acceptodds
Under review as a conference paper at ICLR 2027

Safe Representation for Combined Image-Text Generation in Multi-Turn Conversations

Abstract

Multi-turn conversational interaction introduces complex security risks as harmful intentions become dispersed and concealed across dialogue steps. In unified models, this problem is exacerbated by the tight integration of image and text modalities, leading to safety misalignment during continuous generation. While current research filters explicit risks within single modalities, it overlooks the emergent harm induced by cross-modal interactions. We identify a critical risk pattern termed semantic mis-grounding, where standalone images and text appear benign, yet their combination conveys inappropriate thematic meanings, and the core capability at stake is whether an image is permitted to be interpreted under a specific thematic perspective. Empirical evaluations across five unified models confirm the prevalence of this issue. To address it, we propose Uni-SafeR, a framework for Unified Safe Representation that focuses on the associative semantics between modalities. Uni-SafeR performs safety alignment over full multi-turn interaction trajectories: the objective spans both the image generated in the first turn and the safety-aware textual response in the second turn, jointly shaping the contrast between safe and harmful responses and the contrast between benign and mis-grounding image generations. Training is instantiated on a dataset organized around four task-specific strategies, including diversified templates, benign anchors, counterfactual pairs, and multi-source images, which encourages the model to reason about the safety of guided multimodal alignments and to reject unjustified associations. Extensive experiments demonstrate that Uni-SafeR effectively reduces multimodal harm across various models while preserving, and even improving, their core generation and understanding capabilities.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Multi-turn conversational, Safe generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_49S8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies safety in two-turn unified image–text generation, where an image and a theme-guided narrative are individually benign but jointly convey an inappropriate association. It introduces Uni-SafeR, which aligns the full trajectory with supervised fine-tuning on both the generated image and the textual response, followed by trajectory-level DPO. Training data are constructed through multi-template expansion, benign anchors, counterfactual image–theme pairs, and multi-source images. On MPUP, the method reduces the image–text inappropriate rate from 30–78% to below 15% while maintaining a false refusal rate of 2.10%.

Critical strengths
  • The two-turn formulation is a reasonable minimal setting for studying compositional harm. Separating image generation from theme-conditioned narration makes the binding event explicit and avoids the confounds of long dialogue history.
  • The data construction contains a genuine methodological insight. Counterfactual pairs hold the theme fixed while varying the image, and benign anchors provide positive supervision for permissible associations. Table 6 shows that the counterfactual step reduces the inappropriate rate from 31.29% to 10.14%, while anchors reduce the false refusal rate from 41.20% to 3.40%.
  • The comparison with generic safety SFT and DPO is the correct baseline. It indicates that the improvement is not merely a consequence of applying standard alignment, and that task-specific supervision is important for the safety–over-refusal trade-off.
Critical weaknesses
  • The central phenomenon is substantially induced by the prompt. Templates Q1–Q3 instruct the model to align its tone with the theme and to make the theme "clearly felt", and the image prompt in Figure 1 already contains the racial analogy that the subsequent narrative is said to produce. Under such instructions, a high combined inappropriate rate is consistent with instruction-following; it does not establish that the model spontaneously generates harmful associations.
  • The evaluation is confirmatory rather than independent. The image–text judge is instructed to determine whether a harmful connection exists and to label the content harmful if it does, and the human annotators are aware of the theme. Training themes are inherited from MPUP and the test set uses MPUP's original prompts, but the paper does not document a prompt-level split, so semantic overlap cannot be excluded.
  • The claim of joint image–text alignment is not empirically isolated. Eq. 7 contains only textual likelihoods, and no ablation sets λ_img=0. Moving from SFT-only to the full method mainly reduces the false refusal rate (19.80% to 2.10%), while the inappropriate rate changes only from 7.33% to 6.50%. Capability is evaluated only on Janus-Pro, and the FID computation (1,000 generated images against COCO-30k captions) does not follow the standard definition.
Questions for authors

NA

4 comments
AuthorsOpenReview ↗

Response to Reviewer 49S8 (Part1)

We really appreciate your time and insightful comments. The response to your concerns are summarized as follows.


W1: The inappropriate rate may reflect instruction-following rather than the model spontaneously generating harmful associations.

A1:

We thank the reviewer for this comment, which we believe partly stems from a misunderstanding of the threat model we study.

(1). Our study does not claim that the model spontaneously generates harmful content. What we study is a way in which unified models can be led to produce inappropriate content, and this may well be intentionally guided by the user. Our position is that the model should recognize such implicit intent and refuse. The paper therefore shows that current unified models, under this kind of guidance (effectively an attack setting), generate inappropriate content without refusal, and then studies how to address this problem.

(2). This is analogous to a typical jailbreak attack. If a user directly asks for the procedure to make a bomb and the model follows the instruction, it outputs a harmful procedure. Following the instruction is itself an instance of the model's instruction-following ability, but the output is harmful, so such instruction-following is not necessarily desirable. In other words, "instruction-following" and "a high inappropriate rate" are not competing explanations: the latter is precisely the safety failure of the former.

