acceptodds
Under review as a conference paper at ICLR 2027

One Calibration, Many Requests: Amortized Black-Box Multimodal Jailbreaking via Cross-Model Refusal Coordinates

Abstract

Multimodal foundation models couple visual and textual information, which can make jailbreak behavior transferable across architectures. Many universal attacks rely on surrogate-side optimization without target-specific adaptation, whereas query-based black-box methods often require fresh target queries for every request. We study whether open models can provide a reusable low-dimensional interface for a stricter setting: one calibration on a closed target, followed by reuse of the same multimodal carrier on unseen requests. We propose UniCal, a black-box multimodal jailbreak method that realizes this calibrate-once, reuse-many paradigm. UniCal first aligns refusal and compliance structures from heterogeneous surrogates into a shared coordinate space and tests whether the learned coordinates predict refusal and compliance behavior under leave-one-model-out registration. A coordinate-conditioned generator then maps a fixed benign seed and a low-dimensional coordinate to a request-independent image and text carrier. For a new target, the pretrained generator is frozen, and only a sparse coordinate offset is selected through an evaluator-mediated three-way decision-label interface. The selected carrier is then reused without per-request updates. Across the five coordinate-pool models, UniCal achieves 89.6% mean strict success under one-time decision-label adaptation. Without any target-side calibration, the same surrogate-prepared carrier transfers to the two external unseen architectures with 80.1% strict success on Molmo and 77.9% on Phi-Vision. On closed targets, independently calibrating each dated API evaluation yields 81.7% mean strict success, retains 93.5% benign utility, and amortizes the 192-query calibration after eight reused requests against label-only Random Search.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
multimodal jailbreak, safety alignment, multimodal large language models

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 3 reviews, above 85% of the others.
Reviewer_RjGEOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a jailbreak method for vision-lanuage models, called UniCal. The authors claimed that most universal jailbreak attacks rely on surrogate-side optimization and query-based attacks often require fresh target queries for every request. Therefore these methods face a trade-off between reuse and target-specific adaptation. To address the limitation, UniCal moves rarget adaptation from the high-dimensional input space to a reusable low-dimensional behavioral coordinate space. First, a held-out-validated cross-model refusal and compliance coordinate space is built, which aligns heterogeneous surrogate states into a shared low-dimensional representation. Second, a request-independent generator maps shared coordinates to image-text carriers, allowing the attack to be calibrated once to a new target. Experimental results on MM-SafetyBench, USB, and VLJailbreakBench show that UniCal supports jailbreak reuse across unseen requests and models with a single calibration.

Critical strengths
  1. The three contributions in this paper are well-structured and the motivation is clear, which is to build a black-box jailbreak attack that is adaptable across many unseen requests.
  2. The authors observed refusal and compliance may admit a compact behavioral representation that can be aligned across models. Then they applied this observation to UniCal and built a reusable low-dimensional behavioral coordinate space. The idea is reasonal and will inspire subsequent jailbreak method development.
  3. This paper covers a wide range of comparison methods, such as FigStep, MML, JailBound, ZO-SPSA, and GAMBIT. And these methods are very new. Under this setup, the experimental results in this paper are convincing.
  4. The method section in this paper is detailed. From compliance coordinates building, universal carrier designing, and decision-label calibration, the authors give key formulations and implementation details. It benefits the reimplementation of this method and the reproduction of the experimental results.
Critical weaknesses
  1. In the experiment section, there are two main tables to show the quantitative results but the corresponding analysis is insufficient. The experimental analysis about Table 1-2 and Figure 3-5 only have one paragraph with four sentences. The reader can get very little information from this. And these analyses are very simple, doesn't have any deep insights and findings. This weakness causes this paper not readable.
  2. The loss function for generator training contains five terms with four trade-off coefficients lambda(s). But the paper does not give the value of each lambda. This is a critical missing for the experiment.
  3. The authors claimed that transfer-based attacks remain fixed when tranferred to a new target, but did not discuss why they are not good. In other words, what are the limitations of transfer-based methods? Have the proposed method (UniCal) addressed these limitations?
  4. The proposed method, UniCal, does not realize state-of-the-art results compared other methods mentioned in the paper. So I think this method does not bring any performance imporvement in this field, which weakens the paper's contribution.
  5. In the method comparison tables, such as Table 1 and Table 2, there are many numbers. Why don't the authors mark the best resutls with highlighted font or color? It is a good way to improve the readability of research papers.
  6. How many categories does the jailbreak have in authors' mind? Two (transfer-based and query-based) or three (transfer-based, query-based, and universal attack)? How should the comparison methods be classified?
  7. Can the learned shared coordinates be visualizaed? I want to know if it is readable and meaningful for human.
Questions for authors
  1. The insights and findings from the experimental analyzes.
  2. What limitations has the proposed method addressed?
  3. The details about the four trade-off coefficients.
7 comments
AuthorsOpenReview ↗

