acceptodds
Under review as a conference paper at ICLR 2027

Do Models Share Safety Representations? Cross-Model Steering for Safe Visual Generation

Abstract

Recent progress in generative modeling has made safety control a central challenge, yet existing approaches remain largely model-specific, requiring retraining or tailored interventions for each new architecture. In this work, we introduce cross-model safety steering, a framework that extracts a safety direction from paired safe and unsafe prompts in a source LLM and maps it to the text-conditioning space of a target generator through an alignment learned exclusively from benign data. Once transferred, the direction is applied at inference time while keeping the generative backbone frozen and without requiring target-side unsafe examples for alignment. The framework also supports category-specific directions, enabling selective control over different types of unsafe content. Across diverse source-target pairs for text-to-image and text-to-video generation, transferred directions consistently reduce unsafe outputs, with the alignment strategy and steering strength determining the balance between safety and generation quality. We further show that steering strength can be selected using benign data alone, reducing unsafe outputs in all 45 evaluated configurations. Together, these results show that safety interventions can transfer across independently trained models, offering a modular alternative to model-specific steering. Source code will be made publicly available.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI safety, activation steering, representation alignment, text-to-image generation, text-to-video generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_VfghOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper targets at a critical question: could a safety direction computed at one source model remains effective when transfers to another different model's representation space. This paper proposes a framework that allows we compute a safety vector at the representation of one LLM and, train a light-weighted aligner to map this vector to the representation space of the target LLM without accessing to unsafe data and finally steer the vector before it passes to the downstream generator model (Text-to-Image and Text-to-Video models). By validating under multiple underlying model types and different alignment methods, the results shows the effectiveness of this framework. What's more, this framework also supports multi-category steering.

Critical strengths
  1. This paper targets at a critical problem, where the governer can not access the unsafe data, with only the safety direction computed from another underlying model, would it be possible to leverage it to the deployed model's safety steering. Also, the paper links LLM-to-LLM transfer with Visual content moderation, which is under-explored.
  2. The evaluation is broad (3 sources, 5 targets across two modalities, 3 maps, two image benchmarks) and include lots of useful controlled studies. These controlled studies helps understanding the contribution of the selection of different components and hyper-parameters.
Critical weaknesses

I feel like the current pipeline is more of a combination of several existing technical methods. For example, Difference-of-means steering, aligning with affine/linear/MLP and additive intervention on text-encoder are all well-established components. Therefore, the combination seems to lead to a predictable result, since although the authors use Image and Video generation to make the setting as multi-modal, the upstream encoder is still an LLM so it's natural to use another LLM as the source model.

Questions for authors
  1. Could the authors clarify the technical contribution of the framework they propose? I want to see more technical design insight under this specific cross-model setting.
  2. Why the authors choose image generation and video generation as the downstream generator without any design that is specific to the visual modality? The target LLM is directly used as the encoder of the generator so I think the claim of first work in this "source-only interventions remain actionable through downstream image and video generators without target-side unsafe data" should be clarified.
  3. Would it be possible to decompose the SVD direction into a component parallel to the native direction and an orthogonal residual, and steer with each one separately under matched norm? That would help tell whether SVD’s lower ASR is really coming from the safety direction or from the rest.
  4. The baselines authors choose do not contain methods specifically designed for this multi-modal generation setting. For example, a very direct baseline could be leverage an LLM as rewriter to rewrite the input unsafe prompt, which only has access to the text input without changing the underlying generator.
3 comments
AuthorsOpenReview ↗

Response to Reviewer Vfgh (Part 1)

We thank the Reviewer for recognizing the importance of constructing interventions without target-side unsafe training pairs and the controlled evaluation. We also appreciate the concrete suggestions on SVD decomposition and prompt rewriting.


[W1/Q1] Technical contribution and design insight

The design separates unsafe source contrasts from benign target alignment. A reusable map transports the contrast once; generation then adds the saved vector to normal target conditioning states, keeping the generator frozen and the mapper outside inference. Although the components are established, benign alignment alone does not ensure a useful downstream safety intervention. The new decomposition explains effects beyond native-direction agreement, while Schnell→Dev demonstrates unchanged reuse with a shared encoder.


[W1/Q2] Visual generation and scope of the priority claim

Why visual generation? Visual generation tests whether a transferred text-space intervention remains useful after passing through a frozen downstream generator. Prior work motivates this interface: SafeText [1] aligns the text encoder, SAFREE [2] filters prompt embeddings in its image/video safeguard, and AdvUnlearn [3] finds the text encoder more suitable than the U-Net for robust concept erasure in its evaluation. These findings motivate text conditioning without making it universally preferable to visual-space interventions.

Applicability to text generation. Our mechanism is not inherently visual. We therefore include transfer from Llama-3-8B-Instruct to Mistral-7B-Instruct-v0.3 [4] and Qwen3.5-9B [5] as text-generating targets. Ridge transfers the source safe/unsafe contrast using benign Alpaca anchors for alignment and magnitude scaling, without target-side unsafe training pairs. We evaluate JailbreakBench responses with greedy decoding. Safety is the fraction classified as safe by Llama-Guard-3-8B; perplexity is measured with the intervened target on Pile and Alpaca.

Targetα\alphaLlamaGuard-safe responses (%) ↑\uparrowPile PPL increase (%) ↓\downarrowAlpaca PPL increase (%) ↓\downarrow
Mistral-7B032.00.00.0
Mistral-7B0.560.03.54.3
Mistral-7B184.028.718.4
Qwen3.5-9B040.00.00.0
Qwen3.5-9B0.550.04.11.8
Qwen3.5-9B196.020.19.5

At α=0.5\alpha=0.5, safe-response rates increase by 28.0 percentage points on Mistral and 10.0 on Qwen, with PPL increases of at most 4.3% on both evaluation corpora. At α=1\alpha=1, safety improves further at a larger language-modeling cost. These examples support applicability beyond visual generation. We will clarify that our claimed contribution concerns source-only safety transfer evaluated through frozen image/video generators, not a modality-specific alignment mechanism.


[Q3] SVD parallel/residual decomposition at matched norm

Decomposition. For unit native direction nn and transferred SVD vector vv, we form

p=(v⊤n)n,r=v−p,p=(v^{\top}n)n, \qquad r=v-p,

and rescale each component to ∥v∥2\lVert v\rVert_2. The positive projection makes the norm-matched parallel intervention identical to the native direction at that norm.

Evaluation protocol. Using a Llama-8B source, we evaluate each component under matched applied norm. Cells report detector-ASR (%) / relative benign CLIP loss (%). Unsteered ASR is 30.7% on Schnell and 30.4% on Z-Image.

Targetα\alphaParallelResidualFull SVD
Schnell128.4 / 0.3726.4 / 1.1724.0 / 0.96
Schnell510.8 / 4.063.4 / 13.353.8 / 3.45
Z-Image125.0 / 1.9616.4 / 1.9716.4 / 2.32
Z-Image313.2 / 9.861.2 / 19.510.2 / 21.94

Interpretation. The residual affects detector-ASR despite native orthogonality. On Z-Image at α=1\alpha=1, it gives lower ASR than the parallel component at nearly identical mean CLIP loss. At higher strengths, the residual and full SVD incur much greater loss. This establishes behavioral effects beyond the native direction, without isolating a pure safety feature or attributing all suppression to incoherence.

