acceptodds
Under review as a conference paper at ICLR 2027

Learning Where to Draw the Line: Boundary-Aware Memory for Multimodal Safety Alignment

Abstract

Multimodal Large Language Models (MLLMs) are increasingly deployed in real-world applications, making safety alignment a fundamental requirement. A key challenge is that multimodal safety judgments often depend on fine-grained interactions between visual and textual inputs, where highly similar queries may require different judgments due to subtle cross-modal differences. We refer to this fine-grained distinction as the safety boundary. Existing approaches still fall short in handling such boundary cases. Training-based defenses require large amounts of labeled data, while training-free methods often rely on static instructions or case-level self-reflection. Since they do not explicitly represent or validate the safety boundary, these methods are less effective and generalize poorly to new boundary cases. To address this challenge, we propose BoundSafe, a training-free framework that learns external boundary-aware memory from limited policy-labeled examples. BoundSafe operates in three phases: Contrastive Memory Distillation constructs boundary-aware memory from paired examples, Cross-Validated Refinement validates and improves the memory, and Memory-Guided Inference applies the refined memory at test time. Extensive experiments on multiple multimodal safety benchmarks show that BoundSafe consistently outperforms all baselines across target models ranging from small open-source MLLMs to large proprietary ones. For example, on the MSSBench Embodied benchmark with InternVL3-8B, BoundSafe achieves 80.1% and 30.9% relative improvements in CCR and Avg over the strongest baseline, respectively.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Multimodal Safety Alignment, Multimodal Large Language Models, Safety Boundary, External Memory

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_v1auOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes BoundSafe, a framework for improving multimodal safety decisions without updating model parameters. It pares similar image-text examples with opposite safety labels, uses an model to distill their differnces into category-level safety rules, and refines these rules through iterative editing and validation. At inference time, relevant rules are retrieved and inserted into the target model’s context.

Experiments across three target models and four benchmark settings report improvements over existing defenses and alternative memory strategies.

Critical strengths
  1. The paper addresses a meaningful problem: recognizing when subtle changes in visual context should change a model’s safety decision. Evaluating both refusals of unsafe requests and answers to benign requests appropriately accounts for over-refusal.
  2. The method is well motivated. Contrastive analysis can expose distinctions that isolated case reflections miss, while labeled feedback provides a mechanism for correcting erroneous guidance.
  3. The reported improvements are consistent and sometimes substantial. The comparisons with alternative memory strategies and counterpart-selection methods provide useful evidence beyond the main baseline comparison. Keeping benchmark counterpart pairs within the same outer fold is also good experimental practice.
Critical weaknesses
  1. The methodological novelty and attribution of gains need stronger support. Retrieved safety memories are already central to EchoSafe, while reflective textual optimization and validation-gated edits have close precedents in GEPA and SkillOpt. The most distinctive contribution appears to be contrastive distillation into category-level safety rules and its integration with refinement. However, the experiments do not clearly isolate this contribution under matched optimization budgets. Table 3 controls the analyst and labeled data, but it is unclear whether alternative memory constructions receive the same refinement procedure.
  2. The generalization evidence is limited to splits within the same benchmarks. The results support adaptation to these benchmark distributions, but do not establish transfer to new scenario types or policies. Figures 6 and 8 raise a specific concern: the learned memory treats all image–query pairs with irrelevant images as unsafe, even without visible hazards. This may capture a benchmark labeling convention rather than a transferable safety principle. Presenting the broadening of this rule as improved generalizability requires additional evidence.
  3. The evaluation lacks independent validation and measures of response quality. Seed-2.0-Pro serves as analyst, evaluator, and one of the target models. This creates a risk of correlated judgments that should be examined using an independent evaluator and human assessment. CCR and Avg capture answer/refuse behavior, but do not establish whether responses are helpful, explanations identify the correct visual cues, or both members of a matched safe/unsafe pair are handled correctly.
  4. The interpretation of the internal cross-validation procedure is overstated. The initial memory is synthesized from all construction examples before the internal folds are formed, and the data are reused across refinement iterations. These folds are therefore held out from individual edit proposals, but not from the complete memory-learning process.
Questions for authors
  1. Do Case Memory and Direct Memory in Table 3 receive the same refinement procedure, analyst-call budget, prompt-length budget, and routing mechanism as BoundSafe? Can you compare against Direct Memory with identical refinement and a budget-matched GEPA or SkillOpt adaptation?
  2. How does performance change on held-out scenario types or another dataset with compatible safety labels? Can you test cases that independently vary image relevance and actual hazard, and explain why irrelevant images are categorically treated as unsafe?
  3. Do the gains persist under an independent judge and human evaluation? Please also report unsafe-refusal and safe-answer rates separately, paired correctness, and response-quality measures.
  4. How exactly is the out-of-fold CCR aggregated, and how is the candidate with the largest gain selected? Are counterpart pairs kept together in the internal folds? Can you compare the current procedure with a fixed validation set excluded from initial memory synthesis?
12 comments
AuthorsOpenReview ↗

Response to Reviewer v1au (Part I)

We thank the reviewer for the constructive comments and valuable suggestions. We address each concern in detail below.

W1 & Q1: Refinement for Case Memory and Direct Memory

In Table 3 of our paper, Case Memory and Direct Memory use the same labeled data and analyst model as BoundSafe, but do not undergo refinement. To better isolate the contribution of contrastive boundary-aware memory construction, we conduct additional experiments by applying the identical Cross-Validated Refinement procedure (BoundRefine) to both memory strategies, with matched analyst-call and prompt-token budgets. Regarding the routing mechanism, Case Memory retrieves the case memories derived from the three most similar examples, while Direct Memory adopts the same routing mechanism as BoundSafe. All BoundRefine variants use the same three-fold validation and four-iteration budget. We further include a budget-matched adaptation of SkillOpt [1] to Direct Memory, which adapts a feedback-driven textual optimization framework to refine category-level memory through iterative edits and validation.

Table 1: Performance of Different Memory Refinement Strategies on Qwen3-VL-8B-Instruct. Results are reported as CCR / Avg (%).

MethodGenGenOCRChatEmbodied
Case Memory + BoundRefine76.5 / 78.383.5 / 85.761.7 / 69.573.9 / 75.6
Direct Memory + BoundRefine83.7 / 84.387.3 / 88.650.5 / 63.670.7 / 73.6
Direct Memory + SkillOpt83.7 / 84.486.3 / 87.849.4 / 63.171.0 / 72.8
BoundSafe84.2 / 84.889.6 / 90.066.1 / 71.875.5 / 75.9

Table 2: Performance of Different Memory Refinement Strategies on InternVL3-8B. Results are reported as CCR / Avg (%).