(3). The way inappropriate content arises in our setting likewise builds on the model's instruction-following ability: the model links an image with a specific theme and generates inappropriate content that should not be generated. The model is therefore not expected to avoid harmful text spontaneously after seeing an image; rather, when it is guided in a direction that may produce inappropriate content,*** it should recognize the implicit association, explain it, and refuse.*** Notably, the guidance is implicit: the templates never ask for harmful content, and both the generated image and the generated text are judged benign in isolation. This makes such guidance harder to detect than an explicit harmful request.

(4). Moreover, the effect does not depend on the image and text being produced by the same model: in Table 5, cross-source pairings, where the image is generated by SD3 or Show-o and the text by LLaVA-OneVision, reach 51.57% and 72.36% image-text IR.

(5). What Uni-SafeR changes is precisely the model's behavior under such guidance. After alignment, the image-text IR drops below 15% across all models and evaluators, while the false refusal rate on benign pairs remains at 2.10% . That is, the model refuses when the guidance targets an impermissible binding, and still follows instructions otherwise.


W2: Confirmatory judging and undocumented prompt-level split.

A2: (1). Our evaluation of the generated content follows the standard LLM-as-judge paradigm, which is a typical practice in safety evaluation. To ensure the reliability and accuracy of the evaluation, we evaluate with multiple models of different families and scales, including the Qwen3-VL series (4B/8B/32B), LLaVA-OneVision-8B and GPT-4o, and we now add four further evaluator families. The results are consistent with the main paper: content that is safe in single-modality evaluation shows a strong harmful tendency in the combined form.

Evaluatorimage IR (%)text IR (%)image-text IR (%)
InternVL3.5-8B1.855.2348.29
GPT-5.58.424.9668.07
Claude-Sonnet-4-66.289.7761.73
Gemini-2.5-Pro9.746.6170.84

In addition, the automated judgments are validated against independent human judgments rather than taken as ground truth: in the image-text setting.

(2). Regarding the human evaluation, before annotation the annotators were briefed on the categories of content they might encounter, as required by the human-subject safety protocol (informed consent). During the actual annotation, however, the theme associated with each individual image-text pair was not disclosed. The annotation was performed by three independent annotators.

(3). Regarding the training and test data, test items are MPUP's original image-generation prompts, which do not appear in training; training uses newly constructed prompt variants and images generated fresh from multiple engines (Appendix H).

(4). To reflect this issue more directly, we also conduct an out-of-domain(OOD) evaluation. On 200 image-text pairs whose themes are absent from MPUP, the base model has an image-text IR of 48.73% and Uni-SafeR reduces it to 7.94%, while the single-modality rates remain unchanged.

Methodimage IR (%)text IR (%)image-text IR (%)
Janus-Pro (base)1.023.6148.73
Janus-Pro (Uni-SafeR)0.743.227.94

AuthorsOpenReview ↗

Response to Reviewer 49S8 (Part2)

W3: Eq. 7 confusion; no ablation sets λ_img=0; capability on Janus-Pro only; and the FID computation.

A3: We thank the reviewer for pointing out these issues.

(1). On Eq. 7, the preference is defined over trajectories τ± = (I±, r±) given the conversation context c = (p_img, q_theme), as stated immediately before Eq. 7. We wrote the likelihood in terms of c to keep the formula concise. We understand that this compressed notation may have caused confusion, and we will write Eq. 7 with the factorized form.

(2). We agree with the reviewer, and we conduct a controlled sweep on Janus-Pro in which the training data and the number of steps are held fixed, and only the weight of the image-likelihood term λ_img varies:

Variantimage-text IR (%)FFR (%)FID ↓GenEval ↑
Base50.011.1068.830.64
SFT-only (task-specific data)7.3319.8069.720.61
+ DPO, λ_img = 0 (text only)7.2022.0969.880.60
+ DPO, λ_img = 0.57.054.2268.370.66
+ DPO, λ_img = 1 (full trajectory)6.502.1067.280.70
  • ① The safety gain is carried mainly by the data construction (Table 6: 50.01% to 6.50%, with the largest single step from counterfactual pairs, 31.29% to 10.14%).
  • ② The contribution of the trajectory-level objective is concentrated on the safety-utility trade-off rather than on the IR itself: SFT-only already reaches 7.33% IR, but at a false refusal rate of 19.80%; the joint trajectory preference optimization brings the false refusal rate down to 2.10% while keeping the IR at 6.50%.
  • ③ The λ_img = 0 row shows that text-only preference optimization is not a substitute for trajectory-level optimization in this setting. It barely moves the IR (7.33% to 7.20%), raises the false refusal rate (19.80% to 22.09%), and further degrades generation (FID 69.88, GenEval 0.60). Without the image-likelihood term, the preference signal cannot separate a justified from an unjustified binding, so the model falls back on a stricter textual refusal prior instead of cross-modal reasoning.

