Pith. sign in

REVIEW 3 major objections 32 references

Image safety is not a property of the image: guardrails must flip decisions when the runtime policy changes.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-11 01:48 UTC pith:JTKKJ7SD

load-bearing objection Solid systems paper: same-image policy flips are the real contribution; SOTA is real on their bench but rests on a small, rule-labeled image set. the 3 major comments →

arxiv 2607.05910 v1 pith:JTKKJ7SD submitted 2026-07-07 cs.CV cs.AIcs.CL

PolicyShiftGuard: Benchmarking and Improving Policy-Adaptive Image Guardrails

Mingyang Song , Luxin Xu , Haoyu Sun , Minzhou Pan , Yu Cheng , Bo Li This is my paper
classification cs.CV cs.AIcs.CL
keywords image guardrailspolicy-adaptive safetyvision-language modelsPolicyShiftBenchboundary-pair trainingcontent moderationpolicy shift score
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Most image guardrails treat safety as if it were baked into the picture. In real products the same photo can be fine for one audience, blocked for another, and newly disallowed when a rule is rewritten. This paper argues that a useful guardrail must read the currently supplied policy, bind it to fine-grained visual evidence, and revise its verdict when the boundary moves. To measure that skill it builds PolicyShiftBench: 2,000 policy-discriminative cases over 265 images, each image judged under multiple policies so the same visual evidence must sometimes pass and sometimes block. It then trains PolicyShiftGuard with a two-stage recipe: first randomized policy supervised fine-tuning so the model actually follows full policy bundles, then Boundary-Pair Policy Adaptation that pairs the identical image under a blocking policy and a passing policy and forces the unsafe score to separate. The resulting 7B model reaches 76.9 average F1 and 72.1 average Policy Shift Score, far above ordinary vision-language models and specialized guardrails that often detect risky cues yet almost never flip with the policy, while also running faster via a five-token decision format.

Core claim

Existing vision-language models and specialized image guardrails largely fail at policy adaptation: they can score nontrivial F1 on unsafe content yet collapse on Policy Shift Score, showing they rely on image-level priors rather than the active rule. A compact policy-conditioned model trained with randomized policy SFT plus matched pass/block boundary-pair adaptation closes most of that gap, reaching state-of-the-art 76.9 Avg. F1 and 72.1 Avg. PSS on PolicyShiftBench and transferring to external safety benchmarks.

What carries the argument

Boundary-Pair Policy Adaptation (BP-Adapt): for a fixed image and risk category, train one policy that must block and one that must pass, combining ordinary label losses with a pairwise margin that makes the blocking policy receive a higher unsafe score than the passing policy under identical visual evidence.

Load-bearing premise

Ground-truth labels are treated as correct once three vision models majority-vote atomic image attributes and the authors’ hand-written policy rules are executed on those attributes.

What would settle it

Hold out a new set of real product policies never seen in training or the benchmark catalog; if PolicyShiftGuard’s Policy Shift Score on same-image pass/block flips falls back toward the near-zero scores of ordinary guardrails, the claimed adaptation fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 0 minor

Summary. The paper argues that image safety is policy-dependent rather than an intrinsic image property, and that current VLMs and specialized guardrails fail under policy shifts. It introduces PolicyShiftBench (2,000 policy-discriminative instances over 265 images, 28 policy variants, Adaptive vs held-out Shift splits) and Policy Shift Score (PSS), which scores same-image pass/block flips. It then proposes PolicyShiftGuard (3B/7B), trained with Randomized Policy SFT followed by Boundary-Pair Policy Adaptation that combines CE with a pairwise margin separating matched pass/block policies. On PolicyShiftBench the 7B model reports 76.9 Avg. F1 and 72.1 Avg. PSS (Table 2), with transfer gains on UnSafeBench and SafeEditBench (Table 3), a favorable latency–performance trade-off (Figure 3), and ablations isolating randomization, no-think supervision, and the pair loss (Tables 4–6; Figure 5).