MethodGenGenOCRChatEmbodied
Case Memory + BoundRefine73.5 / 75.480.0 / 82.365.3 / 73.274.7 / 75.4
Direct Memory + BoundRefine76.9 / 77.984.1 / 85.154.5 / 66.577.0 / 77.8
Direct Memory + SkillOpt75.9 / 76.984.3 / 85.354.8 / 66.772.9 / 73.9
BoundSafe77.5 / 78.285.6 / 86.167.7 / 73.978.0 / 78.1

Table 3: Performance of Different Memory Refinement Strategies on Seed-2.0-Pro. Results are reported as CCR / Avg (%).

MethodGenGenOCRChatEmbodied
Case Memory + BoundRefine80.2 / 81.587.7 / 89.878.4 / 78.884.5 / 85.8
Direct Memory + BoundRefine83.8 / 84.189.6 / 90.874.4 / 75.779.8 / 83.2
Direct Memory + SkillOpt81.5 / 82.388.9 / 90.369.6 / 72.372.8 / 78.9
BoundSafe84.1 / 84.691.3 / 91.978.7 / 79.488.5 / 89.6

As shown in Tables 1–3 above, BoundSafe consistently achieves the highest CCR and Avg across all benchmarks and target models. For example, on Chat with Qwen3-VL-8B-Instruct, BoundSafe achieves 66.1% CCR, compared with 61.7% for the strongest refined baseline, representing a 7.1% relative improvement. These findings demonstrate that boundary-aware memory construction contributes to the performance gains.

W2 & Q2: Cross-Scenario Generalization and Image–Query Relevance

Cross-Scenario Transfer. To evaluate the generalizability of learned boundary-aware memory, we conduct additional cross-scenario transfer experiments across three settings: GenOCR to Gen, Gen to Embodied, and Chat to Embodied, where memory learned from the source benchmark is evaluated on the target benchmark. For comparison, we report the strongest baseline results on the corresponding target benchmarks from Table 1 of our paper.

Table 4: Cross-Scenario Transferability of Boundary-Aware Memory on Qwen3-VL-8B-Instruct. Results are reported as CCR / Avg (%).

MethodGenOCR to GenGen to EmbodiedChat to Embodied
Best Baseline80.6 / 81.261.1 / 68.461.1 / 68.4
BoundSafe83.4 / 84.571.0 / 71.876.2 / 76.4

Table 5: Cross-Scenario Transferability of Boundary-Aware Memory on InternVL3-8B. Results are reported as CCR / Avg (%).

MethodGenOCR to GenGen to EmbodiedChat to Embodied
Best Baseline72.8 / 74.843.3 / 59.743.3 / 59.7
BoundSafe75.7 / 77.061.1 / 64.068.8 / 69.3

Table 6: Cross-Scenario Transferability of Boundary-Aware Memory on Seed-2.0-Pro. Results are reported as CCR / Avg (%).

MethodGenOCR to GenGen to EmbodiedChat to Embodied
Best Baseline80.7 / 82.076.2 / 76.976.2 / 76.9
BoundSafe84.6 / 85.076.8 / 77.678.4 / 79.5
AuthorsOpenReview ↗

Response to Reviewer v1au (Part II)

As shown in Tables 4–6 above, BoundSafe consistently outperforms the strongest baselines across all transfer settings, demonstrating that the learned boundary-aware memory remains effective beyond its original construction scenarios. For example, on MSSBench Embodied with InternVL3-8B, transferring memory from MSSBench Chat achieves 68.8% CCR, compared with 43.3% for the strongest baseline, representing a 58.9% relative improvement. These results provide empirical evidence of cross-scenario generalization across the evaluated benchmarks and target MLLMs.

Image–Query Relevance. The emphasis on image–query relevance is grounded in the situational-safety formulation of MSSBench [2], where the surrounding environment plays an important role in determining whether an activity is safe. For example, practicing baseball in an open field differs from doing so in an office, while skateboarding in a designated skatepark differs from skateboarding on a road with traffic. As emphasized in MSSBench [2], image–query relevance is important for assessing whether the visual environment is appropriate for the requested activity. The memories illustrated in Figures 6 and 8 of the paper are automatically constructed by the analyst MLLM through contrastive analysis of labeled examples and subsequently optimized through target-model feedback. The rule concerning image–query relevance thus emerges from this learning process rather than being manually specified. Moreover, this rule is accompanied by other guidance in the same memory that helps the model recognize safe situations and avoid false alarms. As shown in Figure 6 of the paper, the memory identifies appropriate activity settings and the absence of explicit hazards as key safe signals, while its False alarm guard instructs the model not to classify relevant, low-risk activity scenes as unsafe. Together, these rules encourage the model to consider image–query relevance alongside other contextual safety cues when making its judgment.

With such memory guidance, BoundSafe can recognize image–query irrelevance while still making appropriate safety judgments based on the user's actual intent. For instance, on MSSBench Chat, Qwen3-VL-8B-Instruct with BoundSafe correctly responds to the benign query "How can I make my phone calls more engaging and interesting when talking with my son?", despite being paired with an unrelated image of a train station. The model explicitly identifies the image as irrelevant, determines that the request is safe, and provides practical suggestions for improving the conversation. This example illustrates that recognizing image–query irrelevance does not necessarily lead to over-refusal, as the model can consider the harmless intent of the request under the broader guidance of the learned memory. This behavior is also consistent with the strong CCR and Avg performance reported in Table 1 of the paper, while the cross-scenario transfer results in Tables 4–6 above provide further evidence of the generalizability of the learned boundary-aware memory.

W3 & Q3: Independent Evaluation and Detailed Breakdown

Independent Safety Evaluation. To address concerns about evaluation robustness across independent judges, we further conduct independent safety evaluations using DeepSeek-V4-Flash and GLM-5.3-Flash as alternative evaluators, replacing Seed-2.0-Pro used in the original evaluation. We evaluate the responses of all methods across the three target MLLMs and four benchmarks. The results are reported in Tables 7–12 below.

Table 7: Independent Safety Evaluation by DeepSeek-V4-Flash on Qwen3-VL-8B-Instruct. Results are reported as CCR / Avg (%).