Response to Reviewer RjGE

We sincerely thank the reviewer for the encouraging assessment of UniCal's motivation, experimental breadth, and methodological detail. We especially appreciate the suggestions to draw clearer findings from the experiments and improve reproducibility. We have refined the presentation accordingly.

1. Experimental insights and performance (W1, W4).

The results reveal three complementary findings. First, one-time calibration improves strict success from 76.6% to 81.7% while retaining fixed-carrier reuse (Table 3). Independently, a carrier prepared on surrogates reaches 80.1% on Molmo and 77.9% on Phi-Vision without target feedback (Table 1), providing evidence of zero-feedback transfer to unseen architectures.

Second, shared behavioral coordinates are useful beyond representation-level prediction. They achieve 0.859 held-out refusal–compliance AUC versus 0.737 for PCA (Figure 5), and 81.7% versus 77.4% strict success in the calibrated attack pipeline (Table 3). Third, benign-utility preservation produces a meaningful trade-off: removing its loss raises strict success slightly from 81.7% to 82.4%, but lowers benign utility from 93.5% to 85.6% (Table 4). Together, these results illuminate the roles of calibration, behavioral parameterization, and utility preservation.

Table 2 also reports higher raw success for certain native-access baselines, including PolyJailbreak (93.2% versus 81.7%). We present these effectiveness results alongside a separate, matched label-only efficiency comparison. In that comparison, UniCal's 192-query calibration reaches its break-even point after eight reused requests relative to label-only Random Search (Figure 4). Section 5 now draws these findings and their evaluation conditions together more explicitly.

2. Generator coefficients and reproducibility (W2).

We appreciate the request for explicit coefficient values. Appendix A.2 now reports λ_b = 1.0, λ_v = 0.5, λ_t = 0.05, and λ_u = 1.0. In Equation (7), the coordinate-realization term has unit weight; these four coefficients weight the behavioral, visual, textual, and benign-utility losses, respectively. This specifies the weighting of the five-term generator objective.

3. Motivation, taxonomy, visualization, and tables (W3, W5–W7).

Zero-feedback transfer enables immediate reuse without incorporating feedback from a new target. Target-interactive methods can adapt to that target, sometimes with repeated request-level optimization. UniCal studies one label-only calibration followed by unchanged carrier reuse, combining target-specific adaptation with reuse across unseen requests. These are not three mutually exclusive attack categories: transfer versus target interaction concerns feedback access, while universal versus request-specific concerns what is reused.

We have clarified the access and adaptation regimes in the text and improved the presentation of Tables 1–2, including the identification of leading results. For interpretability, Figure 5(a) shows the first two dimensions of the eight-dimensional shared space, with model identities and refusal/compliance labels. Figure 5(b–c) further shows cross-model alignment and held-out predictability. These views make the organization of learned behavior visible at the distributional level, without assigning a separate semantic name to every coordinate axis.

4. Ethical considerations.

We appreciate the reviewer's attention to dual-use risks. We have added Section 7, Ethical Considerations, describing the potential for misuse and our reporting of aggregate evaluation results rather than harmful model completions. This positions the work as an analysis of multimodal safety failure modes intended to inform stronger safeguards.