This diagnostic experiment uses native target information, including the native-derived Z-Image scale. It controls applied norms and generation parameters within each target; it does not demonstrate construction without target-side unsafe training pairs or matched utility across targets.

AuthorsOpenReview ↗

Response to Reviewer Vfgh (Part 2)

[Q4] LLM prompt-rewriting baseline

We agree that prompt rewriting is a relevant baseline. We use Llama-3-8B-Instruct with greedy decoding and no generated-image feedback. Ridge uses strength 5 on Schnell, Dev, and Qwen-Image and 3 on Z-Image. CLIP-Sim is measured against the original prompt; ASR is conditional on valid images, with image availability reported alongside it.

LLM prompt rewriting: safety and generation quality.

Schnell.

MethodASR (%) ↓Benign CLIP-Sim ↑Valid images, I2P / benign (%)
Original30.70.319100.0 / 100.0
Transfer Ridge11.00.306100.0 / 100.0
LLM prompt rewriting21.00.31678.7 / 97.3

Dev.

MethodASR (%) ↓Benign CLIP-Sim ↑Valid images, I2P / benign (%)
Original28.60.309100.0 / 100.0
Transfer Ridge12.40.297100.0 / 100.0
LLM prompt rewriting20.10.30777.1 / 97.3

Qwen-Image.

MethodASR (%) ↓Benign CLIP-Sim ↑Valid images, I2P / benign (%)
Original38.40.332100.0 / 100.0
Transfer Ridge17.10.319100.0 / 100.0
LLM prompt rewriting24.40.32378.7 / 97.3

Z-Image.

MethodASR (%) ↓Benign CLIP-Sim ↑Valid images, I2P / benign (%)
Original30.40.319100.0 / 100.0
Transfer Ridge10.50.287100.0 / 100.0
LLM prompt rewriting18.00.31278.7 / 97.3

Rewriting largely preserves mean benign CLIP-Sim among requests that produce valid images, while transfer achieves lower conditional ASR in these comparisons. Refusals and invalid responses leave 21.3–22.9% of I2P requests and 2.7% of benign requests without an image; these requests receive no safety credit. Consequently, the ASR denominators differ across methods. On the Dev prompts that yield images after rewriting, the paired original ASR is 28.5%, versus 20.1% after rewriting (difference: −8.4 percentage points; 95.0% paired interval [−11.2, −5.8]).

References

[1] Yuepeng Hu, Zhengyuan Jiang, and Neil Zhenqiang Gong. SafeText: Safe Text-to-image Models via Aligning the Text Encoder. arXiv preprint arXiv:2502.20623, 2025.

[2] Jaehong Yoon, Shoubin Yu, Vaidehi Patil, Huaxiu Yao, and Mohit Bansal. SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation. International Conference on Learning Representations (ICLR), 2025.

[3] Yimeng Zhang, Xin Chen, Jinghan Jia, Yihua Zhang, Chongyu Fan, Jiancheng Liu, Mingyi Hong, Ke Ding, and Sijia Liu. Defensive Unlearning with Adversarial Training for Robust Concept Erasure in Diffusion Models. Advances in Neural Information Processing Systems (NeurIPS), 2024.

[4] Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. Mistral 7B. arXiv preprint arXiv:2310.06825, 2023.

[5] Qwen Team. Qwen3.5: Towards Native Multimodal Agents. Qwen technical blog, 2026.

Thank the authors for the detailed response and the additional experiments. I will keep my current positive score. Please update the revised version of the paper to include the additional clarification and experimental results.

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

This paper proposes a cross-model safety steering framework that transfers safety directions from source LLMs to image and video generators using benign-only alignment. Experiments across multiple models demonstrate reduced unsafe generations while revealing trade-offs between safety and generation quality.

Critical strengths
  • The paper investigates an interesting question of whether safety representations can be transferred across independently trained models. The setting of transferring safety directions without target-side unsafe data is practically relevant.

  • The experiments cover multiple source LLMs and several text-to-image and text-to-video generators, demonstrating the feasibility of cross-model safety steering.

  • The paper provides extensive analyses of alignment strategies, steering strengths, and safety-utility trade-offs. The benign-only steering strength selection is also a useful addition.

Critical weaknesses
  • The paper's central claim is that independently trained models share safety representations. Although the authors provide geometric evidence of alignment between transferred and native safety directions, this does not necessarily establish the existence of shared safety-specific representations. The observed transfer could also reflect general semantic features associated with safe and unsafe content. More targeted experiments are needed to distinguish these explanations.

  • The proposed framework mainly combines existing techniques, including contrastive activation differences, standard representation alignment methods (SVD, Ridge, and MLP), and additive activation steering. While transferring these techniques across language and visual generation models is interesting, the methodological contribution appears incremental. It remains unclear what new mechanism or insight the framework introduces beyond combining established methods.

  • The main baselines include native steering, random directions, and string-based steering, while direct comparisons with relevant methods such as AcT and CAT are missing. Although these methods may operate under different data-access assumptions, stronger comparisons would help establish the practical advantages of the proposed approach. Moreover, Table 2 shows that transferred steering does not consistently outperform native steering at matched utility, with Ridge and MLP often performing worse.

  • Some substantial safety improvements come at the cost of considerable generation quality degradation. For example, in Table 1, SVD steering on Z-Image-Turbo reduces ASR from 0.304 to 0.002, but increases FID from 31.7 to 73.5 and decreases CLIP similarity from 0.319 to 0.249.

Questions for authors
  1. How can the authors distinguish the transfer of safety-specific representations from that of general semantic features? Could controlled experiments with non-safety semantic directions provide stronger evidence for the central hypothesis?

  2. How does the proposed framework compare with existing activation-based safety methods such as AcT and CAT under comparable safety-utility constraints?

5 comments
AuthorsOpenReview ↗

Response to Reviewer RUpw (Part 1)

We thank the Reviewer for recognizing the relevance of constructing interventions without target-side unsafe training pairs and the breadth of the safety–utility analyses. We address the distinction between functional transfer and safety specificity, the contribution, and the requested comparisons.


[W1/Q1] Safety-specific representations versus general semantics

We agree that geometric agreement with native directions does not establish a shared safety-specific representation; indeed, cosine alone does not predict suppression (Table 2). Our claim is functional: a direction estimated from source-side safe/unsafe contrasts remains effective after alignment learned only from benign anchors. Existing controls support this:

Random directions. On Qwen-Image, 20 norm-matched random directions match transfer in benign CLIP-Sim (0.309±0.003 vs. 0.310) but not in ASR (0.347±0.034 vs. 0.175; unsteered 0.366) (Table 18). Suppression is therefore not explained by perturbation magnitude alone. Category-name steering. Steering away from the six unsafe-category names gives 0.407 and 0.283 ASR on Qwen-Image and Z-Image (unsteered 0.384 and 0.304), versus 0.171 and 0.105 for Llama3.1/Ridge (Table 1). The learned contrast outperforms this simple semantic baseline. Benign uses of associated concepts. On benign prompts requiring such concepts (e.g., swords in Olympic fencing), CLIP-Sim loss at α=1 is 0.5%, versus −0.1% on ordinary prompts (Appendix C.11). In a separate human evaluation of SVD steering, 95.2% of required attributes are present on Schnell at α=5, compared with 97.1% without steering; the corresponding figures on Z-Image at α=3 are 88.9% and 96.7%. Most prompt-required attributes therefore remain present in both evaluated sets.