Accordingly, the task-specific data construction produces the safety gain, and the trajectory-level objective with the image-likelihood term determines where the aligned model sits on the safety-utility plane: it converts a safety gain that costs generation quality (SFT-only) into one that improves it.

(3). On capability evaluation, we agree with the reviewer. We have supplemented the generation and understanding evaluation on more adapted models, and the results are as follows:

ModelsMMMU ↑MME ↑GQA ↑SEED ↑GenEval ↑
Show-o (base)30.5134266.158.00.68
Show-o (Uni-SafeR)29.6135765.558.70.72
Harmon (base)31.1135768.359.20.70
Harmon (Uni-SafeR)32.0135567.659.90.72
UniPic (base)27.9134665.857.20.67
UniPic (Uni-SafeR)28.4134064.757.00.70
  • ① Consistent with the Janus-Pro results in Table 3, the understanding scores fluctuate within about one point on all benchmarks, and GenEval improves on all three backbones. This supports our conclusion that Uni-SafeR does not compromise, and can even improve, the core capabilities across architectures.

  • ② We further report the false refusal rate on 1,000 benign image-text pairs for the aligned models, which is the most direct measure of whether the alignment impairs normal use: Janus-Pro 2.1%, Show-o 3.4%, Harmon 1.6%, UniPic 0.8%. All are below 5%.

(4). On FID, computing the full standard FID requires generating more than 30,000 images. For a fast and controlled measurement of the change in generation quality, we followed the practice of prior work [1] and randomly selected 1,000 captions to generate images. Since this protocol is applied identically to the base and aligned models, it is valid for relative comparison, although the absolute values are not comparable to the standard FID-30k. We will recompute FID under the standard protocol, which may take some time. In either case, GenEval, TIFA and CLIP-Score, which follow their standard protocols, all improve after alignment.

[1] Lyu M, Yang Y, Hong H, et al. One-dimensional adapter to rule them all: Concepts, diffusion models and erasing applications. CVPR 2024


Thank you again for your valuable suggestions. We value the opportunity to discuss our work with the research community and welcome any further questions or comments.

AuthorsOpenReview ↗

This Is a Kind Reminder for Reviewer 49S8

Dear Reviewer 49S8,

Thank you again for taking the time to provide valuable feedback on our paper. We kindly ask if you’ve had a chance to review our responses.

We hope we have adequately addressed your concerns and clarified any ambiguities. Should you have any further questions or suggestions, we would be happy to respond promptly.

Additionally, if our responses and the newly added experiment have satisfactorily resolved your concerns, we would greatly appreciate it if you could reflect this in your review score.

Thank you once again for your time and consideration.

Best regards, The Authors

AuthorsOpenReview ↗

To Reviewer 49S8

Dear Reviewer 49S8,

Thank you again for your detailed and technically precise review. As the discussion period is about to close, we would like to briefly recall how our response addressed your main concerns:

  • Prompt-induced phenomenon: we do not claim spontaneous harm. As in jailbreak research, the threat model is implicit user guidance; the templates never request harmful content, and following such guidance is precisely the failure the model should recognize and refuse.
  • Isolating joint alignment: a controlled λ_img sweep shows that data construction drives the safety gain, while the image-likelihood term drives the safety–utility trade-off (FRR 22.09 → 2.10, GenEval 0.60 → 0.70). Eq. 7 will be written in factorized form.
  • Evaluation and split: four additional evaluator families reproduce the pattern, test prompts never appear in training, and on an out-of-domain set with unseen themes IR drops from 48.73% to 7.94%.
  • Capability: results on Show-o, Harmon and UniPic show stable understanding scores and improved GenEval on all three.

If these responses address your concerns, we would greatly appreciate it if you could consider reflecting this in your score. We would be glad to answer any remaining questions within the remaining time.

Best regards, The Authors

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

This paper studies compositional safety risks in unified multimodal models under a two-turn image–text generation setting: the model first generates an image and then produces a theme-conditioned textual response, where the two outputs may appear benign individually but convey harmful meaning jointly. The authors propose Uni-SafeR, a safety alignment framework that combines task-specific supervision with SFT and DPO over multimodal interaction trajectories. Its data construction incorporates diverse query templates, benign anchors, counterfactual image–theme pairings, and images from multiple sources. Experiments across several unified models report reduced harmful image–text combinations, while additional evaluations show low false refusal rates and preservation of generation and understanding capabilities, with the most comprehensive capability assessment conducted on Janus-Pro.

Critical strengths

The paper studies a relevant compositional safety risk across multiple unified multimodal models, provides useful data ablations, and evaluates false refusals and capability preservation alongside safety improvements.