We sincerely appreciate the reviewer's constructive assessment. We hope the clearer experimental findings, explicit training coefficients, and revised presentation help address the concerns, and would be grateful if the reviewer could reconsider the overall assessment in light of these revisions.

AuthorsOpenReview ↗

We Are Truly Grateful to Reviewer RjGE for the Thoughtful Comments and Generous Scholarly Guidance

Dear Reviewer RjGE,

Thank you again for your encouraging and constructive feedback. As the discussion period is approaching its deadline, we would be grateful if you could revisit our responses and revisions and share whether your concerns have been adequately addressed. We warmly welcome any further questions or discussion.

We sincerely appreciate the time and care you have devoted to reviewing our work and hope you have a wonderful life filled with happiness!

Best regards,

The Authors

Thanks for the rebuttal. The response has answered most of questions and concerns I rised. I still have one question: what limitations does this paper address? Please clarify them clearly.

Reviewer_R4khOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies black-box multimodal jailbreaking where an attack is calibrated once on a target model and then reused across unseen requests. The proposed method, UniCal, aligns refusal and compliance representations across open surrogate models and trains a generator to produce a request-independent image–text carrier. Target adaptation searches only a low-dimensional coordinate offset using coarse decision labels, while keeping the generator fixed. Experiments on open and closed models demonstrate effective reuse and transfer, but performance is subpar comparing to some baselines.

Critical strengths
  1. This paper addresses a practically meaningful problem: adapting a multimodal jailbreak to a black-box target once and reusing it across unseen requests. This setting makes the trade-off between attack effectiveness and repeated target-query costs explicit.

  2. The shared refusal–compliance coordinate space provides an appealing connection between representation analysis and black-box adaptation. Using this space to guide a request-independent carrier generator gives the method a coherent rationale beyond directly searching over image pixels or text.

Critical weaknesses
  1. Some important baselines are missing, such as HADES [1], and FORCE [2].

  2. While I do agree this paper has its technical novelty, I also feel the novelty is weaker than its currently claimed, specifically, this paper did not mentioned its linage to the representation similarlty based transferability attack, but in fact it is really directly based on those methods. Equation 1 is generalized Procrustes analysis: a shared latent Z, per-model linear maps W_k, a near-orthogonality penalty. That lineage running through CCA variants, CKA, and model stitching. The only thing separating it from vanilla RSA is λ_p·L_pair

  3. Most importantly, UniCal undeperforms PolyJailbreak consistently.

[1] Images are Achilles' Heel of Alignment: Exploiting Visual Vulnerabilities for Jailbreaking Multimodal Large Language Models

[2] FORCE: Transferable Visual Jailbreaking Attacks via Feature Over-Reliance CorrEction

Questions for authors

please refer to weaknesses

4 comments
AuthorsOpenReview ↗

Response to Reviewer R4kh

We sincerely thank the reviewer for recognizing the practical significance of one-time target adaptation and the technical novelty of linking behavioral representations to black-box calibration. We appreciate the suggestions on related work, methodological positioning, and performance evaluation.

1. HADES and FORCE.

We thank the reviewer for pointing out HADES and FORCE, both of which have been added to Section 2.1. HADES investigates image-mediated vulnerabilities in multimodal safety alignment, while FORCE improves the transferability of visual attacks by mitigating reliance on model-specific features. Both contribute important perspectives on transferable attack construction.

UniCal examines a complementary question: can a multimodal carrier prepared using open surrogates be calibrated once with discrete target feedback and then reused unchanged across unseen requests? This setting combines target-specific adaptation with fixed-carrier reuse, rather than focusing solely on zero-feedback transfer. Tables 1–2 report baseline effectiveness under the methods' stated access conditions, while Figure 4 separately evaluates query efficiency against label-only Random Search under a matched feedback interface. We have clarified these evaluation scopes without treating results from different protocols as directly interchangeable.