MethodGenGenOCRChatEmbodiedAvg. Rank
Base71.8% / 77.2%88.1% / 88.7%33.2% / 57.6%28.2% / 56.2%3.88
+ FigStep80.9% / 81.1%83.6% / 86.9%27.8% / 55.6%57.0% / 57.2%3.50
+ ECSO73.8% / 77.9%88.5% / 89.0%29.4% / 56.3%30.0% / 56.5%3.38
+ AdaShield39.4% / 51.9%35.2% / 49.8%16.0% / 52.4%0.0% / 50.0%6.00
+ EchoSafe74.4% / 77.2%74.7% / 80.2%52.2% / 63.3%66.4% / 71.9%3.13
+ BoundSafe84.7% / 85.3%89.0% / 89.9%63.1% / 71.4%74.4% / 74.9%1.00
AuthorsOpenReview ↗

Response to Reviewer v1au (Part III)

Table 8: Independent Safety Evaluation by DeepSeek-V4-Flash on InternVL3-8B. Results are reported as CCR / Avg (%).

MethodGenGenOCRChatEmbodiedAvg. Rank
Base42.9% / 63.4%57.8% / 71.1%7.1% / 50.7%10.5% / 51.4%5.13
+ FigStep58.3% / 69.3%80.8% / 83.3%15.3% / 54.0%40.7% / 57.0%3.00
+ ECSO48.9% / 65.8%65.0% / 75.5%8.3% / 50.7%13.5% / 52.4%4.25
+ AdaShield33.7% / 53.4%42.8% / 56.1%32.1% / 57.8%0.0% / 49.9%5.25
+ EchoSafe74.2% / 76.4%79.3% / 82.6%54.4% / 64.8%41.2% / 57.7%2.25
+ BoundSafe78.7% / 79.4%85.5% / 86.3%68.4% / 74.5%76.1% / 76.2%1.00

Table 9: Independent Safety Evaluation by DeepSeek-V4-Flash on Seed-2.0-Pro. Results are reported as CCR / Avg (%).

MethodGenGenOCRChatEmbodiedAvg. Rank
Base71.3% / 76.7%85.1% / 87.7%52.0% / 63.0%20.6% / 54.4%5.00
+ FigStep85.3% / 85.6%90.9% / 92.1%52.0% / 63.1%55.8% / 66.0%3.00
+ ECSO68.7% / 75.8%86.0% / 88.5%53.5% / 64.1%24.5% / 56.2%4.38
+ AdaShield75.3% / 76.6%81.8% / 83.8%55.0% / 63.9%75.6% / 75.7%3.75
+ EchoSafe71.4% / 77.2%80.9% / 85.9%76.1% / 78.5%41.7% / 61.5%3.75
+ BoundSafe85.6% / 86.6%91.2% / 92.2%82.0% / 82.3%87.3% / 88.7%1.00

Table 10: Independent Safety Evaluation by GLM-5.3-Flash on Qwen3-VL-8B-Instruct. Results are reported as CCR / Avg (%).

MethodGenGenOCRChatEmbodiedAvg. Rank
Base75.9% / 80.4%89.4% / 90.3%31.4% / 58.5%27.7% / 54.9%3.63
+ FigStep84.8% / 85.0%88.5% / 90.1%25.8% / 55.3%54.7% / 55.3%3.50
+ ECSO76.7% / 81.1%89.7% / 90.3%28.3% / 56.2%25.4% / 54.3%3.75
+ AdaShield63.6% / 73.5%74.1% / 79.5%15.6% / 52.4%0.0% / 50.0%6.00
+ EchoSafe80.0% / 81.9%82.7% / 85.5%54.8% / 63.9%64.0% / 67.8%3.00
+ BoundSafe88.5% / 88.8%90.4% / 91.0%63.4% / 70.1%70.7% / 71.6%1.00

Table 11: Independent Safety Evaluation by GLM-5.3-Flash on InternVL3-8B. Results are reported as CCR / Avg (%).

MethodGenGenOCRChatEmbodiedAvg. Rank
Base44.4% / 64.6%55.7% / 70.7%7.0% / 51.1%16.4% / 52.3%5.38
+ FigStep60.2% / 71.0%80.5% / 83.6%17.6% / 54.4%21.8% / 52.6%3.25
+ ECSO48.4% / 66.3%62.2% / 74.4%6.7% / 51.1%17.6% / 51.5%4.88
+ AdaShield57.5% / 69.5%62.8% / 73.0%31.7% / 57.7%0.0% / 49.9%4.38
+ EchoSafe78.2% / 79.8%84.5% / 86.1%54.8% / 64.5%40.2% / 57.0%2.00
+ BoundSafe80.0% / 80.5%86.1% / 86.6%66.3% / 72.7%70.1% / 70.5%1.00

Table 12: Independent Safety Evaluation by GLM-5.3-Flash on Seed-2.0-Pro. Results are reported as CCR / Avg (%).

MethodGenGenOCRChatEmbodiedAvg. Rank
Base68.3% / 75.3%82.9% / 86.0%48.0% / 65.2%13.2% / 52.6%5.13
+ FigStep85.8% / 86.4%90.5% / 91.5%51.4% / 65.9%52.4% / 65.3%2.63
+ ECSO65.4% / 73.7%82.9% / 86.3%48.5% / 64.9%21.4% / 55.1%5.00
+ AdaShield80.0% / 80.9%84.5% / 86.6%53.2% / 64.2%75.9% / 76.4%3.13
+ EchoSafe74.1% / 78.4%80.5% / 85.4%77.5% / 79.1%46.7% / 63.1%4.00
+ BoundSafe87.4% / 87.8%90.9% / 91.9%81.8% / 82.4%87.4% / 88.6%1.00

As shown in Tables 7–12 above, BoundSafe consistently achieves the best CCR and Avg across all benchmarks and target MLLMs under both independent evaluators, maintaining an average rank of 1.00 in every setting. The improvements are particularly substantial on the more challenging Chat and Embodied benchmarks. For example, on the Embodied benchmark with InternVL3-8B as the target model, BoundSafe achieves 76.1% and 70.1% CCR under DeepSeek-V4-Flash and GLM-5.3-Flash evaluation, respectively, outperforming the second-best method, EchoSafe (41.2% and 40.2%), by relative improvements of 84.7% and 74.4%. These results provide strong evidence that BoundSafe's performance gains are robust across independent evaluators.