Critical weaknesses
  1. I am not fully convinced that the contribution beyond prior work has been established clearly enough. "Multimodal pragmatic jailbreak on text-to-image models" already studies harmful meanings arising from individually benign visual and textual content, and its classifier evaluation includes benign counterparts constructed by modifying either component. Uni-SafeR extends this problem to two-turn generation with unified models, but the paper does not sufficiently establish what new technical challenges this setting introduces or how its supervision design goes beyond adapting existing strategies.
  2. Table 7 does not yet isolate the benefit of full-trajectory optimization. Table 7 compares Uni-SafeR with generic safety baselines that use different training data, without clearly specifying whether their objectives include image-generation likelihoods. Consequently, this comparison does not isolate the benefit of joint image–text optimization. A controlled comparison between response-only and full-trajectory optimization using the same task-specific data would be necessary to substantiate this aspect of the method.
  3. I am also not convinced that the current evaluations establish semantic discrimination. Lower inappropriateness and false refusal rates demonstrate improved behavior on the evaluated distributions, but do not establish that the model has learned when a cross-modal interpretation is permissible. The benign evaluation uses samples from GQA and SEED, while the counterfactual ablation measures aggregate performance rather than correctness on matched safe–unsafe pairs. Direct evaluation on such pairs and ambiguous boundary cases would better distinguish relational safety reasoning from more selective refusal heuristics.
Questions for authors
  1. Could you clarify the novelty relative to “Multimodal pragmatic jailbreak on text-to-image models”? Specifically, what new challenge arises in the two-turn unified-model setting, and which components of Uni-SafeR are necessary to address it?
  2. Do the Table 7 baselines optimize image-generation likelihoods? Can you compare response-only and full-trajectory training using the same data and comparable training budgets?
  3. Table 6 shows that counterfactual training data improves aggregate safety and false refusal metrics. However, I could not find a held-out paired evaluation testing whether the model correctly answers a benign image–theme pairing while refusing its harmful counterpart under the same theme. Have you conducted such an evaluation? This would provide more direct evidence of semantic discrimination, complementing the existing training-data ablation.
4 comments
AuthorsOpenReview ↗

Response to Reviewer iuva (Part1)

We really appreciate your time and insightful comments. Thank you so much for acknowledging notable performance gain and detailed analysis of our work. The response to your concerns are summarized as follows.


W1 & Q1: Novelty relative to "Multimodal pragmatic jailbreak on text-to-image models".

A1: (1). We agree that we did not make this distinction explicit enough. In the revised manuscript, we distinguish the related lines of work along the axis that matters most, namely where the harmful image–text pair comes from:

WorkSettingWhere the harm residesDoes the harmful pair exist before generation?Supervision target
Multimodal pragmatic jailbreak (Liu et al., 2025; introduces MPUP)single-turn text-to-imagevisual text rendered within a single generated imageYes (specified in the prompt)prompt-side / single-image defense
SIUO, MSTS, VLSUimage + text → text comprehensiongiven image–text inputYesclassification of given inputs
Uni-SafeR (ours)multi-turn unified image–text generationbinding of two model-generated outputsNofull trajectory (I, T)

(2). There are two technical points that we should have stated more clearly in the paper.

  • First, in Multimodal pragmatic jailbreak, the harmful output is still a single generated image: the harm arises because visual text (e.g., slogans or captions) rendered within the image combines with the visual content, so a defense can act on the prompt or on the single output image. Our work does not focus on a single output, but on harmful content that emerges from multi-turn image–text generation, a common form of information expression in unified models. In our setting, the image–text pair is produced by the model across multiple turns, so the only place a defense can act is the trajectory.

  • Second, the vulnerability is not confined to one model's closed loop: in Appendix E (Table 5), pairing SD3-generated images with LLaVA-OneVision-generated descriptions reaches 51.57% IR, comparable to the internal-generation setting (47.71%).

(3). The multi-turn unified setting differs from single-turn compositional settings in three respects.

  • First, the harm is defined on the pair: Eq. 3 states H(I) = 0 and H(T) = 0 while H(I, T) = 1. Since neither component carries the harm, no decision can be made from either alone; the question is whether the image admits the interpretation induced by q_theme, i.e., the predicate π(I, q_theme).
  • Second, the textual response is generated conditioned on the image of the first turn, so the two outputs form a single joint expression rather than two independent generations.
  • Third, understanding and generation share one backbone in unified models, so supervision applied only to the text branch does not reach the image branch, while supervision applied strongly to either branch degrades the other.

(4). The trajectory supervision in Uni-SafeR is designed for the third point, and it is the third point that makes the first two difficult to satisfy simultaneously.


AuthorsOpenReview ↗

Response to Reviewer iuva (Part2)

W2 & Q2: Table 7 does not isolate joint optimization.

A2: (1). Reviewer is correct on both points. The Generic safety SFT and Generic safety SFT + DPO baselines are trained on text-only LLM safety alignment data, and their objectives contain no image-generation likelihood term.