Significance. Policy-adaptive visual guardrailing is a genuine deployment gap: product, age, region, and institutional policies routinely re-label the same image. The paper’s main contributions are concrete and useful: (i) a compositional, scenario-grounded benchmark with executable rules over atomic attributes; (ii) PSS, which cleanly separates ordinary unsafe recognition from policy-flip sensitivity; (iii) a compact two-stage recipe whose pair-loss ablation is large and interpretable (Table 6: +17.7 Avg. F1 for 7B when L_pair is added). Strengths include image-disjoint train/eval construction, Adaptive vs held-out Shift policies, multi-model baselines including closed-source systems, transfer evaluation, latency reporting, and rerun variation in Appendix E.2. If the benchmark and labels are accepted as a reasonable proxy for runtime policy bundles, the result is a clear advance over fixed-taxonomy image guardrails and a practical training recipe for compact policy-conditioned models.

major comments (3)
  1. Table 2 / §2.1: The headline SOTA (76.9 Avg. F1, 72.1 Avg. PSS) is measured on only 265 unique images (2,000 instances). PSS further macro-averages over same-image, same-category flip groups (§E.1), so a modest number of boundary images can dominate the metric. The manuscript should report image-level or group-level variance (e.g., bootstrap over images/groups) and discuss how sensitive Avg. PSS is to removing high-leverage images; without this, the SOTA claim is harder to interpret than ordinary F1 on large fixed-taxonomy sets.
  2. §2.2–2.3 and Appendix B–C: Ground-truth labels are produced by three-VLM majority-voted attributes plus author-written executable rules. Human audits (88%/90% blind accuracy; 95% qualified rate) check consistency with that same rule catalog, not external product/legal policies. This is load-bearing for both the benchmark and BP-Adapt: if attributes or exemptions (medical, news, cultural, OCR, etc.) systematically mis-specify real boundaries, models optimize the authors’ rule engine. The paper should more explicitly scope claims as ‘policy-rule following under this catalog,’ add failure analysis of attribute errors that flip labels, and avoid overstating fidelity to real deployment policies without external policy-expert validation.
  3. Table 2, Adaptive vs Shift: PolicyShiftGuard-7B remains strong on Shift PSS (70.4) but drops sharply in Shift F1 (86.8 → 67.0), while Gemini-3-Flash-Preview leads Shift F1 (74.2). The average SOTA therefore mixes in-distribution policy following with held-out generalization. The main claim should more carefully separate ‘best average policy-adaptive score on this bench’ from ‘best held-out policy generalization,’ and discuss why F1 degrades more than PSS on Shift.

Circularity Check

0 steps flagged

No derivation-by-construction circularity; mild shared-construction risk only because train and eval labels both come from the authors' rule engine over the same attribute schema.

full rationale

This is an empirical systems paper, not a first-principles derivation. Labels are produced by majority-voted VLM attributes plus deterministic author-written policy rules (Sec. 2.2–2.3; App. B–C), then used as supervised targets; the trained model does not define those labels. RP-SFT trains on Adaptive-family policies, BP-Adapt uses training-only augmentation policies never seen at test time, and the Shift split uses held-out policy definitions; training images are image-disjoint from evaluation (Table 7). Transfer to UnSafeBench and SafeEditBench is external. L_pair optimizes a soft margin on train boundary pairs while PSS scores hard flip accuracy on held-out image/policy groups—aligned training, not identity by construction. No self-definitional equations, no fitted parameter renamed as a prediction, and no load-bearing uniqueness theorem imported from overlapping-author citations. The residual score-1 concern is only that headline SOTA is measured primarily against an author-constructed rule catalog (so success means matching that catalog), which is a validity/generalization issue rather than circular derivation. Honest finding: no significant circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 4 invented entities