Detailed Safety Metrics. We further report unsafe-refusal rate (URR), safe-answer rate (SAR), and paired correctness (PC) to provide a more detailed assessment of safety performance. Paired correctness measures the proportion of contrastive pairs for which both the safe and unsafe instances are handled correctly. We report these metrics for all methods under the three evaluators.

Reviewer_Q2DvOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes BoundSafe, a framework for improving multimodal safety judgments without updating model parameters. It constructs textual safety memory by contrasting similar examples with opposite safety labels, refines this memory through an iterative procedure using rotating validation folds, and retrieves relevant category-level guidance during inference. Experiments evaluate three target models across four benchmark settings, reporting improvements in the balance between refusing unsafe requests and answering safe requests. Additional analyses examine memory construction, refinement, counterpart selection, routing, and inference efficiency.

Critical strengths
  1. Addressing the problem of visually similar inputs can require different safety decisions, making generic safety instructions insufficient by learning explicit distinctions from contrasting examples is well-motivated.
  2. The method produces inspectable textual memory and requires no target-model parameter updates, making it applicable to both open and proprietary models.
Critical weaknesses
  1. The interpretation of the refinement procedure as validation on unseen data is overstated. The initial memory is synthesized from all of DcD_c, including examples subsequently designated as held-out folds, and those examples are reused across iterations. Furthermore, CCRoofCCR_{\mathrm{oof}} aggregates evaluations of different candidate memories, whereas only one candidate is ultimately selected. This aggregate therefore does not directly estimate the selected memory’s generalization performance. The separate outer test folds remain valuable, but the internal validation claim needs qualification.
  2. The experiments do not fully isolate the contribution of boundary-aware memory from supervised prompt optimization. Table 3 controls the labeled data and analyst model, but it is unclear whether alternative memory strategies receive equivalent refinement, context lengths, and model-call budgets.
Questions for authors
  1. How is CCRoofCCR_{\mathrm{oof}} computed, and why should an aggregate over different candidate memories predict the performance of the single selected candidate?
  2. What annotation policy justifies treating every irrelevant image–query pair as unsafe? How does the learned memory behave on benign requests paired with irrelevant images, and does performance transfer to new scenarios or policies?
  3. Do Case Memory and Direct Memory receive the same refinement procedure, prompt-token budget, and analyst-call budget as BoundSafe?
5 comments
AuthorsOpenReview ↗

Response to Reviewer Q2Dv (Part I)

We thank the reviewer for the constructive comments and valuable suggestions. We address each concern in detail below.

W1 & Q1: CCRoofCCR_{oof} Computation and Candidate Selection

Role of Held-Out Validation. As described in Section 4.2, during each refinement round, candidate edits are proposed based on error patterns identified from the remaining folds, without using errors from the corresponding held-out fold. Since these edits are derived from specific observed failures, they may overfit to those error patterns. We therefore use the held-out folds to examine whether the proposed edits can also improve performance on other examples.

The ablation results in Table 2 of our paper provide supporting evidence for this design. Without held-out rotation, CCR improves only marginally over the variant without refinement (61.7% vs. 61.4% on Chat with Qwen3-VL-8B-Instruct), while the full BoundSafe achieves 66.1%. This suggests that held-out validation helps select generalizable edits.

Out-of-Fold CCR Aggregation. In our Cross-Validated Refinement procedure, we first evaluate each candidate memory on its corresponding held-out fold. We then pool the held-out evaluation results across all three folds and compute the harmonic mean of the overall safe-answer and unsafe-refusal rates to obtain CCRoofCCR_{oof}.

We use CCRoofCCR_{oof} to identify promising refinement rounds rather than directly estimate the performance of the selected candidate. Three-fold validation allows us to make better use of limited labeled data and explore different refinement directions, increasing the chance of finding effective memory edits. Since the selected candidate will undergo further evaluation on the full construction set DcD_c, we aim to prioritize promising refinements while avoiding unnecessary computational overhead. Specifically, we proceed with candidate selection only when CCRoof\mathrm{CCR}_{\mathrm{oof}} exceeds CCRcsaved\mathrm{CCR}^{\mathrm{saved}}_c, indicating that the proposed refinements collectively show an improvement across the held-out folds. We then select the candidate with the largest held-out CCR gain and evaluate it on the full DcD_c. This procedure helps balance refinement effectiveness and computational efficiency.

Q2: Image–Query Relevance and Cross-Scenario Generalization

Image–Query Relevance. The emphasis on image–query relevance is grounded in the situational-safety formulation of MSSBench [1], where the surrounding environment plays an important role in determining whether an activity is safe. For example, practicing baseball in an open field differs from doing so in an office, while skateboarding in a designated skatepark differs from skateboarding on a road with traffic. As emphasized in MSSBench [1], image–query relevance is important for assessing whether the visual environment is appropriate for the requested activity. The memories illustrated in Figures 6 and 8 of the paper are automatically constructed by the analyst MLLM through contrastive analysis of labeled examples and subsequently optimized through target-model feedback. The rule concerning image–query relevance thus emerges from this learning process rather than being manually specified. Moreover, this rule is accompanied by other guidance in the same memory that helps the model recognize safe situations and avoid false alarms. As shown in Figure 6 of the paper, the memory identifies appropriate activity settings and the absence of explicit hazards as key safe signals, while its False alarm guard instructs the model not to classify relevant, low-risk activity scenes as unsafe. Together, these rules encourage the model to consider image–query relevance alongside other contextual safety cues when making its judgment.

With such memory guidance, BoundSafe can recognize image–query irrelevance while still making appropriate safety judgments based on the user's actual intent. For instance, on MSSBench Chat, Qwen3-VL-8B-Instruct with BoundSafe correctly responds to the benign query "How can I make my phone calls more engaging and interesting when talking with my son?", despite being paired with an unrelated image of a train station. The model explicitly identifies the image as irrelevant, determines that the request is safe, and provides practical suggestions for improving the conversation. This example illustrates that recognizing image–query irrelevance does not necessarily lead to over-refusal, as the model can consider the harmless intent of the request under the broader guidance of the learned memory. This behavior is also consistent with the strong CCR and Avg performance reported in Table 1 of the paper, while the cross-scenario transfer results in Tables 1–3 below provide further evidence of the generalizability of the learned boundary-aware memory.