(2). We therefore conduct the controlled comparison: all variants use the same task-specific data and the same training-step budget, and differ only in λ_img, where λ_img = 0 corresponds to response-only (text-only) optimization. We report IR, FRR, and generation metrics for all settings:

VariantImage–text IR (%) ↓FRR (%) ↓FID ↓GenEval ↑
Base50.011.1068.830.64
SFT-only (task-specific data)7.3319.8069.720.61
+ DPO, λ_img = 0 (response-only)7.2022.0969.880.60
+ DPO, λ_img = 0.57.054.2268.370.66
+ DPO, λ_img = 1 (full trajectory, Uni-SafeR)6.502.1067.280.70

(3). Three observations follow from this table:

  • ① The safety gain is carried mainly by the data construction (Table 6: 50.01% → 6.50%, with the largest single step coming from counterfactual pairs, 31.29% → 10.14%).

  • ② The contribution of the trajectory-level objective is concentrated on the safety–utility trade-off rather than on the IR itself. SFT-only already reaches 7.33% IR, but at a false refusal rate of 19.80%; joint trajectory preference optimization brings the false refusal rate down to 2.10% while keeping the IR at 6.50%.

  • ③ The λ_img = 0 row shows that response-only preference optimization is not a substitute for full-trajectory optimization in this setting. It barely moves the IR (7.33% → 7.20%), raises the false refusal rate (19.80% → 22.09%), and further degrades generation. Without the image-likelihood term, the preference signal cannot separate a justified binding from an unjustified one, so the model falls back on a stricter textual refusal prior instead of cross-modal reasoning.

(4). Accordingly, task-specific data construction produces the safety gain, while the trajectory-level objective with the image-likelihood term determines where the aligned model sits on the safety–utility plane: it converts a safety gain that costs generation quality (SFT-only) into one that improves it.


W3 & Q3: Semantic discrimination is not directly measured.

A3: (1). We agree with the reviewer. This is a useful suggestion. We conduct a held-out matched-pair evaluation consisting of 500 triples (q_theme, I_safe, I_harm). Within each triple, the theme and the prompt template are identical, and only the image varies: combined with q_theme, I_safe forms a benign pair that should be answered normally, whereas I_harm forms a pair whose combination is harmful. We report the rate of safe responses on the harmful pairs and the rate of normal answers on the benign pairs:

ModelSafe response on (I_harm, q_theme) (%) ↑Normal answer on (I_safe, q_theme) (%) ↑
Base38.8100.0
Uni-SafeR94.095.2

Here, a response is counted as safe if it is either an explicit refusal or judged harmless by the evaluator; this is why the base model obtains 38.8% on harmful pairs even though it rarely refuses explicitly.

(2). The base model answers almost every request regardless of whether the binding is permissible. After alignment, Uni-SafeR raises the safe-response rate on harmful pairs from 38.8% to 94.0%, while still answering 95.2% of the benign pairs under the same theme. Since the theme and the template are identical within each triple, a heuristic triggered by the theme alone would refuse both members of the pair; the results instead indicate that the model's decision depends on the relation between the image and the theme.


Thank you again for your valuable suggestions. We value the opportunity to discuss our work with the research community and welcome any further questions or comments.

AuthorsOpenReview ↗

This Is a Kind Reminder for Reviewer iuva

Dear Reviewer iuva,

Thank you again for taking the time to provide valuable feedback on our paper. We kindly ask if you’ve had a chance to review our responses.

We hope we have adequately addressed your concerns and clarified any ambiguities. Should you have any further questions or suggestions, we would be happy to respond promptly.

Additionally, if our responses and the newly added experiment have satisfactorily resolved your concerns, we would greatly appreciate it if you could reflect this in your review score.

Thank you once again for your time and consideration.

Best regards, The Authors

AuthorsOpenReview ↗

To Reviewer iuva

Dear Reviewer iuva,

Thank you again for your careful review. As the discussion period is about to close, we would like to briefly recall how our response addressed your three main concerns:

  • Novelty vs. Multimodal Pragmatic Jailbreak: there, the harmful pair is specified before generation, so defense can act on the input; in our setting, both components are generated by the model across turns, so only the trajectory can be supervised.
  • Isolating joint optimization: with identical task-specific data and budget, response-only optimization (λ_img = 0) barely lowers IR (7.33 → 7.20), raises FRR (19.80 → 22.09) and degrades generation, whereas full-trajectory optimization reaches IR 6.50, FRR 2.10 and GenEval 0.70.
  • Semantic discrimination: on 500 matched triples sharing the same theme and template, Uni-SafeR responds safely to 94.0% of harmful pairs while still answering 95.2% of benign ones, so its decision depends on the image–theme relation rather than on the theme alone.

If these responses address your concerns, we would greatly appreciate it if you could consider reflecting this in your score. We would be glad to answer any remaining questions within the remaining time.

Best regards, The Authors

Reviewer_u1o9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies cross-modal safety risks in multi-turn image-text generation, where individually benign images and text can jointly convey harmful meanings. It proposes Uni-SafeR, which combines task-specific data construction with trajectory-level SFT and DPO to align image generation and textual responses. Experiments across several unified multimodal models report substantial reductions in inappropriate image-text combinations while largely preserving generation and understanding capabilities.