The same contrast also transfers to text generation: on Mistral-7B, LlamaGuard-safe JailbreakBench responses rise from 32% to 60% at α=0.5, with ≤4.3% perplexity increase (response to Reviewer Vfgh, Q2).

Disentangling safety from content semantics remains open, and we do not claim a mechanism independent of them. We will state this explicitly and list it among the limitations.


[W2] Methodological contribution and new insight

We agree that the individual components (difference-of-means directions, representation alignment, and additive steering) are established. Our contribution is the setting and what it reveals: safety supervision stays entirely on the source side, while constructing the target intervention requires only benign anchors.

Mapping a direction between text encoders does not guarantee control over visual generation. The transferred edit reaches the output through the conditioning of a frozen generative backbone, which was not trained with that edit. We show that the effect survives this pipeline and reduces unsafe generated content across image and video models, while its safety-utility trade-off varies substantially across generators.

Our analyses also clarify what makes transfer effective. The new SVD decomposition shows that a component orthogonal to the estimated native direction can reduce ASR with limited utility loss; matching the native direction is therefore not necessary for effective control. The Schnell→Dev experiment shows that an unchanged map and vector can be reused across different generative backbones sharing a conditioning space. We will make these empirical insights, and the distinction between component techniques and our contribution, clearer in the paper.

AuthorsOpenReview ↗

Response to Reviewer RUpw (Part 2)

[W3/Q2] Comparison with AcT, CAT, and native steering

AcT and CAT. We agree that AcT [1] and CAT [2] are relevant comparisons despite different data access. The additional evaluation identifies their target-side supervision and intervention locations; CAT is the text-only variant. We compare safety alongside benign utility and the fraction of prompts producing a valid image on the same generator. A shared coefficient or fixed operating point does not establish matched utility.

Native steering. The practical benefit of transfer is constructing useful interventions when target-side unsafe training pairs are unavailable. Native steering remains a strong reference when those pairs can be encoded by the target, and transfer does not consistently outperform it at matched utility. The comparisons quantify what can be achieved under this data constraint; reuse can also amortize the cost of benign alignment.

Comparison protocol. We report ASR on valid generated images, benign CLIP-Sim against the original prompt, relative CLIP loss, and image availability. At the fixed settings, transfer uses strength 5 for Schnell, Dev, and Qwen-Image and 3 for Z-Image. We then examine representative configurations and coefficient sweeps. SAFREE is evaluated on SDXL; AcT and CAT are evaluated on Schnell, Dev, Qwen-Image, and Z-Image.

Comparison with prompt rewriting and SAFREE [3].

Fixed operating points. Results are grouped by target. For Schnell, Dev, Qwen-Image, and Z-Image, Original and transfer rows report full-set results from the paper, whereas prompt rewriting is evaluated on a 300-prompt benign subset, with 292 valid outputs. Rewriting CLIP loss compares these outputs with unsteered images for exactly the same prompts and seeds, using the same recomputed CLIP metric for both conditions. The full-set CLIP-Sim values shown in the Original rows are therefore not the denominators used for rewriting loss. For example, on Qwen-Image the paired original and rewriting means are 0.324 and 0.323, yielding 0.43% loss. For SAFREE, loss likewise uses paired unsteered outputs restricted to valid benign pairs. Negative loss indicates higher mean similarity.

Schnell.

MethodStrength / thresholdASR (%) ↓CLIP-Sim ↑CLIP loss (%) ↓Valid images, I2P / benign (%)
Original030.70.3190.00100.0 / 100.0
Transfer Ridge511.00.3064.08100.0 / 100.0
Transfer SVD53.80.3083.45100.0 / 100.0
Transfer MLP511.40.3083.45100.0 / 100.0
LLM prompt rewriting—21.00.3160.4078.7 / 97.3

Dev.

MethodStrength / thresholdASR (%) ↓CLIP-Sim ↑CLIP loss (%) ↓Valid images, I2P / benign (%)
Original028.60.3090.00100.0 / 100.0
Transfer Ridge512.40.2973.88100.0 / 100.0
Transfer SVD53.50.2848.09100.0 / 100.0
Transfer MLP512.40.2983.56100.0 / 100.0
LLM prompt rewriting—20.10.3070.1477.1 / 97.3

Qwen-Image.

MethodStrength / thresholdASR (%) ↓CLIP-Sim ↑CLIP loss (%) ↓Valid images, I2P / benign (%)
Original038.40.3320.00100.0 / 100.0
Transfer Ridge517.10.3193.92100.0 / 100.0
Transfer SVD58.70.29710.57100.0 / 100.0
Transfer MLP518.40.3213.31100.0 / 100.0
LLM prompt rewriting—24.40.3230.4378.7 / 97.3

Z-Image.

MethodStrength / thresholdASR (%) ↓CLIP-Sim ↑CLIP loss (%) ↓Valid images, I2P / benign (%)
Original030.40.3190.00100.0 / 100.0
Transfer Ridge310.50.28710.03100.0 / 100.0
Transfer SVD30.20.24921.94100.0 / 100.0
Transfer MLP312.10.28610.34100.0 / 100.0
LLM prompt rewriting—18.00.3120.3678.7 / 97.3

SDXL.

MethodStrength / thresholdASR (%) ↓CLIP-Sim ↑CLIP loss (%) ↓Valid images, I2P / benign (%)
Original031.40.2870.00100.0 / 100.0
Transfer Ridge (three maps)324.30.2744.59100.0 / 100.0
Transfer SVD (three maps)317.10.2763.63100.0 / 100.0
SAFREE broadthreshold 0.0122.90.287-0.02100.0 / 100.0
SAFREE nuditythreshold 0.0125.70.2792.59100.0 / 100.0
SAFREE broadthreshold 0.0021.90.2870.1097.9 / 100.0
SAFREE nuditythreshold 0.0029.90.2812.1197.9 / 100.0
AuthorsOpenReview ↗

Response to Reviewer RUpw (Part 3)

Paired coefficient comparison. On Schnell, native and transfer strengths are selected from {1, 3} using benign calibration. AcT is shown at its smallest tested coefficient, 0.1, and CAT at coefficient 1. All conditions produce valid images. The full coefficient sweeps follow under their respective methods.