AuthorsOpenReview ↗

Response to Reviewer Q2Dv (Part II)

Cross-Scenario Transfer. To evaluate the generalizability of learned boundary-aware memory, we further conduct cross-scenario transfer experiments across three settings: GenOCR to Gen, Gen to Embodied, and Chat to Embodied, where memory learned from the source benchmark is evaluated on the target benchmark. For comparison, we report the strongest baseline results on the corresponding target benchmarks from Table 1 of our paper.

Table 1: Cross-Scenario Transferability of Boundary-Aware Memory on Qwen3-VL-8B-Instruct. Results are reported as CCR / Avg (%).

MethodGenOCR to GenGen to EmbodiedChat to Embodied
Best Baseline80.6 / 81.261.1 / 68.461.1 / 68.4
BoundSafe83.4 / 84.571.0 / 71.876.2 / 76.4

Table 2: Cross-Scenario Transferability of Boundary-Aware Memory on InternVL3-8B. Results are reported as CCR / Avg (%).

MethodGenOCR to GenGen to EmbodiedChat to Embodied
Best Baseline72.8 / 74.843.3 / 59.743.3 / 59.7
BoundSafe75.7 / 77.061.1 / 64.068.8 / 69.3

Table 3: Cross-Scenario Transferability of Boundary-Aware Memory on Seed-2.0-Pro. Results are reported as CCR / Avg (%).

MethodGenOCR to GenGen to EmbodiedChat to Embodied
Best Baseline80.7 / 82.076.2 / 76.976.2 / 76.9
BoundSafe84.6 / 85.076.8 / 77.678.4 / 79.5

As shown in Tables 1–3 above, BoundSafe consistently outperforms the strongest baselines across all transfer settings, demonstrating that the learned boundary-aware memory remains effective when transferred to different scenarios. For example, on MSSBench Embodied with InternVL3-8B, transferring memory from MSSBench Chat achieves 68.8% CCR, compared with 43.3% for the strongest baseline, representing a 58.9% relative improvement. These results provide empirical evidence of cross-scenario generalization across different benchmarks and target MLLMs.

W2 & Q3: Refinement for Case Memory and Direct Memory

In Table 3 of our paper, Case Memory and Direct Memory use the same labeled data and analyst model as BoundSafe, but do not undergo memory refinement. To isolate the contribution of boundary-aware memory construction, we conduct additional experiments by applying the same Cross-Validated Refinement procedure (BoundRefine) to both memory strategies, with matched analyst-call and prompt-token budgets. All BoundRefine variants use the same three-fold validation and four-iteration budget. We also evaluate a budget-matched adaptation of SkillOpt [2] to Direct Memory, a text-space optimization framework that iteratively improves external skill instructions through feedback-driven edits and held-out validation.

Table 4: Performance of Different Memory Refinement Strategies on Qwen3-VL-8B-Instruct. Results are reported as CCR / Avg (%).

MethodGenGenOCRChatEmbodied
Case Memory + BoundRefine76.5 / 78.383.5 / 85.761.7 / 69.573.9 / 75.6
Direct Memory + BoundRefine83.7 / 84.387.3 / 88.650.5 / 63.670.7 / 73.6
Direct Memory + SkillOpt83.7 / 84.486.3 / 87.849.4 / 63.171.0 / 72.8
BoundSafe84.2 / 84.889.6 / 90.066.1 / 71.875.5 / 75.9

Table 5: Performance of Different Memory Refinement Strategies on InternVL3-8B. Results are reported as CCR / Avg (%).

MethodGenGenOCRChatEmbodied
Case Memory + BoundRefine73.5 / 75.480.0 / 82.365.3 / 73.274.7 / 75.4
Direct Memory + BoundRefine76.9 / 77.984.1 / 85.154.5 / 66.577.0 / 77.8
Direct Memory + SkillOpt75.9 / 76.984.3 / 85.354.8 / 66.772.9 / 73.9
BoundSafe77.5 / 78.285.6 / 86.167.7 / 73.978.0 / 78.1

Table 6: Performance of Different Memory Refinement Strategies on Seed-2.0-Pro. Results are reported as CCR / Avg (%).

MethodGenGenOCRChatEmbodied
Case Memory + BoundRefine80.2 / 81.587.7 / 89.878.4 / 78.884.5 / 85.8
Direct Memory + BoundRefine83.8 / 84.189.6 / 90.874.4 / 75.779.8 / 83.2
Direct Memory + SkillOpt81.5 / 82.388.9 / 90.369.6 / 72.372.8 / 78.9
BoundSafe84.1 / 84.691.3 / 91.978.7 / 79.488.5 / 89.6

As shown in Tables 4–6 above, BoundSafe consistently achieves the highest CCR and Avg across all benchmarks and target models. For example, on Chat with Qwen3-VL-8B-Instruct, BoundSafe achieves a CCR of 66.1%, compared with 61.7% for the strongest refined alternative. These results further demonstrate that the effectiveness of BoundSafe stems from its boundary-aware memory construction combined with refinement.

AuthorsOpenReview ↗

Response to Reviewer Q2Dv (Part III)

[1] Zhou K, Liu C, Zhao X, et al. Multimodal situational safety[C]//International Conference on Learning Representations. 2025, 2025: 10658-10688.

[2] Yang Y, Gong Z, Huang W, et al. Skillopt: Executive strategy for self-evolving agent skills[J]. arXiv preprint arXiv:2605.23904, 2026.

Reviewer_5FNzOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes BoundSafe, a training-free framework for fine-grained multimodal safety alignment. It extracts boundary-aware rules from highly similar image-text pairs with opposite safety labels, refines the resulting memory through cross-validated error correction, and retrieves relevant memory to guide inference without updating model parameters. Experiments on four multimodal safety benchmarks across three MLLMs demonstrate consistent improvements over existing baselines, supported by ablation studies and comparisons with alternative memory strategies.

Critical strengths
  1. Well-motivated and practical framework.

The paper addresses an important challenge in multimodal safety alignment: distinguishing visually similar inputs that require different safety judgments. The proposed boundary-aware memory provides explicit, reusable safety guidance without modifying model parameters, offering a practical alternative to training-based approaches and case-level memory methods.

  1. Strong empirical performance and comprehensive evaluation.

BoundSafe consistently outperforms existing baselines across four benchmarks and three target MLLMs, with particularly substantial improvements on context-dependent safety tasks. The ablation studies and comparisons with alternative memory strategies further support the effectiveness of the proposed design.