The central empirical claim rests on a small set of modeling and data-construction choices rather than free physical constants: loss weights and training hyperparameters, the assumption that safety is a relation between image attributes and executable policy text, and the reliability of multi-VLM attribute voting plus author-written rules as gold. Invented constructs are the benchmark, PSS metric, and training stages; they are operationally defined and partially validated by human audit and external transfer, not pure postulates.

free parameters (3)
  • BP-Adapt loss weights λ_l, λ_p, λ_w = 0.10, 0.20, 0.05
    Stage-2 objective uses L_CE + λ_l L_label + λ_p L_pair + λ_w L_cat; main recipe uses 0.10 / 0.20 / 0.05. Figure 5 shows mild sensitivity, but reported SOTA depends on these chosen weights.
  • RP-SFT / BP-Adapt learning rates and epochs = LR 1e-5 (4 ep), LR 1e-6 (1 ep)
    Stage-1 uses LR 1e-5 for 4 epochs; Stage-2 uses LR 1e-6 for 1 epoch on 1,908 boundary groups. These are hand-chosen training hyperparameters that affect final F1/PSS.
  • Pair margin m in L_pair
    The hinge margin that forces higher unsafe score under the blocking policy than the passing policy is a design hyperparameter of BP-Adapt.
axioms (4)
  • domain assumption Image safety is not an intrinsic image property but a relation between observed attributes and the currently active policy bundle.
    Stated in the introduction and used to justify same-image multi-policy evaluation and PSS.
  • domain assumption Majority vote over three multimodal annotators yields sufficiently accurate atomic attributes for deterministic rule labeling.
    Section 2.2–2.3 and Appendix B.5 report 97.5% unanimous field decisions; labels and training targets depend on this.
  • ad hoc to paper The authors’ seven-category, scenario-grounded executable policy catalog is a valid proxy for realistic product/region/age moderation policies.
    Appendix C defines all Adaptive/Shift/Augmentation rules; both benchmark gold and training supervision are generated from this catalog.
  • domain assumption A concise final-token decision format (true|c / false) is an adequate and preferable guardrail interface versus free-form reasoning for deployment metrics.
    Used throughout training and evaluation; ablations claim no-think beats think-mode on Avg. F1/PSS.
invented entities (4)
  • PolicyShiftBench independent evidence
    purpose: Provide policy-discriminative same-image evaluation with Adaptive and held-out Shift splits.
    New benchmark with 2,000 instances over 265 images and 28 evaluation policy variants; independent of prior fixed-taxonomy image safety sets.
  • Policy Shift Score (PSS) independent evidence
    purpose: Score only correct handling of same-image pass/block policy flips.
    Formally defined in Appendix E.1 as a group macro-average over flip pairs; not a standard metric in prior image safety benchmarks.
  • Boundary-Pair Policy Adaptation (BP-Adapt) no independent evidence
    purpose: Train matched pass/block prompts for the same image and category with CE plus pairwise margin.
    Stage-2 recipe is the paper’s main methodological invention; ablations attribute large F1/PSS gains to L_pair.
  • PolicyShiftGuard-3B/7B independent evidence
    purpose: Compact policy-conditioned image guardrail models.
    Models are fine-tunes of Qwen2.5-VL; the entity is the trained system plus output format, not a new physical object.

pith-pipeline@v1.1.0-grok45 · 48890 in / 3685 out tokens · 43200 ms · 2026-07-11T01:48:09.309600+00:00 · methodology