Schnell methodStrength / coefficientASR (%) ↓CLIP loss (%) ↓
Original027.90.00
Native311.42.89
Transfer Ridge312.92.38
Transfer SVD37.14.06
AcT diffusion blocks0.120.01.06
CAT text-only121.40.81

In the paired Schnell comparison, SVD gives 7.1% ASR with 4.06% CLIP loss, versus 20.0% / 1.06% for AcT and 21.4% / 0.81% for CAT. Transfer therefore offers greater suppression at these settings, with a larger utility cost. CLIP loss is relative to paired unsteered outputs for the same valid benign requests; negative values indicate higher mean similarity. These operating points do not imply equal utility across methods.

Since SDXL was not evaluated in the original manuscript, we adapt our method to its two text encoders, CLIP-L and CLIP-bigG, and the projected pooled output of CLIP-bigG, which also conditions generation. We fit three separate source-to-target maps on benign anchors, one for each component, and apply the mapped directions to their respective positive-conditioning components.

AcT inside diffusion blocks: effect of intervention strength.

We apply the official LinearOTHook inside the diffusion generator, fitting target-specific transformations on safe/unsafe pairs. FLUX follows the official AcT module selection; Qwen-Image and Z-Image adapt the intervention to their core diffusion blocks. These are architecture adaptations with target-side safety supervision.

AcT is fitted independently for each target using the same 1,000 safe/unsafe prompt pairs used to construct our safety directions. We report results for coefficients 0.1, 0.25, 0.5, and 1 without calibration-based selection. Evaluation uses 1,000 I2P prompts for ASR and 300 benign prompts for CLIP-Sim. The unsteered model (coefficient 0) is evaluated on these same prompt sets, providing a common reference for measuring safety improvements and benign-fidelity costs.

Schnell.

AcT coefficientASR (%) ↓Benign CLIP-Sim ↑CLIP loss (%) ↓
027.90.3140.00
0.120.00.3111.06
0.2517.10.23026.69
0.54.30.20235.51
10.00.20634.38

Dev.

AcT coefficientASR (%) ↓Benign CLIP-Sim ↑CLIP loss (%) ↓
030.70.3040.00
0.130.70.307-0.98
0.2529.30.2875.52
0.527.10.19735.13
10.00.19436.18

Qwen-Image.

AcT coefficientASR (%) ↓Benign CLIP-Sim ↑CLIP loss (%) ↓
035.00.3280.00
0.10.70.20836.81
0.250.00.20637.26
0.50.70.20437.86
10.00.20138.69

Z-Image.

AcT coefficientASR (%) ↓Benign CLIP-Sim ↑CLIP loss (%) ↓
023.60.3160.00
0.123.60.3121.14
0.2525.70.25718.79
0.55.70.20634.87
10.00.20235.99

On Schnell, coefficient 0.1 reduces ASR from 27.9% to 20.0% with 1.06% benign CLIP loss. Stronger interventions further reduce ASR but incur 26.69–35.51% CLIP loss. Thus, the 0.0% ASR at coefficient 1 comes with substantial loss of prompt fidelity and does not establish a favorable safety–utility trade-off.

Reviewer_zb8GOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether internal safety representations learned by a source language model can be effectively transferred to regulate the text-conditioning space of independant generative models. Recent studies focus either on safety alignment for specific model architectures or on achieving safety during the model training process. This paper goes beyond being limited to specific model architectures and aims to transfer the safety direction of the source LLM by training representation space mappings (SVD, Ridge, MLP) to other target generation models. The experiments were conducted using various T2I models and a single T2V model. And the privacy-utility trade-off varies depending on each mapping.

Critical strengths
  • The motivation of this paper is very clear, and I think the approach is meaningful. Many existing studies on safety alignment and guardrails are limited to specific model architectures. Although such model specific approaches can achieve strong performance, I agree with the direction of this work from the perspective of costs.
  • The proposed method is very intuitive. Authors extract a safety direction from a source LLM, transfer it into the target model's representation space through a mapping, and achieve meaningful safety improvements simply by adding the transferred vector during inference.
Critical weaknesses

Major

  • The proposed framework appears somewhat incremental, as it combines several techniques, including safety direction extraction, represenation alignment and activation steering.
  • This study also relies on activation steering, so a comparison with recent activation steering methods would be valuable. Of course, given the nature of this work, I would not necessarily expect it to outperform methods designed for specific architectures. But I believe such comparisons would serve as complementary evaluations. I do not intend to argue that authors should achieve SotA results.
  • There seems to be a gap between the title and the main claim of this paper. From the title, "Do Models Share Safety Representations?", readers might expect an investigation into whether safety representations are shared among different generative models. However, the main contribution focuses on transferring safety directions from a source LLM to the language space of a target generative model. Authors should more clearly distinguish these two interpretations and clarify the scope of the claimed shared safety representations.
  • Then, can the same alignment mapping and safety representation be directly reused across different generative models that share the same text encoder? If this is possible, it would provide additional evidence supporting the reusability of the proposed method. Authors have already evaluated two Flux models based on T5-XXL. So, I wonder whether the safety vector and alignment mapping obtained from a source LLM and Flux.1-schnell can be directly applied to Flux.1-dev without additional alignment.

Minor

  • It seems somewhat difficult to claim broad generalization to T2V models, as the evaluation covers only Wan 2.2. Furthermore, video generation involves safety risks arising from temporal dynamics that do not exist in images. It would be valuable to consider such temporal aspects rather than relying on frame-level safety evaluation.
  • From the Table 2, SVD shows a lower mean cos sim in terms of geometric agreement than Ridge and MLP. However, in Table 1, SVD achieves the lowest ASR. While I can understand that lower geometric agreement might be associated with degradation in generation quality, it is less clear why SVD achieves lower ASR than methods with higher cos sim.
Questions for authors

See the weaknesses.

7 comments
AuthorsOpenReview ↗

Response to Reviewer zb8G (Part 1)

We thank the Reviewer for recognizing the clear motivation and practical value of the approach. We appreciate the suggestions on direct reuse, comparative evaluation, and the scope of the claims.


[W1] Contribution beyond established components

The contribution is the separation of source-side safety supervision from target-side benign alignment, allowing a reusable map to transport interventions into frozen visual generators. The estimators and additive steering are established, but fitting benign anchors does not itself ensure that a source safety direction will usefully control visual generation. Our evaluation demonstrates this transfer and characterizes its utility cost. The SVD decomposition and Schnell→Dev reuse further explain how geometric agreement and shared encoders affect its behavior.


[W2] Comparison with recent activation-steering methods

We now include AcT [1] and CAT [2] as target-trained references, alongside SAFREE [3] and prompt rewriting as comparisons without target-side unsafe training pairs. The tables distinguish method variants, actual benign utility, and setup/inference costs. Fixed steering strengths alone do not constitute a matched-utility comparison.

Comparison protocol. We compare safety alongside benign prompt fidelity and image availability. Transfer uses strength 5 on Schnell, Dev, and Qwen-Image and 3 on Z-Image. AcT and CAT coefficient sweeps are reported below. Prompt rewriting uses greedy decoding. ASR is conditional on valid images, and CLIP-Sim is measured against the original prompt. AcT and CAT are evaluated on the four image targets; SAFREE is evaluated on SDXL.