Critical weaknesses

1. Limited evidence of memory generalizability and dependence on the analyst model.

BoundSafe constructs and refines memory based on target-model-specific errors, but the paper does not investigate whether the learned memory can transfer across different target MLLMs without reconstruction. Moreover, all experiments rely on Seed-2.0-Pro as the analyst model, without evaluating sensitivity to analyst capability. Since the same model is also used for response safety evaluation, the lack of independent validation further limits confidence in the robustness and generalizability of the reported improvements.

2. Insufficient evaluation of deployment overhead and general-purpose utility.

BoundSafe introduces an additional 2B-parameter embedding model and injects safety memory during inference. Although the paper reports inference latency comparisons, it does not fully characterize the additional computational and memory costs. Furthermore, the proposed pipeline appears to apply memory guidance to every incoming query, including benign ones. The paper does not evaluate whether such guidance affects general-purpose multimodal capabilities (e.g., VQA) or introduces unnecessary overhead for non-safety-related requests.

Questions for authors

1. Cross-model memory transferability.

Does each target MLLM require separately constructed and refined memory? Could the authors evaluate whether memory learned for one target model can be directly transferred to other models without reconstruction?

2. Computational overhead.

Could the authors provide a breakdown of the additional inference latency, GPU memory usage, and prompt token overhead introduced by the embedding model and memory injection? It would also be helpful to report the offline cost of memory construction and refinement.

3. General-purpose capability preservation.

Does BoundSafe inject safety memory for all queries, including clearly benign VQA requests? If so, does this affect general-purpose multimodal performance? Could the authors evaluate performance on standard benign benchmarks or clarify whether a mechanism exists to bypass unnecessary memory injection?

4. Analyst model dependence and evaluation reliability.

How sensitive is BoundSafe to the choice of analyst model? Could the authors compare different analyst MLLMs? Additionally, since Seed-2.0-Pro is used for both memory construction and safety evaluation, could the authors provide independent evaluation results to rule out potential evaluator-specific bias?

13 comments
AuthorsOpenReview ↗

Response to Reviewer 5FNz (Part I)

We thank the reviewer for the constructive comments and valuable suggestions. We address each concern in detail below.

W1 & Q1: Cross-model Memory Transferability

In our framework, we construct and refine memory for each target MLLM to better accommodate its specific safety capabilities and failure patterns. As described in Section 4.1, BoundSafe probes the target model to identify which safety boundaries it already handles well and which remain challenging, allowing subsequent boundary analysis to focus on the actual gaps. This target-specific feedback allows the resulting memory to address the particular weaknesses of each target MLLM.

Nevertheless, the learned memory captures generalizable cross-modal safety cues and decision rules, which can be transferred across different MLLMs. To investigate this, we conduct additional cross-model memory transfer experiments using memories constructed and refined with three source models: Qwen3-VL-8B-Instruct, InternVL3-8B, and Seed-2.0-Pro. We evaluate two target models, InternVL3-8B and Seed-2.0-Pro, on all four benchmarks. For each transfer setting, we directly apply the memory learned from a different source model to the target MLLM, without reconstructing or refining the memory using feedback from the target model. We compare the transferred memory with both the target-specific memory (i.e., memory constructed and refined using feedback from the target MLLM itself) and the strongest baseline reported in Table 1 of the paper.

Table 1: Cross-Model Transferability of Boundary-Aware Memory on InternVL3-8B. Results are reported as CCR / Avg (%).

Memory Source / MethodGenGenOCRChatEmbodied
Best Baseline72.8 / 74.881.1 / 83.453.3 / 63.643.3 / 59.7
InternVL3-8B (Target-Specific)77.5 / 78.285.6 / 86.167.7 / 73.978.0 / 78.1
Qwen3-VL-8B-Instruct76.9 / 77.586.0 / 86.468.1 / 74.178.2 / 78.3
Seed-2.0-Pro78.5 / 79.184.4 / 84.965.1 / 73.672.3 / 72.7

Table 2: Cross-Model Transferability of Boundary-Aware Memory on Seed-2.0-Pro. Results are reported as CCR / Avg (%).

Memory Source / MethodGenGenOCRChatEmbodied
Best Baseline80.7 / 82.089.7 / 90.471.8 / 75.576.2 / 76.9
Seed-2.0-Pro (Target-Specific)84.1 / 84.691.3 / 91.978.7 / 79.488.5 / 89.6
Qwen3-VL-8B-Instruct81.4 / 82.491.5 / 91.875.8 / 76.482.2 / 85.4
InternVL3-8B82.1 / 82.990.2 / 91.076.6 / 77.379.2 / 84.3

As shown in Tables 1 and 2 above, the transferred memories consistently outperform the strongest baselines in both CCR and Avg across all cross-model transfer settings, demonstrating their effectiveness across different MLLMs. Notably, when InternVL3-8B uses memory transferred from Qwen3-VL-8B-Instruct, it achieves CCRs of 86.0%, 68.1%, and 78.2% on GenOCR, Chat, and Embodied, respectively, slightly exceeding its target-specific results of 85.6%, 67.7%, and 78.0%. This suggests that the learned boundary-aware memory captures generalizable safety knowledge that can be effectively reused across different MLLMs. Meanwhile, target-specific memory still provides additional benefits in some settings by addressing the actual safety gaps and failure patterns of the target MLLM. For example, on Embodied, Seed-2.0-Pro achieves 88.5% CCR with its own memory, compared with 82.2% and 79.2% using memories transferred from Qwen3-VL-8B-Instruct and InternVL3-8B, respectively, both of which still outperform the strongest baseline. These findings demonstrate that BoundSafe's boundary-aware memory can be directly reused across different MLLMs without reconstruction or refinement, while target-specific feedback can further improve performance by addressing model-dependent failure patterns.

W2 & Q2: Computational Overhead

Inference Overhead. As suggested, we conduct a detailed breakdown of the inference latency, GPU memory usage, and prompt token overhead across all four benchmarks. Specifically, we measure the computational cost of Qwen3-VL-Embedding-2B and compare target-model inference with and without memory injection for Qwen3-VL-8B-Instruct and InternVL3-8B.

Table 3: Computational Cost of Qwen3-VL-Embedding-2B.

MetricGenGenOCRChatEmbodied
Latency (s/query)0.2160.2180.2750.170
GPU Memory (MB)939094411125611453

Table 4: Inference Overhead on Qwen3-VL-8B-Instruct. Results are reported as latency (s/query) / GPU memory (MB) / prompt tokens.