2. Representation alignment and technical contribution.

We appreciate the observation that Equation 1 is related to classical alignment methods, including Procrustes-style mappings. In UniCal, the alignment objective supplies behavior-aligned coordinates for generator conditioning and subsequent black-box adaptation.

Specifically, behavior-paired anchors shape a shared refusal–compliance coordinate space across heterogeneous surrogates. A coordinate-conditioned generator learns to realize coordinate shifts through request-independent image–text carriers. For a new target, the generator is frozen, and adaptation searches only sparse offsets using three-way decision labels, without target representations or generator retraining. The technical contribution is this connection from cross-model behavioral alignment to a realizable, reusable calibration interface.

The existing results support its practical role. Shared coordinates achieve 0.859 held-out behavioral AUC versus 0.737 for PCA (Figure 5), and 81.7% strict success versus 77.4% for PCA and 74.9% for a single-direction control within the calibrated pipeline (Table 3). Removing the coordinate-realization loss lowers strict success to 75.9% (Table 4). We have made the relationship to prior alignment methods and the scope of our contribution clearer in Section 2.2.

3. Comparison with PolyJailbreak.

We appreciate the reviewer highlighting PolyJailbreak as a strong baseline. Both methods exploit reuse, but differ in what is reused and how target feedback is incorporated. PolyJailbreak uses reusable Atomic Strategy Primitives in target-adaptive attack construction. UniCal instead fixes the same image–text carrier after one calibration using three-way decision labels; subsequent requests require no carrier update or request-specific optimization.

Table 2 reports the higher closed-target strict success of PolyJailbreak (93.2% versus 81.7% for UniCal). These figures describe effectiveness under different native-access protocols, rather than a matched comparison of target-query costs. UniCal additionally studies a constrained setting in which one-time label-only calibration improves strict success from 76.6% to 81.7% (Table 3), and its 192-query calibration cost breaks even after eight reused requests relative to label-only Random Search (Figure 4). Thus, native-access attack effectiveness and matched-feedback amortized efficiency provide complementary views of performance.

We thank the reviewer again for the constructive assessment. We hope the clarified technical contribution, related-work positioning, and evaluation protocols will be helpful in reconsidering the overall assessment.

AuthorsOpenReview ↗

Our Deepest Appreciation to Reviewer R4kh for the Careful Review and Invaluable Constructive Feedback

Dear Reviewer R4kh,

Thank you again for your valuable comments and suggestions. We would sincerely appreciate it if you could take another look at our rebuttal and revised manuscript and let us know whether our clarifications have addressed your concerns. We would be very happy to discuss any remaining questions or suggestions.

We truly appreciate your time and thoughtful consideration. Wishing you continued success in your research, and hoping you have a wonderful life filled with happiness and fulfillment!

Best regards,

The Authors

I thank the authors for their comprehensive response, most of my concerns have been well-addressed, I will adjust my evaluation accordingly.

Dear Reviewer R4kh,

Thank you very much for your positive follow-up and for taking the time to carefully revisit our rebuttal and revised manuscript!

We are truly grateful that our clarifications have addressed most of your concerns. We sincerely appreciate your openness, thoughtful feedback, and willingness to reconsider our work. Your constructive suggestions have been invaluable in helping us improve the clarity and technical positioning of our paper.

Thank you once again for your time and thoughtful consideration. We sincerely hope you have a wonderful life filled with happiness and fulfillment!

Warm regards,

The Authors

Reviewer_H3vdOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces UniCal (Universal Calibration), a black-box multimodal jailbreak attack framework, to achieve a "calibrate once, reuse multiple times" attack paradigm. By mapping the refusal and compliance behaviors of heterogeneous VLMs into a shared, low-dimensional coordinate space, the method generates a transferable, request-independent image-text attack carrier, attached to the original request. This method addresses the trade-off between target adaptability and query efficiency inherent in existing attack approaches.

Critical strengths
  1. Meaningful problem formulation. The paper investigates an interesting setting, balanceing target-specific adaptation and query efficiency.

  2. Reasonable framework design. This method integrates cross-model refusal/compliance representation learning, coordinate-conditioned carrier generation, and low-dimensional black-box calibration into a unified approach.