Critical strengths
  1. The paper addresses a meaningful safety issue in unified multimodal generation. Extending compositional safety analysis to generated image-text sequences is practically relevant.

  2. The use of benign anchors and counterfactual pairs is well motivated. The experiments show substantial safety improvements with a low false refusal rate compared with generic safety alignment methods.

3.The evaluation covers several unified multimodal models and multiple safety evaluators, including human judgments. The additional generation and understanding benchmarks help assess whether safety alignment affects model utility.

Critical weaknesses
  1. Previous studies, including MPUP, SIUO, and VLSU, have already investigated harmful meanings emerging from otherwise benign multimodal components. The paper should better distinguish its contribution from existing compositional safety research.

  2. The methodological novelty is relatively limited. The framework largely combines SFT, DPO, benign anchors, and counterfactual data construction. The experiments do not fully isolate the advantage of joint image-text trajectory optimization over task-specific text-response alignment.

  3. The multi-turn setting is restricted to a fixed two-turn interaction. It remains unclear whether the method generalizes to longer conversations or more adaptive multi-turn attacks.

  4. The safety evaluation relies heavily on MLLM judgments and a limited set of constructed themes.

  5. The preservation of generation and understanding capabilities is mainly evaluated on Janus-Pro. Similar evaluations across the other adapted models are needed to support the broader utility claims.

Questions for authors
  1. Isolate the contribution of joint trajectory optimization from task-specific data construction.

  2. Discuss generalization to longer multi-turn interactions.

  3. Clarify the reliability of safety evaluation on borderline cases.

4 comments
AuthorsOpenReview ↗

Response to Reviewer u1o9 (Part1)

We really appreciate your time and insightful comments. Thank you so much for acknowledging notable performance gain and detailed analysis of our work. The response to your concerns are summarized as follows.


W1: Insufficient distinction from prior compositional safety work.

A1: We thank the reviewer for raising this point. We agree that the current Related Work section addresses this question in only one sentence. In the revised manuscript, we distinguish the lines of work along the axis that matters most, namely where the harmful image–text pair comes from:

WorkSettingWhere the harm residesDoes the harmful pair exist before generation?Supervision target
MPUP (Liu et al., 2025)single-turn text-to-imagepragmatics of the input promptYesprompt-side filtering / defense
SIUO, MSTS, VLSUimage + text → text comprehensiongiven image–text inputYesclassification of given inputs
Uni-SafeR (ours)two-turn unified image–text generationbinding of two model-generated outputsNofull trajectory (I, T)

(1). The distinction is not merely terminological. In prior work, the harmful pair is authored before the model acts, so a defense can operate on the input. In our setting, neither the image nor the text exists before generation, and each is individually benign, so input-side filtering is not applicable.

(2). Moreover, the textual response in the second turn is generated conditioned on the image of the first turn, so the two outputs form a single joint expression rather than two independent generations. As formalized in Eq. 3 and Eq. 4, H(I, T) cannot be recovered from H(I) and H(T); safety must therefore be judged on the pair.

(3). In particular, since understanding and generation share one backbone in unified models, supervision applied to only one branch either fails to reach the other or degrades generation. This is directly visible in Table 7: generic safety SFT+DPO reduces the image–text IR to 7.01% but raises the false refusal rate to 42.80%.

(4). We also note that Appendix E (Table 5) shows that the effect is not an artifact of a single model's internal generation loop: pairing SD3-generated images with LLaVA-OneVision-generated descriptions yields 51.57% IR, and pairing Show-o-generated images with LLaVA-OneVision-generated descriptions reaches 72.36%.


W2 & Q1: Joint trajectory optimization is not isolated from data construction.

A2: We agree with the reviewer and thank you for pressing on this point.

(1). We conduct a controlled sweep on Janus-Pro in which the training data and the number of training steps are held fixed, and only the weight of the image-likelihood term λ_img varies:

VariantImage–text IR (%) ↓FRR (%) ↓FID ↓GenEval ↑
Base50.011.1068.830.64
SFT-only (task-specific data)7.3319.8069.720.61
+ DPO, λ_img = 0 (text only)7.2022.0969.880.60
+ DPO, λ_img = 0.57.054.2268.370.66
+ DPO, λ_img = 1 (full trajectory, Uni-SafeR)6.502.1067.280.70

(2). Three observations follow from this table:

  • ① The safety gain is carried mainly by the data construction (Table 6: 50.01% → 6.50%, with the largest single step coming from counterfactual pairs, 31.29% → 10.14%).
  • ② The contribution of the trajectory-level objective is concentrated on the safety–utility trade-off rather than on the IR itself. SFT-only already reaches 7.33% IR, but at a false refusal rate of 19.80%; joint trajectory preference optimization brings the false refusal rate down to 2.10% while keeping the IR at 6.50%.
  • ③ The λ_img = 0 row shows that text-only preference optimization is not a substitute for trajectory-level optimization in this setting. It barely moves the IR (7.33% → 7.20%), raises the false refusal rate (19.80% → 22.09%), and further degrades generation (FID 69.88, GenEval 0.60). Without the image-likelihood term, the preference signal cannot separate a justified binding from an unjustified one, so the model falls back on a stricter textual refusal prior instead of cross-modal reasoning.