MethodGenGenOCRChatEmbodied
BoundSafe w/o Memory2.095 / 88876 / 1401.72.068 / 88876 / 1401.71.960 / 90058 / 1410.70.562 / 90059 / 811.4
BoundSafe2.406 / 88876 / 2244.92.677 / 88876 / 2243.01.998 / 90060 / 2250.20.772 / 90060 / 1576.6
AuthorsOpenReview ↗

Response to Reviewer 5FNz (Part II)

Table 5: Inference Overhead on InternVL3-8B. Results are reported as latency (s/query) / GPU memory (MB) / prompt tokens.

MethodGenGenOCRChatEmbodied
BoundSafe w/o Memory1.603 / 85226 / 2937.71.565 / 85225 / 2937.71.120 / 85225 / 2519.40.555 / 85226 / 1207.4
BoundSafe1.867 / 85226 / 3781.21.843 / 85226 / 3702.61.215 / 85228 / 3315.20.754 / 85228 / 1912.2

As shown in Table 3 above, Qwen3-VL-Embedding-2B introduces only about 0.2 seconds of latency per query, substantially lower than the target MLLMs' inference latency reported in Tables 4 and 5 above. In terms of GPU memory, the embedding model requires approximately 9.4–11.5 GB, substantially less than the 85–90 GB consumed by the target MLLMs, representing only about 10–13% of their GPU memory usage. These results indicate that the lightweight embedding model enables efficient memory routing with relatively modest computational overhead.

Tables 4 and 5 above further show that memory injection adds approximately 700–850 prompt tokens per query, yet increases inference latency by less than 0.3 seconds in most settings. Meanwhile, the target models' GPU memory usage remains virtually unchanged. These results demonstrate that BoundSafe can effectively integrate boundary-aware memory into target MLLMs with modest inference latency overhead and minimal additional GPU memory consumption.

Offline Construction and Refinement Costs. We also measure the offline cost of memory construction and refinement. As described in Sections 4.1 and 4.2, these stages construct boundary-aware memory and iteratively improve it through cross-validated editing. Both stages are performed offline, and the resulting memory can be reused during inference without repeating the construction or refinement process.

Table 6: Offline Cost of Memory Construction and Refinement for Qwen3-VL-8B-Instruct. All values are reported in minutes.

StageGenGenOCRChatEmbodied
Construction9.810.48.42.2
Refinement46.664.737.011.9

Table 7: Offline Cost of Memory Construction and Refinement for InternVL3-8B. All values are reported in minutes.

StageGenGenOCRChatEmbodied
Construction7.98.15.31.7
Refinement33.432.824.310.9

As shown in Tables 6 and 7 above, memory construction takes only a few minutes, while refinement accounts for most of the offline computational cost. However, as demonstrated by the ablation study in Table 2 of the paper, BoundSafe without refinement (w/o iteration) already outperforms all baselines in most benchmarks, indicating that the initially constructed memory is already effective, while refinement further improves its performance. Moreover, both memory construction and refinement are performed offline and only need to be completed once. The resulting memory can then be reused across large-scale inference queries without additional construction or refinement, allowing the offline computational cost to be amortized over deployment.

W2 & Q3: General-Purpose Capability Preservation.

As illustrated by the example in Figure 1 of our paper, multimodal safety judgments often depend on subtle cross-modal cues. Even a seemingly benign query, such as asking how to maintain balance while skateboarding, can become unsafe when paired with an image showing a road with traffic in the background. Therefore, BoundSafe applies memory guidance to all incoming queries, including those that appear benign, to ensure that potentially unsafe queries are not overlooked.

To examine whether this design affects general-purpose multimodal capabilities, we conduct additional experiments on the MME benchmark [1], which evaluates both perception and cognition abilities across 14 subtasks, including 10 perception tasks and 4 cognition tasks. Following the official evaluation protocol, we report the Perception, Cognition, and Total scores. We evaluate Qwen3-VL-8B-Instruct and Seed-2.0-Pro, comparing their original performance (Base) with performance under BoundSafe's memory guidance.

Table 8: General-Purpose Multimodal Performance on MME with Qwen3-VL-8B-Instruct.

MethodPerceptionCognitionTotal
Base1750.4676.12426.5
+ BoundSafe1678.8672.12351.0

Table 9: General-Purpose Multimodal Performance on MME with Seed-2.0-Pro.

MethodPerceptionCognitionTotal
Base1791.1744.62535.7
+ BoundSafe1817.0751.82568.7
AuthorsOpenReview ↗

Response to Reviewer 5FNz (Part III)

As shown in Tables 8 and 9 above, BoundSafe largely preserves the general-purpose multimodal capabilities of both target MLLMs. On Qwen3-VL-8B-Instruct, BoundSafe incurs only a marginal decrease in the MME score. Notably, BoundSafe even improves the performance of Seed-2.0-Pro, increasing its total score from 2535.7 to 2568.7, with gains in both Perception and Cognition. That’s because our boundary-aware memory encourages the model to process queries more carefully, which ultimately improves its performance.

W1 & Q4: Analyst Model Dependence and Evaluation Reliability

Analyst Model Sensitivity. To investigate the sensitivity of BoundSafe to the choice of analyst model, we conduct additional experiments using two alternative analyst MLLMs, GLM-5.3-Flash and Qwen3.8-27B, in addition to the original Seed-2.0-Pro. We evaluate all three target MLLMs across four benchmarks, following the same experimental protocol as in the paper. The results are presented in Tables 10–12 below.

Table 10: Effect of Different Analyst Models on Qwen3-VL-8B-Instruct. Results are reported as CCR / Avg (%).

Analyst modelGenGenOCRChatEmbodied
Best Baseline80.6 / 81.288.5 / 89.752.2 / 62.461.1 / 68.4
Seed-2.0-Pro84.2 / 84.889.6 / 90.066.1 / 71.875.5 / 75.9
GLM-5.3-Flash86.3 / 86.589.1 / 89.766.2 / 71.877.6 / 77.7
Qwen3.8-27B84.9 / 85.290.6 / 90.869.0 / 74.178.6 / 78.8

Table 11: Effect of Different Analyst Models on InternVL3-8B. Results are reported as CCR / Avg (%).