Critical weaknesses
  1. Unclear presentation and organization.

In the method sections, the dependencies among shared-coordinate learning, carrier generation, and one-time calibration are not well-structured in the illustratrion. Despite the relatively complex representation learning and generator training procedures, the actual target-side optimization is essentially a low-dimensional candidate search. Figure 2 is too dense with many icons and small texts, while the caption is too simple, making it very difficult to understand the overall pipeline.

Section 4.1 introduces representation-level and attack-level leave-one-model-out (LOMO) protocols immediately after describing the shared coordinate space. However, these are evaluation protocols rather than essential steps of the attack pipeline. Representation-level LOMO allows access to unlabeled hidden states of the held-out model for registration, but attack-level LOMO excludes access to the target model's internal representations entirely. Introducing both protocols within the methodology obscures the distinction between offline representation validation and actual black-box target adaptation.

In the experiment sections, Table 1 Open-model strict success contains two different settings, first five are attack-level LOMO, allowing one label-only calibration, while the latter two are zero-feedback transfer.

  1. Insufficient justification for the refusal/compliance coordinates. Section 4.1 learns a shared refusal/compliance subspace, while Section 4.2 trains a generator to induce the desired hidden-state displacement. However, predicting refusal does not necessarily imply the ability to control refusal, and reducing refusal does not necessarily lead to successful jailbreaks. In particular, the paper groups both l=1 (partial or non-substantive compliance) and l=2 (substantive policy-violating compliance) into the same compliance category when constructing behavior anchors. It is therefore unclear whether the learned coordinates genuinely capture the behavior relevant to the final attack objective. And whether the observed improvements primarily arise from the effectiveness of the universal carrier generator itself, or from the learned cross-model refusal coordinates that are intended to enable efficient one-time calibration is not fully justified.
Questions for authors

Q1. Since both partial compliance and substantive policy-violating compliance are grouped into the same compliance category, how do the authors justify that the learned coordinates capture behavior relevant to successful jailbreaks rather than merely predicting or reducing refusals?

Q2. Table 3 compares shared coordinates against PCA, single-direction, and random alternatives. And Table 10 in the supplementary provides more results on representation level LOMO. However, the relationship between refusal/compliance prediction and strict jailbreak success remains insufficiently justified.

Q3. It is also unclear whether the observed improvements primarily stem from the learned coordinates or the effectiveness of the carrier generator itself. Could the authors provide more controlled comparisons to disentangle the contributions of coordinate learning and carrier generation?

Q4. Could the authors additionally report one-time calibration results on Molmo and Phi-Vision to isolate the benefit of target-side adaptation on completely unseen architectures?

4 comments
AuthorsOpenReview ↗

Response to Reviewer H3vd

We thank the reviewer for recognizing the value of the calibrate-once, reuse-many setting and UniCal's integrated design. We appreciate the thoughtful questions about behavioral coordinates, carrier generation, and strict success, and address them using our method and existing evidence.

1. Presentation, pipeline, and LOMO protocols.

UniCal has four operations: (i) learning shared behavioral coordinates from open surrogates, (ii) training a coordinate-conditioned carrier generator, (iii) one-time target calibration via sparse coordinate selection, and (iv) reusing the fixed carrier on unseen requests.

Lightweight target-side search is enabled by the preceding surrogate-side learning. Shared coordinates define a behavior-informed search space, and the generator maps coordinates to multimodal carriers. This leaves only sparse coordinate selection at the target, without generator retraining.

The LOMO protocols have different roles. Representation-level LOMO uses unlabeled held-out representations to assess cross-model coordinate validity. Attack-level LOMO excludes the target from surrogate-side construction and uses only three-way decision labels during calibration; representation-level registration is not used. We clarified this distinction in Section 4.1 and the Figure 2 caption. Table 1 also distinguishes five attack-level LOMO targets from two external zero-feedback targets.