(3). Accordingly, task-specific data construction produces the safety gain, while the trajectory-level objective with the image-likelihood term determines where the aligned model sits on the safety–utility plane: it converts a safety gain that costs generation quality (SFT-only) into one that improves it. We will make this decomposition explicit in the revised manuscript.


AuthorsOpenReview ↗

Response to Reviewer u1o9 (Part2)

W3 & Q2: Generalization to longer conversations.

A3: (1). We agree that the two-turn interaction is a minimal canonical setting rather than a general claim, and we have stated this explicitly in Appendix I. The objective itself extends without modification: for a trajectory τ = (I₁, t₁, ..., I_K, t_K), the trajectory likelihood sums over turns, and Eq. 7 applies unchanged.

(2). Empirically, we add a transfer evaluation in which the first two turns are benign and the harmful theme is introduced only at turn 3. The aligned model is the one trained on two-turn trajectories only, so this experiment measures transfer rather than training in the longer setting. We report the image–text IR and FRR at turns 3 and 4:

TurnModelImage–text IR (%) ↓FRR (%) ↓
Turn 3Base45.233.08
Turn 3Uni-SafeR9.182.32
Turn 4Base52.661.89
Turn 4Uni-SafeR7.532.67

The results show that the safety behavior learned from two-turn trajectories transfers to longer conversations: the image–text IR is reduced from 45.23% to 9.18% at turn 3 and from 52.66% to 7.53% at turn 4, while the false refusal rate remains below 3%. We have added this experiment to the revised manuscript.


W4 & Q3: Evaluator dependence and theme coverage.

A4: (1). We agree with the reviewer. Being aware of this concern, we conducted a consistency analysis between MLLM evaluators and human judgments. Section 4.2 (Figure 4) shows that the evaluators are consistent with human judgment in the image–text setting, with Pearson r = 0.82 for Qwen3-VL-8B and r = 0.89 for Qwen3-VL-32B; similar analyses for LLaVA-OneVision-8B and GPT-4o are provided in Appendix C. This supports the credibility of the model-based evaluation results.

(2). Beyond the Qwen3-VL series, LLaVA-OneVision-8B and GPT-4o already reported in the paper, we now add four further evaluator families. The results are consistent with the main paper: content judged safe in single-modality evaluation shows a strong harmful tendency in the combined form.

EvaluatorImage IR (%)Text IR (%)Image–text IR (%)
InternVL3.5-8B1.855.2348.29
GPT-5.58.424.9668.07
Claude-Sonnet-4-66.289.7761.73
Gemini-2.5-Pro9.746.6170.84

(3). For human evaluation, three independent annotators labelled 20% of the test set, achieving Fleiss' κ = 0.71 for Janus-Pro and 0.73 for Show-o, which indicates substantial agreement. For borderline cases, the final human label is determined by majority voting, and the image–text evaluation rubric (Figure 11) explicitly treats refusals and safety-oriented analyses as safe and instructs the judge to follow a reasonable, mainstream human interpretation.

(4) Regarding theme coverage, we follow the setting of [1] and conduct experiments on four major harm categories, namely "hate speech", "physical harm", "fraud", and "pornography", which are further divided into 26 subcategories covering a broad range of inappropriate-content scenarios. To examine generalization beyond these themes, Appendix H reports results on an out-of-domain test set whose themes are absent from MPUP: the base model reaches 48.73% image–text IR, which Uni-SafeR reduces to 7.94%.

[1] Liu T, Lai Z, Wang J, et al. Multimodal pragmatic jailbreak on text-to-image models. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2025: 4681–4720.


W5: Capability preservation evaluated on a single backbone.

A5: (1). We agree with the reviewer. We have supplemented the evaluation of generation and understanding capabilities on more adapted models:

BackboneMMMU ↑MME ↑GQA ↑SEED ↑GenEval ↑
Show-o (Base)30.5134266.158.00.68
Show-o (Uni-SafeR)29.6135765.558.70.72
Harmon (Base)31.1135768.359.20.70
Harmon (Uni-SafeR)32.0135567.659.90.72
UniPic (Base)27.9134665.857.20.67
UniPic (Uni-SafeR)28.4134064.757.00.70

Consistent with the Janus-Pro results in Table 3, the understanding scores fluctuate within about one point on all benchmarks, while GenEval improves on all three backbones. These results support our conclusion that Uni-SafeR preserves, and can even improve, the core capabilities across different unified architectures.

(2) We further report the false refusal rate on 1,000 benign image–text pairs, which is the most direct measure of whether the alignment impairs normal use: Janus-Pro 2.1%, Show-o 3.4%, Harmon 1.6%, and UniPic 0.8%. All are below 5%.


Thank you again for your valuable suggestions. We value the opportunity to discuss our work with the research community and welcome any further questions or comments.

