen3.8-27B0.7250.688
Llama-2-7B-chat0.7810.774
Llama-2-13B-chat0.7880.828
Llama-3.2-3B-Instruct0.7540.815
Gemma-2-9B-it0.8030.579
Gemma-3-12B-it0.6630.659
Gemma-4-12B-it0.4790.460
Mistral-7B-Instruct-RR0.4730.494
GPT-oss-20B0.7690.750

The correlations are positive across all eleven models, with Pearson coefficients ranging from 0.473 to 0.803 and Spearman coefficients from 0.460 to 0.828. Thus, although our constraint considers only the first token, it remains informative about divergence over a ten-token prefix. These results provide empirical support for using first-token KL as a computationally efficient approximation of early-generation deviation.


Q3. Were the baselines on Gemma-4-12B run correctly? Please show MMLU and ARC-E near-chance scores of SD, RDO, and RCO. Besides, the gains without the HarmBench classifier in the loop need to clarify.

Authors’ response. (UPDATED) We thank the reviewer for raising this concern. We checked the SD, RDO, and RCO runs on Gemma-4-12B and confirmed that the interventions substantially disrupt generation. The near-chance MMLU and ARC-E scores reported in Tables 1 and 3 therefore reflect a loss of model utility following steering. To make this behaviour explicit, we have included in Appendix C representative HarmBench completions and representative MMLU/ARC-E answers.



Summary. We thank the reviewer for their patience and contribution to our work. We have posted responses to several points raised and are conducting further evaluations to strengthen and complete our response. We will share these additional results as they become available.

In the meantime, we would be grateful if the reviewer could consider the clarifications and results provided so far. We hope that this evidence helps address the concerns raised and supports a more positive assessment of our contribution. Should any point remain unclear or require further discussion, we would welcome the opportunity to clarify it.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of the rebuttal

We wish to thank the reviewers for recognizing the relevance of our findings to the safety audit, the breadth of our assessment, and the substantial advantages over existing steering methods, as well as for providing constructive feedback that substantially improved the quality and clarity of our work. In particular, after the refutation, the paper now:

  • Clarifies the operational meaning of non-stationarity, namely that the refusal vocabulary, the prompts that still elicit refusal, and the resulting layer-wise directions no longer remain adequate after an intervention. It also distinguishes the behavioral evidence (changes in refusal expression, the shrinking set of refused prompts, and the ablations) from the geometric analysis, which is now interpreted as consistent with a multidimensional account of refusal. (See Section 1, Section 5 "Mechanistic Interpretability" and Appendix B).

  • Quantifies non-stationarity directly. On Gemma-2-9B, the fraction of harmful training prompts that still elicit refusal falls from 0.48 to 0.003 within six iterations, and the dominant refusal token changes at each of them, providing practical signals that re-estimation is needed. (See Appendix A.9, Table 9)

  • Provides an analysis of ASR against the number of interventions across all eleven models. We now show that our method reaches its ASR with 7 to 95 single-layer ablations, whereas the baselines apply between 48 and 903, showing that the gains stem from where the method intervenes rather than from how much. (See Section 5 "Uncensoring effectiveness" and Appendix A.3, Figure 5)

  • Strengthens the statistical evidence. Ablations are now reported as means and standard deviations over five generation runs and extended to four additional models, and the sensitivity analysis of the KL threshold τ\tau now covers four models. (See Table 2, Appendix A.7, Table 7, and Appendix A.8, Table 8)

  • Corrects and fully documents the utility evaluation protocol. It resolves a chat-template scoring issue that affected MMLU and ARC-E on the Llama-2 and GPT-oss models (e.g., base MMLU of Llama-2-7B rises from 0.22 to 0.39), while all ASR and JS results remain unchanged. (See Section 4 "Utility evaluation", Tables 1 and 3, and Appendix A.6)

  • Expanded the capability evaluation with IFEval on all eleven models, with reasoning enabled and complete responses graded. Our method stays within 0.06 of the original model on ten of the eleven models. (See Section 5 "Model utility" and Appendix A.6, Table 6)

  • Clarifies the fairness and scope of the evaluation. We have now revised the experimental setup to further explain how each data split is used, apply the same validation criterion to all methods, reserve the StrongREJECT judge for independent evaluation, and expand the discussion of the threat model and the ethics statement. (See Section 3 "Threat model", Section 4, and the Ethics statement)

All new and modified parts of the revised manuscript are highlighted in blue for ease of reference.

We believe these revisions address the points raised in the reviews. Some interpretations proposed by the reviewers, such as the multidimensional view of refusal, turned out to be complementary to our findings rather than in conflict with them, and helped us sharpen how we present non-stationarity. The additional experiments therefore have strengthened our conclusions. We are sincerely grateful for the time and care the reviewers have put into this discussion. If they feel these revisions address their concerns, we would be grateful if they could consider a more positive assessment of our work.

We remain available to the reviewers for any further clarification or additional analysis if needed.

Dear reviewers,

As the discussion period is coming to an end, we would like to take this last opportunity to ask whether there are any remaining aspects we could clarify or address together.

We would also like to thank you all once again for your valuable feedback, which has helped us significantly improve our work. We hope that the revised version addresses your concerns, and that it may be reflected in a more positive assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.