Comparison with prompt rewriting and SAFREE. Results at the reported strengths are grouped by target.

Schnell.

MethodStrength / thresholdASR (%) ↓Benign CLIP-Sim ↑Valid images, I2P / benign (%)
Original030.70.319100.0 / 100.0
Transfer Ridge511.00.306100.0 / 100.0
Transfer SVD53.80.308100.0 / 100.0
Transfer MLP511.40.308100.0 / 100.0
LLM prompt rewriting—21.00.31678.7 / 97.3

Dev.

MethodStrength / thresholdASR (%) ↓Benign CLIP-Sim ↑Valid images, I2P / benign (%)
Original028.60.309100.0 / 100.0
Transfer Ridge512.40.297100.0 / 100.0
Transfer SVD53.50.284100.0 / 100.0
Transfer MLP512.40.298100.0 / 100.0
LLM prompt rewriting—20.10.30777.1 / 97.3

Qwen-Image.

MethodStrength / thresholdASR (%) ↓Benign CLIP-Sim ↑Valid images, I2P / benign (%)
Original038.40.332100.0 / 100.0
Transfer Ridge517.10.319100.0 / 100.0
Transfer SVD58.70.297100.0 / 100.0
Transfer MLP518.40.321100.0 / 100.0
LLM prompt rewriting—24.40.32378.7 / 97.3

Z-Image.

MethodStrength / thresholdASR (%) ↓Benign CLIP-Sim ↑Valid images, I2P / benign (%)
Original030.40.319100.0 / 100.0
Transfer Ridge310.50.287100.0 / 100.0
Transfer SVD30.20.249100.0 / 100.0
Transfer MLP312.10.286100.0 / 100.0
LLM prompt rewriting—18.00.31278.7 / 97.3

SDXL.

MethodStrength / thresholdASR (%) ↓Benign CLIP-Sim ↑Valid images, I2P / benign (%)
Original031.40.287100.0 / 100.0
Transfer Ridge (three maps)324.30.274100.0 / 100.0
Transfer SVD (three maps)317.10.276100.0 / 100.0
SAFREE broadthreshold 0.0122.90.287100.0 / 100.0
SAFREE nuditythreshold 0.0125.70.279100.0 / 100.0
SAFREE broadthreshold 0.0021.90.28797.9 / 100.0
SAFREE nuditythreshold 0.0029.90.28197.9 / 100.0

On Schnell, Dev, Qwen-Image, and Z-Image, transfer achieves lower ASR at the displayed strengths, while rewriting largely preserves benign similarity and leaves some requests without an image. These results expose the practical safety–utility trade-off rather than imply equal utility across methods.

Since SDXL was not evaluated in the original manuscript, we adapt our method to its two text encoders, CLIP-L and CLIP-bigG, and the projected pooled output of CLIP-bigG, which also conditions generation. We fit three separate source-to-target maps on benign anchors, one for each component, and apply the mapped directions to their respective positive-conditioning components.

AuthorsOpenReview ↗

Response to Reviewer zb8G (Part 2)

AcT inside diffusion blocks: effect of intervention strength.

We apply the official LinearOTHook inside the diffusion generator, fitting target-specific transformations on safe/unsafe pairs. FLUX uses the official AcT module selection; Qwen-Image and Z-Image adapt the intervention to their core diffusion blocks. These are adaptations of AcT's intervention approach, with target-side safety supervision.

For each target, we fit AcT using the same 1,000 safe/unsafe prompt pairs used to estimate our safety directions. We evaluate coefficients 0.1, 0.25, 0.5, and 1 without a calibration filter, measuring ASR on 1,000 I2P prompts and CLIP-Sim on 300 benign prompts. The unsteered reference (coefficient 0) is evaluated on the same prompts, allowing safety gains and benign-fidelity costs to be assessed within this evaluation subset.

Schnell.

AcT coefficientASR (%) ↓Benign CLIP-Sim ↑CLIP loss (%) ↓
027.90.3140.00
0.120.00.3111.06
0.2517.10.23026.69
0.54.30.20235.51
10.00.20634.38

Dev.

AcT coefficientASR (%) ↓Benign CLIP-Sim ↑CLIP loss (%) ↓
030.70.3040.00
0.130.70.307-0.98
0.2529.30.2875.52
0.527.10.19735.13
10.00.19436.18

Qwen-Image.

AcT coefficientASR (%) ↓Benign CLIP-Sim ↑CLIP loss (%) ↓
035.00.3280.00
0.10.70.20836.81
0.250.00.20637.26
0.50.70.20437.86
10.00.20138.69

Z-Image.

AcT coefficientASR (%) ↓Benign CLIP-Sim ↑CLIP loss (%) ↓
023.60.3160.00
0.123.60.3121.14
0.2525.70.25718.79
0.55.70.20634.87
10.00.20235.99

On Schnell, coefficient 0.1 reduces ASR from 27.9% to 20.0% with 1.06% benign CLIP loss. Stronger interventions further reduce ASR but incur 26.69–35.51% CLIP loss. Thus, the 0.0% ASR at coefficient 1 comes with substantial loss of prompt fidelity and does not establish a favorable safety–utility trade-off.

CAT in text encoders: effect of intervention strength.

We adapt CAT to the second half of CLIP and T5 in FLUX and of the language encoder in Qwen-Image. Each target uses residual MLP maps fitted on safe/unsafe pairs, with safe-identity regularization 0.5 and no additional conditioning gate. Z-Image uses a different text-only configuration: regularization 0 and Mahalanobis threshold 0.25.

For Schnell, Dev, and Qwen-Image, all four CAT coefficients are reported without coefficient selection. For Z-Image, test results are available at 0.1 and 1. Negative CLIP loss indicates higher mean similarity. CAT is evaluated on the same 1,000 I2P prompts and 300 benign prompts used for the AcT sweeps.

Schnell.

CAT coefficientASR (%) ↓Benign CLIP-Sim ↑CLIP loss (%) ↓
027.90.3140.00
0.125.70.3140.00
0.2521.40.315-0.22
0.522.10.3130.27
121.40.3110.81

Dev.

CAT coefficientASR (%) ↓Benign CLIP-Sim ↑CLIP loss (%) ↓
030.70.3040.00
0.130.70.3030.29
0.2527.90.3010.94
0.526.40.3030.23
123.60.3020.60

Qwen-Image.

CAT coefficientASR (%) ↓Benign CLIP-Sim ↑CLIP loss (%) ↓
035.00.3280.00
0.137.10.329-0.20
0.2538.60.3280.22
0.536.40.332-1.12
137.10.3270.49

Z-Image.

CAT coefficientASR (%) ↓Benign CLIP-Sim ↑CLIP loss (%) ↓
023.60.3160.00
0.122.10.3140.73
112.90.27812.02

CAT reduces ASR on both FLUX models with little change in benign CLIP-Sim. At coefficient 1, Schnell improves from 27.9% to 21.4% ASR and Dev from 30.7% to 23.6%, with 0.81% and 0.60% CLIP loss, respectively. Schnell also reaches 21.4% ASR at coefficient 0.25, with -0.22% CLIP loss. On Qwen-Image, no tested coefficient improves the unsteered 35.0% ASR (36.4–38.6% after steering).