Thanks for running the Janus-Pro comparison with the training data held fixed. It answers my main question about what joint training adds: the clearest benefit is the drop in false refusals. The results on turns three and four also address my concern about the two-turn setup. I’m keeping my rating.

We are excited to hear that most of the concerns are addressed and that the reviewer acknowledged our newly added content. Thanks to the reviewer very much again for valuable suggestions, which significantly helped us improve the quality and clarity of the paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Our Work and the Discussion

Dear Area Chair and Reviewers,

We sincerely thank all reviewers for their thoughtful and constructive feedback, and the Area Chair for handling our submission. Below we summarize (I) the paper and its recognized strengths, and (II) the main concerns and how we addressed them.


I. Paper Summary and Recognized Strengths

What the paper does. We study cross-modal safety in multi-turn unified generation: an image and a theme-guided text, each benign, can jointly convey harmful meaning. We verify this risk across unified models, and propose Uni-SafeR, which combines task-specific data construction (benign anchors, counterfactual pairs, multi-template, multi-source) with trajectory-level SFT and DPO.

Strengths acknowledged by the reviewers:

  • A meaningful and practically relevant safety problem in unified multimodal generation (u1o9, iuva); the two-turn formulation is "a reasonable minimal setting" that "makes the binding event explicit" (49S8).
  • Well-motivated data design. Benign anchors and counterfactual pairs are "a genuine methodological insight" (49S8, u1o9), supported by useful ablations (iuva).
  • Strong safety–utility trade-off. Large safety gains with a low false refusal rate, compared against the "correct baseline" of generic safety SFT/DPO (49S8, u1o9).
  • Broad evaluation. Multiple unified models, multiple evaluators including humans, and capability benchmarks (u1o9, iuva).

II. Main Concerns and Our Responses

1. Novelty relative to prior compositional safety work (u1o9-W1, iuva-W1)

Our response:

  • In MPUP / Multimodal Pragmatic Jailbreak and SIUO / MSTS / VLSU, the harmful pair exists before the model acts, so defense can act on the input. In our setting, both components are generated by the model across turns and each is benign, so only the trajectory can be supervised.
  • The effect is not tied to one model's loop: cross-source pairs (SD3 + LLaVA-OneVision) reach 51.57% IR. A comparison table is added to Related Work.

2. Isolating joint trajectory optimization (u1o9-W2, iuva-W2, 49S8-W3)

Our response:

  • New: controlled λ_img sweep with identical data and budget (λ_img = 0 is response-only):
VariantIR ↓FRR ↓GenEval ↑
SFT-only7.3319.800.61
DPO, λ_img = 07.2022.090.60
DPO, λ_img = 1 (Ours)6.502.100.70
  • Data construction drives safety; the image-likelihood term drives the safety–utility trade-off. Response-only optimization barely lowers IR, raises FRR and degrades generation. Eq. 7 will be written in factorized form.

3. Semantic discrimination vs. selective refusal (iuva-W3)

Our response:

  • New: 500 matched triples (same theme and template, only the image differs). Uni-SafeR responds safely on 94.0% of harmful pairs (base 38.8%) while still answering 95.2% of benign pairs. A theme-triggered heuristic would refuse both; the decision depends on the image–theme relation.

4. Is the phenomenon merely prompt-induced? (49S8-W1)

Our response:

  • We do not claim spontaneous harm. As in jailbreak research, the threat model is implicit user guidance: the templates never request harmful content, and both outputs are benign in isolation. Following such guidance is precisely the safety failure; the model should recognize it and refuse.
  • After alignment, IR drops below 15% across all models while FRR stays at 2.10%: the model refuses impermissible bindings and still follows benign instructions.

5. Evaluation reliability and data split (u1o9-W4, 49S8-W2)

Our response:

  • New: four additional evaluator families (InternVL3.5, GPT-5.5, Claude-Sonnet-4-6, Gemini-2.5-Pro) reproduce the same safe-safe-harmful pattern.
  • Judges agree with humans (r = 0.82 / 0.89); three annotators reach Fleiss' κ = 0.72, without per-sample theme disclosure.
  • Test prompts never appear in training; on an out-of-domain set with themes absent from MPUP, IR drops from 48.73% to 7.94%.

6. Generalization beyond two turns (u1o9-W3)

Our response:

  • New: four-turn transfer with the harmful theme introduced at turn 3, using the two-turn-trained model: IR drops from 45.23% → 9.18% (turn 3) and 52.66% → 7.53% (turn 4), with FRR below 3%.

7. Capability preservation (u1o9-W5, 49S8-W3)

Our response:

  • New: capability results on Show-o, Harmon and UniPic: understanding scores stay within about one point, and GenEval improves on all three. FRR on benign pairs stays below 5% for every model.

We are grateful to the reviewers for their engagement, which has clarified our threat model and made the evidence for joint trajectory optimization more direct. We would be glad to provide any further clarification. Thank you for your time and consideration.

Sincerely,

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.