Analyst modelGenGenOCRChatEmbodied
Best Baseline72.8 / 74.881.1 / 83.453.3 / 63.643.3 / 59.7
Seed-2.0-Pro77.5 / 78.285.6 / 86.167.7 / 73.978.0 / 78.1
GLM-5.3-Flash81.9 / 82.086.3 / 86.870.0 / 75.379.1 / 79.2
Qwen3.8-27B80.2 / 80.485.2 / 87.174.7 / 78.274.9 / 75.5

Table 12: Effect of Different Analyst Models on Seed-2.0-Pro. Results are reported as CCR / Avg (%).

Analyst modelGenGenOCRChatEmbodied
Best Baseline80.7 / 82.089.7 / 90.471.8 / 75.576.2 / 76.9
Seed-2.0-Pro84.1 / 84.691.3 / 91.978.7 / 79.488.5 / 89.6
GLM-5.3-Flash84.2 / 84.891.4 / 91.778.7 / 79.292.1 / 92.6
Qwen3.8-27B84.6 / 85.092.2 / 92.374.7 / 75.692.7 / 93.1

As shown in Tables 10–12 above, BoundSafe consistently outperforms the strongest baselines across all benchmarks and target MLLMs, regardless of the analyst model used. Notably, alternative analyst models often achieve comparable or even better performance than Seed-2.0-Pro. For example, using Qwen3.8-27B as the analyst improves the Chat CCR of InternVL3-8B from 67.7% to 74.7%, while increasing the Embodied CCR of Seed-2.0-Pro from 88.5% to 92.7%. These findings demonstrate that BoundSafe is robust to different analyst models and does not depend on a specific MLLM for effective memory construction and refinement.

Independent Evaluator Robustness. To address the concern about evaluator-specific bias, we further conduct independent safety evaluations using DeepSeek-V4-Flash and GLM-5.3-Flash as alternative evaluators, replacing Seed-2.0-Pro used in the original evaluation. We evaluate the responses of all methods across the three target MLLMs and four benchmarks. The results are reported in Tables 13–18 below.

Table 13: Independent Safety Evaluation by DeepSeek-V4-Flash on Qwen3-VL-8B-Instruct. Results are reported as CCR / Avg (%).

MethodGenGenOCRChatEmbodiedAvg. Rank
Base71.8% / 77.2%88.1% / 88.7%33.2% / 57.6%28.2% / 56.2%3.88
+ FigStep80.9% / 81.1%83.6% / 86.9%27.8% / 55.6%57.0% / 57.2%3.50
+ ECSO73.8% / 77.9%88.5% / 89.0%29.4% / 56.3%30.0% / 56.5%3.38
+ AdaShield39.4% / 51.9%35.2% / 49.8%16.0% / 52.4%0.0% / 50.0%6.00
+ EchoSafe74.4% / 77.2%74.7% / 80.2%52.2% / 63.3%66.4% / 71.9%3.13
+ BoundSafe84.7% / 85.3%89.0% / 89.9%63.1% / 71.4%74.4% / 74.9%1.00

Table 14: Independent Safety Evaluation by DeepSeek-V4-Flash on InternVL3-8B. Results are reported as CCR / Avg (%).

MethodGenGenOCRChatEmbodiedAvg. Rank
Base42.9% / 63.4%57.8% / 71.1%7.1% / 50.7%10.5% / 51.4%5.13
+ FigStep58.3% / 69.3%80.8% / 83.3%15.3% / 54.0%40.7% / 57.0%3.00
+ ECSO48.9% / 65.8%65.0% / 75.5%8.3% / 50.7%13.5% / 52.4%4.25
+ AdaShield33.7% / 53.4%42.8% / 56.1%32.1% / 57.8%0.0% / 49.9%5.25
+ EchoSafe74.2% / 76.4%79.3% / 82.6%54.4% / 64.8%41.2% / 57.7%2.25
+ BoundSafe78.7% / 79.4%85.5% / 86.3%68.4% / 74.5%76.1% / 76.2%1.00

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

We sincerely thank the Area Chair, Senior Area Chairs, Program Chairs, and all reviewers for their time, careful consideration, and constructive feedback throughout the review and rebuttal process.

A central challenge for multimodal safety alignment is correctly distinguishing ambiguous cases near the safety boundary, where highly similar multimodal inputs may require opposite safety judgments. To address this challenge, our paper proposes BoundSafe, a training-free framework for fine-grained multimodal safety alignment by learning generalizable safety boundaries from limited labeled examples. BoundSafe consistently outperforms existing baselines across four benchmarks and three target MLLMs. For example, on MSSBench Embodied with InternVL3-8B, BoundSafe achieves 80.1% and 30.9% relative improvements in CCR and Avg, respectively, over the strongest baseline. During the rebuttal, we provided detailed clarifications and additional experiments to address the reviewers' concerns.

Reviewer 5FNz. We demonstrated cross-model memory transferability, evaluated sensitivity to different analyst models, and confirmed the robustness of our findings using independent safety evaluators. We also provided detailed computational overhead measurements, general-purpose evaluations, and safety–utility–efficiency comparisons with strong baselines. The reviewer explicitly acknowledged that our rebuttal had addressed most of the concerns and subsequently raised the rating.

Reviewer Q2Dv. We clarified the computation and role of out-of-fold CCR, candidate selection, and internal held-out validation. We conducted matched-refinement experiments to better isolate the contribution of boundary-aware construction. We also provided cross-scenario transfer experiments and clarified the role of image–query relevance in situational safety. The reviewer explicitly confirmed that our clarifications and experiments had addressed the main concerns and recommended acceptance.

Reviewer v1au. We provided matched-refinement comparisons, cross-scenario transfer experiments, independent safety evaluations, detailed safety metrics, response-quality evaluations, and fixed-validation comparisons. The reviewer acknowledged that these experiments addressed several questions and provided stronger evidence for boundary-aware construction, while raising three remaining concerns. We responded to all three follow-up concerns by clarifying the benefits of boundary-aware construction, conducting an additional ablation of the image–query relevance rule, extending response-quality evaluation to Chat and Embodied, and evaluating visual safety cue grounding across all four benchmarks.

Overall, through detailed clarifications and additional experiments, our rebuttal provides stronger support for BoundSafe's effectiveness, efficiency, generalizability, robustness, and general-purpose utility. We sincerely appreciate the Area Chair, Senior Area Chairs, Program Chairs, and reviewers for their time and thoughtful consideration of our work.

Best regards,

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.