For Z-Image, CAT coefficient 0.1 is selected using benign calibration and gives 22.1% ASR with 0.73% CLIP loss. Coefficient 1 gives 12.9% ASR with 12.02% loss. Coefficients 0.25 and 0.5 were evaluated only during calibration, so no test ASR is reported. This illustrates the benefit of calibrating CAT strength to limit benign degradation.

AuthorsOpenReview ↗

Response to Reviewer zb8G (Part 3)

Paired Schnell comparison.

Schnell methodStrength / coefficientASR (%) ↓CLIP loss (%) ↓
Original027.90.00
Native311.42.89
Transfer Ridge312.92.38
Transfer SVD37.14.06
AcT diffusion blocks0.120.01.06
CAT text-only121.40.81

Native and transfer strengths are selected from {1, 3} using benign calibration. AcT is shown at 0.1, its smallest tested coefficient and the setting with the lowest observed CLIP loss on Schnell; CAT uses coefficient 1 as a fixed reference. Their full sweeps appear above. SVD reaches 7.1% ASR with 4.06% CLIP loss, compared with 20.0% for AcT and 21.4% for CAT. The baselines preserve more benign similarity, while transfer offers greater suppression at these settings without target-side unsafe training pairs.

Steering and prompt-rewriting costs.

TargetAcT block preparation, summed worker time (s)AcT blocks, coefficient 1 (s/image)LLM rewriting + generation (s/image)CAT text-only, strength 1 (s/image)
Schnell309.862.522.872.65
Dev4,225.2731.5330.6530.48
Qwen-Image844.295.405.124.89
Z-Image527.254.404.413.78

Generation times are medians over valid outputs on an NVIDIA A100, excluding model loading and scoring. AcT preparation includes activation extraction and fitting, excluding generator loading; it sums worker times rather than elapsed parallel runtime. Rewriting includes the LLM call and image generation and requires no fitting. Transfer adds a saved vector during inference; its alignment cost and reuse saving are reported in the Schnell→Dev comparison below.


[W3] Title and scope of shared safety representations

The title motivates the question of whether safety-relevant interventions remain actionable across models. Our results demonstrate functional transfer, while their safety specificity remains unresolved. We will distinguish these two interpretations explicitly in the manuscript, preserving the demonstrated transfer contribution without asserting a universal safety representation.


[W4] Direct reuse from FLUX.1-schnell to FLUX.1-dev

Direct reuse. Yes. The two generators share their T5 and CLIP encoder weights and tokenizers. We apply the Schnell Ridge map and its transferred safety vector unchanged to Dev, without refitting or rescaling. The intervention acts on T5 conditioning. For comparison, we independently fit a Dev map using benign anchors only.

Evaluation protocol. We retain the original Llama-3-8B-Instruct checkpoint, representation extraction, and anchor-based scaling. Independently fitted vectors have cosine similarity 0.999987 despite the original numerical precisions differing between Schnell (float16) and Dev (bfloat16). We compare unchanged reuse with independent fitting at α=3\alpha=3 and α=5\alpha=5. This experiment uses a dedicated evaluation subset of 1,000 I2P prompts for ASR and 150 benign prompts for CLIP-Sim and relative CLIP loss. All models and interventions are evaluated on the same prompts using a single generation seed (42). Steering strength α is selected using an additional set of 100 benign calibration prompts, disjoint from the test set.

Target and interventionα\alphaDetector-ASR (%) ↓\downarrowBenign CLIP loss (%) ↓\downarrow
Schnell, original030.00.00
Schnell, transferred Ridge317.31.19
Schnell, transferred Ridge511.02.42
Dev, original026.70.00
Dev, unchanged Schnell vector314.01.77
Dev, independently fitted Ridge map312.01.75
Dev, unchanged Schnell vector57.34.25
Dev, independently fitted Ridge map58.04.15

At α=3\alpha=3, unchanged reuse preserves a 12.7-point ASR reduction on both generators; refitting on Dev improves it by a further 2.0 points. At α=5\alpha=5, reuse gives 0.7 points lower ASR than refitting. Thus, the Reviewer's proposed reuse is effective without additional alignment, with no consistent advantage from refitting at the reported strengths.

Alignment cost. Reuse avoids Dev anchor extraction and map fitting.

Dev alignmentBenign anchor extraction (s)Ridge fitting and scaling (s)
Independent fit238.644.22
Reuse Schnell bundleNot requiredNot required

These timings exclude model loading, source preparation, utility calibration, and image generation.

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

This paper investigates whether safety directions from a source language model can be transferred to image and video generators. The authors construct directions from paired safe and unsafe prompts in the source model, then map them to a target text encoder using a cross-model mapping learned solely from benign text anchors. The transferred directions are applied without updating the generator's parameters. The paper compares mapping methods, global and category-specific directions, and steering strengths, evaluating safety and benign generation quality on multiple image generators and one video generator. It also explores selecting a deployment strength using benign data alone.

Critical strengths
  1. Useful restricted-data setting. The paper explores reusing safety directions when unsafe pairs are unavailable on the target side, keeping those pairs in the source environment and using only benign data to align with the target encoder.

  2. Broad cross-model evaluation. The experiments cover multiple source models, image generators, and a video generator, and include native target-side directions, utility-matched comparisons, benign-only selection, and a blinded vision-language model evaluation. These results support safety direction transfer in the evaluated settings.

Critical weaknesses
  1. Limited comparison under restricted access. SAFREE [1] adaptively filters unsafe concepts without target-side safe–unsafe pairs, while the paper's restricted-access baseline concatenates six category names into one vector. Their inputs and interventions differ, but without a matched-utility comparison on compatible models, the safety benefits and costs of transfer remain unclear.

  2. Safety under benign-only selection. Under a 5% CLIP-Sim loss budget, detector-based attack success rate (ASR) decreases in all 45 tested configurations, but selection relies only on benign CLIP-Sim loss and its uncertainty. Safety gains and residual risk on new models or risk categories therefore remain unclear.

  3. Benign content preservation. The global direction may suppress benign content related to safety categories. The targeted evaluation uses 60 prompts, one generation seed, and CLIP-Sim, leaving it unclear whether required objects, actions, or scene attributes are removed or altered.

Reference

[1] Jaehong Yoon, Shoubin Yu, Vaidehi Patil, Huaxiu Yao, and Mohit Bansal. SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation. ICLR 2025. https://arxiv.org/abs/2410.12761

Questions for authors
  1. At matched benign utility, how does transfer compare with SAFREE and similar methods in safety and computational or deployment costs on compatible models?

  2. Without target-side unsafe data, what can be inferred about the safety gains and residual risk of benign-only selection across models and risk categories?

  3. How do category directions affect target and other categories, how do selection errors affect safety and benign quality, and does global steering preserve task-relevant benign objects and attributes?

  4. What do transfer and agreement with native directions reveal about shared safety representations and the role of general semantics, such as unsafe concepts, in steering?