2. Q1–Q2: Behavioral coordinates and strict success.

Binary refusal/compliance anchors capture shared behavioral structure, not the final success criterion. Target calibration retains the distinction between partial and substantive compliance through ψ(0) = 0, ψ(1) = 0.5, and ψ(2) = 1; strict success requires substantive policy-violating compliance (ℓ = 2). The coordinates thus parameterize carrier generation and calibration, while success is evaluated under the stricter criterion.

Figure 5 reports held-out refusal–compliance AUC of 0.859 for shared coordinates versus 0.737 for PCA. Crucially, Table 3 compares them in the calibrated attack pipeline: shared coordinates achieve 81.7% strict success, versus 77.4% for PCA, 74.9% for a single direction, and 65.8% for random coordinates. The advantage is thus observed in downstream strict success as well as behavioral prediction.

3. Q3: Coordinate learning versus carrier generation.

The two components serve complementary purposes: coordinate learning establishes a cross-model behavioral parameterization; the generator realizes its image–text carriers. Table 3 examines representation choice within the calibrated pipeline. Table 4 evaluates generator objectives: removing coordinate realization reduces strict success from 81.7% to 75.9%, while removing behavioral loss reduces it to 70.8%. Together, these ablations support the contributions of both the learned coordinates and their generator-based realization.

4. Q4: Completely unseen architectures.

Molmo and Phi-Vision were reserved for zero-feedback transfer: one surrogate-prepared carrier was fixed before either target was accessed. Their strict success rates of 80.1% and 77.9% demonstrate cross-architecture transfer without target calibration.

One-time adaptation is evaluated separately by attack-level LOMO and Table 3's calibration ablation, where strict success increases from 76.6% to 81.7%. A paired calibration test on Molmo/Phi-Vision would quantify an additional architecture-specific benefit, distinct from zero-feedback transfer. Our current results establish both one-time adaptation on held-out targets and zero-feedback transfer to external architectures; Section 5.2 and Table 1 now distinguish these protocols.

We thank the reviewer for helping us clarify the method and its supporting evidence.

Dear Reviewer H3vd,

Thank you again for your thoughtful feedback and positive response! We sincerely appreciate your time, openness, and willingness to revisit our work. We are very glad that our clarifications have addressed your concerns and are truly grateful for your constructive engagement throughout the review process. Wishing you continued success in your research, happiness in your personal life, and all the very best!

Warm regards,

The Authors

Thank you for your detailed response and clarifications. The distinction between the LOMO protocols, and the ablation evidence have addressed my main concerns.

Dear Reviewer H3vd,

We are delighted to hear that our clarifications regarding the LOMO protocols and ablation evidence have resolved your main concerns.

Thank you once again for your thoughtful engagement and encouraging follow-up. We truly appreciate it! We sincerely hope you have a wonderful life filled with happiness and fulfillment!

Warm regards,

The Authors

Comments on the paper

3 comments
AuthorsOpenReview ↗

Overall Response and Summary Part 1

Dear Area Chair, Senior Area Chair, and Reviewers,

We sincerely thank all reviewers for their careful assessments, constructive suggestions, and recognition of the practical significance of our work. We have carefully considered the feedback, revised the manuscript, and provided detailed responses to clarify our methodological contributions, experimental evidence, and evaluation protocols.

Main contribution and limitations addressed. UniCal addresses a practical limitation of existing multimodal jailbreak methods: the difficulty of combining target-specific adaptation with cross-request reuse under restricted black-box feedback. Zero-feedback transfer attacks cannot adapt their fixed carriers to newly encountered targets, while request-dependent black-box attacks may require repeated optimization for new requests. UniCal bridges this gap through shared behavioral coordinates, a request-independent carrier generator, and one-time decision-label calibration, enabling subsequent reuse without request-specific optimization.