4 comments
AuthorsOpenReview ↗

Response to Reviewer XdfN (Part 1)

We thank the Reviewer for recognizing the value of constructing interventions without target-side unsafe training pairs and the breadth of the evaluation. We appreciate the questions on practical comparisons, benign preservation, and the interpretation of transfer.


[W1/Q1] Comparison with SAFREE and deployment costs

Comparison protocol. We agree that the category-string baseline does not replace SAFREE. We therefore compare transferred Ridge/SVD directions with SAFREE [1] on SDXL under shared generation settings. Both SAFREE concept lists are evaluated at the default token-selection threshold sf_alpha=0.01 and at 0.00. We report safety, benign utility, and deployment costs.

Deployment cost. Transfer requires anchor extraction and fitting, but inference only adds a saved vector to the target conditioning states. In the controlled Schnell→Dev experiment, reuse avoids 238.64 seconds of Dev anchor extraction and 4.22 seconds of Ridge fitting and scaling. Thus, shared encoders allow the initial alignment cost to be amortized across generators; target-specific benign calibration remains necessary.

Three mappings for SDXL. Since SDXL was not evaluated in the original manuscript, we explain here how we adapt our method to its conditioning architecture. SDXL uses two text encoders, CLIP-L and CLIP-bigG, whose token representations are concatenated. The projected pooled output of CLIP-bigG also conditions image generation. We therefore learn three independent mappings from the source representation space: one to CLIP-L token representations, one to CLIP-bigG token representations, and one to the projected pooled output of CLIP-bigG. All three maps are fitted using benign anchors. The two mapped token directions are concatenated and added to the positive token conditioning; the third direction is added to the positive pooled conditioning. Negative conditioning is unchanged.

SDXL comparison. The table reports I2P ASR, benign CLIP-Sim, and relative benign CLIP loss. For each method, CLIP loss is computed as 100 × (1 − method mean / paired original mean), using the same benign prompts and seeds for both means and retaining only pairs with valid images. Negative loss indicates higher mean similarity.

MethodSteering strength / SAFREE thresholdI2P ASR (%) ↓Benign CLIP-SimBenign CLIP loss (%)
Original031.40.2870.00
Transfer Ridge324.30.2744.59
Transfer SVD317.10.2763.63
SAFREE broadthreshold 0.0122.90.287-0.02
SAFREE nuditythreshold 0.0125.70.2792.59
SAFREE broadthreshold 0.0021.90.2870.10
SAFREE nuditythreshold 0.0029.90.2812.11

At threshold 0.00, SAFREE broad and nudity each produce invalid images for 2.1% of I2P requests and none of the benign requests. At threshold 0.01, both produce valid images for all test requests. ASR and CLIP-Sim are computed only on valid outputs.

Safety and utility. SVD and Ridge reduce ASR from 31.4% to 17.1% and 24.3%, respectively, with benign CLIP losses of 3.63% and 4.59%. SAFREE broad reaches 22.9% ASR at its default threshold while preserving mean similarity. Lowering the threshold to 0.00 gives 21.9% conditional ASR with more invalid outputs. These results support transfer as a practical alternative, with a different safety–utility trade-off.

Under a common 5% benign calibration budget, SVD at strength 3 reduces ASR from 31.4% to 17.1% with 3.63% CLIP loss; SAFREE broad gives 22.9% ASR with −0.02% loss. SVD has the lower ASR point estimate within this budget, although the paired difference of −5.7 percentage points (95% interval: −12.9 to +1.4) leaves the ranking uncertain. Ridge at its selected strength 1 gives 32.1% ASR with 0.09% loss, illustrating that the mapping and selected strength both matter.

SDXL setup and inference costs.

MethodTarget-side unsafe training pairsEmbedding preparation (s)Mapping fitting (s)Median generation (s/image)
OriginalNo——2.98
Transfer RidgeNo300.410.783.06
Transfer SVDNo300.410.693.07
SAFREE broad, threshold 0.01No——3.21
SAFREE nudity, threshold 0.01No——3.06

Timings are measured on an NVIDIA A100. Embedding preparation includes model loading and extraction of anchor representations; source and token-encoder representations are reused. Fitting sums the three corresponding fits, including Ridge validation. Generation latency includes the active intervention.

AuthorsOpenReview ↗

Response to Reviewer XdfN (Part 2)

[W2/Q2] Selecting strength on benign prompts and evaluating safety

Selection and evaluation are separate. Benign-only selection determines the steering strength; the safety benefit is measured through unsafe-output evaluation. For each target and direction, we select the largest tested α\alpha whose one-sided 95% upper confidence bound on benign CLIP loss is within 5%. With α\alpha fixed, we evaluate safety and verify benign utility. Calibration is repeated for each target; the manuscript's Llama-3.1/Ridge results in Table 3 illustrate this procedure.

Interpretation and remaining risk. ASR decreases in all 45 manuscript source–target–mapping configurations after benign-only strength selection. Benign verification remains within 5% in 44/45 cases, with 5.25% loss in the remaining case. This is consistent empirical evidence that useful interventions can be selected without unsafe target prompts for calibration. Safety improvement is established by the output evaluation, not by CLIP similarity alone; it must still be measured for a new model or risk category.

Additional video evidence. On Bernini-R, strength 3 is selected using benign calibration. Detector-ASR then decreases from 51.7% to 39.2%, and Qwen3.5-9B [2] VLM-ASR from 40.8% to 30.8% (10.0-point reduction; 95% paired interval: 3.3–16.7), with −0.21% benign CLIP loss. This extends the observed benefit to an additional video target while making the residual risk explicit.


[W3/Q3] Benign preservation, category effects, and selection errors

Benign attribute preservation: human evaluation. To check whether steering removes required objects or actions, a human annotator assesses original and SVD-steered images. This covers 210 attributes in 70 Schnell examples (40 sensitive-benign, 30 ordinary) and 180 attributes in 60 sensitive-benign Z-Image examples. Each required attribute is marked present, absent, or uncertain. Percentages refer to the attributes required by the prompt in each condition.

ModelInterventionSteering strength α\alphaPresent attributes (%)Absent attributes (%)Uncertain attributes (%)
SchnellWithout steering097.12.90.0
SchnellTransferred SVD steering595.24.30.5
Z-ImageWithout steering096.73.30.0
Z-ImageTransferred SVD steering388.910.01.1

After steering, 95.2% of required attributes are present on Schnell and 88.9% on Z-Image, versus 97.1% and 96.7% originally. Most required objects and actions therefore remain visible, providing evidence of prompt-relevant content preservation beyond aggregate CLIP similarity. The larger decrease on Z-Image also identifies a meaningful benign-content cost.

Category effects. We evaluate each category-conditioned direction across all six prompt categories and assess benign utility on a common set of prompts.

Cross-category effects.

Each direction is applied separately on Schnell with Ridge at α=5\alpha=5, with magnitudes scaled using centered benign anchors. We evaluate each direction across all six I2P prompt categories and measure benign utility relative to paired original images.

Steering directionSexualIllegal activityViolenceHateHarassmentShockingBenign CLIP-Sim loss (%)
Original ASR (%, reference)34.421.934.434.431.331.3—
Sexual12.53.115.63.112.512.53.42
Illegal activity18.812.525.012.515.621.91.18
Violence12.53.121.96.318.818.81.55
Hate12.56.318.812.518.821.93.83
Humiliation18.86.318.818.815.621.94.20
Disturbing9.43.125.06.315.618.83.57

The six category columns report ASR (%; lower is better): the first row is the unsteered reference, and each subsequent row gives ASR after applying that direction. ASR measures overall detected unsafe content within each prompt category. Harassment/Shocking are I2P labels associated with the Humiliation/Disturbing directions; the two taxonomies are not identical.

Every direction reduces ASR on its corresponding prompt category by 9.4–21.9 percentage points, with benign CLIP losses of 1.18–4.20%. Reductions on other categories are often comparable or larger, showing that the effects overlap.

This overlap should be interpreted in light of the prompts themselves. Both the SafeSteer prompts used to construct directions and the I2P prompts used to generate images often combine multiple safety-related concepts, making it difficult to isolate a single category. Even the single recorded label of each selected I2P prompt does not imply that its content belongs exclusively to that category. Such semantic overlap can contribute to each vector affecting other categories. We therefore describe these as category-conditioned directions, without claiming independent category-selective control.

AuthorsOpenReview ↗

Response to Reviewer XdfN (Part 3)

Sensitivity to simulated category-selection errors. Category selection is external to our steering mechanism: a user or an upstream classifier chooses which directions to apply. To address the Reviewer's question, we isolate three possible errors: missing an intervention, selecting another category, and unnecessarily steering a benign prompt. In the single-category setting above, the reference policy selects the recorded category for risky prompts and leaves benign prompts unsteered. Its ASR is 15.6%, compared with 31.3% without steering.

Using the cross-category outputs, we simulate errors affecting 25% or 50% of the relevant requests while keeping other decisions at the reference policy. Wrong categories are chosen uniformly from the five alternatives; unnecessary benign interventions use a uniformly chosen direction from all six. The table reports mean outcomes across random assignments.

Simulated selection errorMetric (%) ↓\downarrow25% affected50% affected
Skip steering on risky promptsASR on all risky prompts19.523.4
Select another category for risky promptsASR on all risky prompts15.214.8
Unnecessarily steer benign promptsCLIP loss on affected benign prompts only2.942.94

Missing half the interventions approximately halves the ASR reduction. Replacing the selected category leaves overall ASR close to the reference, consistent with the overlapping effects in the preceding matrix; this does not establish suppression of the intended risk category. Unnecessary benign intervention incurs about 2.94% mean CLIP loss on the prompts actually affected. Averaged over the entire benign set, including untouched prompts, the loss is 0.74%/1.47% at the two error rates. This distinguishes the cost to affected requests from its average impact across requests.

These results quantify sensitivity to the specified selection errors on fixed outputs. They do not measure an upstream classifier's accuracy or establish robustness to errors concentrated on difficult prompts.


[Q4] Shared safety representations and general semantics

Transferred directions align closely with native target directions: mean cosine similarity is 0.961 for Ridge and 0.949 for MLP, while the 95th percentile of the absolute cosine for isotropic random directions is below 0.039 (Appendix C.1). This shows that the source-derived contrast retains a consistent orientation after alignment across independently trained encoders. It does not establish a content-independent safety representation. Nor is native agreement sufficient to explain effectiveness: SVD has lower mean cosine (0.718) yet often achieves stronger suppression, and its component orthogonal to the estimated native direction also reduces ASR at limited utility cost (response to Reviewer Vfgh, Q3).

Safe/unsafe prompt pairs inevitably differ in content (for example, clothing versus nudity) and our experiments do not disentangle that content from safety. Norm-matched random directions, category-name steering, and tests on benign prompts containing related concepts argue against arbitrary perturbation and indiscriminate concept removal (response to Reviewer RUpw, W1/Q1). They do not rule out richer semantic explanations. We will therefore frame our conclusion as the transfer of safety-relevant interventions, and make this distinction explicit in the title and limitations.

References

[1] Jaehong Yoon, Shoubin Yu, Vaidehi Patil, Huaxiu Yao, and Mohit Bansal. SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation. International Conference on Learning Representations (ICLR), 2025.

[2] Qwen Team. Qwen3.5: Towards Native Multimodal Agents. Qwen technical blog, 2026.

Thank the authors for the detailed response. My concerns have been largely addressed. Therefore, I keep my positive overall assessment.

Comments on the paper

2 comments

We thank all reviewers for their thoughtful feedback and constructive suggestions. We have posted individual responses addressing their comments, including additional experiments and analyses. We are incorporating these results and clarifications into the paper and will upload a revised version in the coming days. In the meantime, we remain available for any further questions or clarifications.

AuthorsOpenReview ↗

Post-rebuttal revisions

We thank all four reviewers for their careful reading and constructive feedback. Their questions helped us clarify what the results show about shared safety representations. Their suggestions also led us to compare additional safeguards, test whether Schnell’s steering vector could be reused on Dev, and check which attributes of benign prompts remain visible after steering. We thank zb8G and RUpw for raising their scores to "Weak acceptance" following the rebuttal, and XdfN and Vfgh for maintaining their positive assessments.

  • Scope of the claim. We revised the abstract, conclusion, and limitations to clarify what transfers across models. A source-derived direction can reduce unsafe outputs after alignment. How far this direction can be separated from the semantic content of the prompts remains open. (Sections 1 and 5, Appendix E)
  • New baselines. We added AcT and CAT on all four image generators, SAFREE on SDXL, and LLM prompt rewriting. (Section 4.2, Appendix C.12)
  • SVD analysis. The component of the transferred SVD direction orthogonal to the native direction also reduces attack success rate (ASR). The main text now reports the random-direction controls and the fidelity cost of steering at lower strength. (Section 4.2, Appendix C.13)
  • Reuse across generators. Applying the Schnell map and vector directly to Dev, which shares its text encoders, reduces Dev ASR from 26.7% to 14.0%. (Appendix C.14)
  • Video evaluation. We added Bernini-R and evaluated it alongside Wan2.2 with an additional VLM judge. On Bernini-R, detector ASR falls after steering with both 4 and 16 sampled frames. (Section 4.3, Appendix C.15)
  • Human evaluation. After SVD steering, 95.2% of attributes requested in benign prompts remain visible on Schnell and 88.9% on Z-Image. (Section 4.4, Appendix C.16)
  • Category overlap and selection errors. We measured how each direction affects all six risk categories. We also simulated choosing the wrong direction, skipping steering on a risky prompt, and steering a benign prompt unnecessarily. (Appendix C.17)
  • Additional targets. We extended transfer to text generation with Mistral-7B and Qwen3.5-9B, and to image generation with the Gemma2 encoder in Lumina-Image 2.0. (Appendices C.18–C.19)

We also updated the references and the AI Use Statement to document the new comparisons and automated evaluations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.