Evidence supporting the shared behavioral coordinates. We clarified the relationship between representation-level prediction and downstream attack effectiveness. The learned coordinates capture shared refusal–compliance behavior, while strict success requires substantive policy-violating compliance. Our shared coordinates achieve a held-out refusal–compliance AUC of 0.859, compared with 0.737 for PCA (Figure 5). More importantly, they achieve 81.7% strict success, compared with 77.4% for PCA, 74.9% for a single-direction control, and 65.8% for random coordinates (Table 3). Removing the coordinate-realization loss reduces strict success to 75.9%, while removing the behavioral loss reduces it to 70.8% (Table 4). These results support the complementary roles of behavioral alignment and carrier generation.

One-time calibration, transferability, and query efficiency. One-time calibration improves strict success from 76.6% to 81.7% compared with the uncalibrated carrier (Table 3). The 192-query calibration cost reaches its break-even point after eight reused requests relative to label-only Random Search under the matched feedback interface (Figure 4). Separately, zero-feedback transfer achieves 80.1% on Molmo and 77.9% on Phi-Vision (Table 1). We clarified the distinction between representation-level LOMO, attack-level LOMO, and zero-feedback external transfer in Sections 4.1 and 5.1 and Table 1.

Effectiveness, utility, and comparison with prior methods. PolyJailbreak achieves higher raw closed-target strict success (93.2% vs. 81.7%), while UniCal achieves higher benign utility (93.5% vs. 88.7%) and lower detection AUC (0.60 vs. 0.70) under the reported evaluation settings (Table 2 and Appendix Table 5). We clarified these trade-offs and the differences in feedback access and adaptation protocols, without treating them as directly matched comparisons.

Additional manuscript revisions.

  • Related Work (Sections 2.1–2.2): Added HADES and FORCE and clarified the connection to Procrustes-style representation alignment.
  • Method (Section 4.1 and Figure 2): Clarified the four-stage pipeline and distinguished the two LOMO protocols.
  • Experiments (Section 5, Tables 1–4): Expanded the interpretation of transfer, calibration, coordinate learning, and generator ablations, and improved table presentation.
  • Reproducibility (Appendix A.2): Reported the generator coefficients λb=1.0\lambda_b=1.0, λv=0.5\lambda_v=0.5, λt=0.05\lambda_t=0.05, and λu=1.0\lambda_u=1.0.
  • Ethical Considerations (Section 7): Added a discussion of potential dual-use risks and responsible reporting.
AuthorsOpenReview ↗

Overall Response and Summary Part 2

Progress during the author–reviewer discussion. We are particularly grateful for the constructive and encouraging engagement from the reviewers and the Area Chair.

Reviewer H3vd has kindly revisited their assessment and explicitly confirmed that our clarification of the LOMO protocols and the ablation evidence addressed their main concerns.

Reviewer R4kh has also reconsidered their assessment and acknowledged that most of their concerns have been well addressed by our comprehensive response.

Reviewer RjGE has acknowledged that our rebuttal addressed most of their questions and concerns. In response to their remaining question, we have provided an additional clarification explaining how UniCal addresses the limitations of zero-feedback transfer and repeated request-specific optimization.

We sincerely appreciate the reviewers' willingness to revisit our work and the Area Chair's efforts in facilitating constructive discussion. Their feedback has been invaluable in improving the clarity, technical positioning, and presentation of UniCal.

We hope that the revised manuscript, supporting experimental evidence, and subsequent clarifications will be carefully considered during the final assessment. We remain happy to address any further questions.

Thank you again for your valuable time and efforts.

Best regards,

The Authors

Dear Area Chair, Senior Area Chair, and Reviewers,

As the author–reviewer discussion comes to a close, we would like to express our heartfelt gratitude to all of you for your time, thoughtful evaluations, and constructive engagement throughout the review process.

We are especially grateful to our Area Chair for kindly facilitating the discussion, and to all three reviewers for their openness, careful reconsideration, and valuable feedback. Your suggestions have helped us significantly improve the clarity, technical positioning, and presentation of UniCal. We deeply appreciate the opportunity to engage in such a constructive scholarly discussion, and we are genuinely grateful for the time, care, and expertise you have generously shared with us.

We sincerely wish you all continued success in your research and a wonderful life filled with happiness, fulfillment, and many rewarding moments ahead!

With our warmest regards and sincere appreciation,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.