0 comments
read the original abstract

Image guardrails are typically trained and evaluated under a fixed safety policy, implicitly treating safety as an intrinsic property of an image. Real deployments are different: the same image may be allowed in one product, restricted in another, and newly disallowed when a policy boundary changes. We study policy-adaptive image guardrailing, where a model must decide whether an image violates the currently supplied policy and generalize to held-out policy definitions. We introduce PolicyShiftBench, a comprehensive benchmark with 2,000 policy-discriminative instances over 265 images, where each image is paired with 7.55 policy-conditioned prompts on average to test whether models adapt to the active policy rather than relying on image-level safety priors. We then propose PolicyShiftGuard, a compact policy-conditioned guardrail trained with a two-stage training recipe that combines Randomized Policy SFT (RP-SFT) with Boundary-Pair Policy Adaptation (BP-Adapt). BP-Adapt trains matched prompts for the same image and risk category using standard label supervision and a pairwise comparison loss that separates blocking policies from passing policies. Experiments show that existing VLMs and specialized guardrails remain brittle under policy shifts, while PolicyShiftGuard substantially improves policy-sensitive performance. The 7B model achieves SOTA performance of 76.9 Avg. F1 and 72.1 Avg. PSS on PolicyShiftBench, transfers well to UnSafeBench and SafeEditBench, and improves the latency-performance trade-off with a concise output format. Ablations confirm that matched pass/block boundary pairs are essential for stable policy adaptation.

Figures

Figures reproduced from arXiv: 2607.05910 by Bo Li, Haoyu Sun, Luxin Xu, Mingyang Song, Minzhou Pan, Yu Cheng.

Figure 1
Figure 1. Figure 1: Structure of POLICYSHIFTBENCH. (a) Dataset statistics organized by risk category and scenarios, highlighting the distribution of different policies across the benchmark. (b) Scenario– category policy matrix. Each policy is defined by combining a risk category with a realistic moderation scenario. Augmentation policies are used for Boundary-Policy Adaptation. • Extensive experiments demonstrate that existin… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the PolicyShiftGuard pipeline. (a) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Cost-performance trade￾off. We plot the main table Avg. F1 against decision speed, computed as 1000/Time(ms). Moving upward means stronger average performance, while mov￾ing right means faster inference [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Models’ performance across various risk categories. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Ablation over loss weights. We study the robustness of Boundary-Pair-SFT by varying three loss weights, λw, λp, and λl . This ablation study is conducted on Qwen2.5-VL-7B. 3B and +12.3 for 7B in stage 2. These results suggest that policy-adaptive guardrailing benefits more from directly optimizing the final decision tokens than free-form rationale. More post-training data alone is not enough [PITH_FULL_IM… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative policy-flip case 1. Nudity and sexual-content example. The same image is permitted under a more context-sensitive policy but blocked under a stricter family-safe policy. Qualitative Result of PolicyShiftGuard Case 02 Same image · different active policy · different gold decision Risk Category 01 · Nudity, Sexual Content & Fetish adaptive split· image 5512 · policy-sensitive sample Two active po… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative policy-flip case 2. Nudity and sexual-content example. The policy change turns a pass decision into a category-01 block decision. 38 [PITH_FULL_IMAGE:figures/full_fig_p038_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative policy-flip case 3 [PITH_FULL_IMAGE:figures/full_fig_p039_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative policy-flip case 4. Violence, hate, terrorism, and self-harm example. The policy pair tests whether the model can distinguish an allowed contextual case from a blocked category-02 case. 39 [PITH_FULL_IMAGE:figures/full_fig_p039_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative policy-flip case 5. Regulated goods and substances example. The policy flip depends on whether the active rule allows static or educational presentation versus blocking the regulated-good cue. Qualitative Result of PolicyShiftGuard Case 06 Same image · different active policy · different gold decision Risk Category 04 · IP, Copyright & Brand Safety shift split· image 2030 · policy-sensitive sa… view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative policy-flip case 6. IP, copyright, and brand-safety example from the Shift Split. The same visual evidence is allowed under one policy and blocked under a held-out IP/brand policy. 40 [PITH_FULL_IMAGE:figures/full_fig_p040_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Qualitative policy-flip case 7 [PITH_FULL_IMAGE:figures/full_fig_p041_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Qualitative policy-flip case 8. Cultural and religious sensitivity example from the Shift Split. The case tests whether the model follows a culturally specific policy rather than a generic safe/unsafe prior. 41 [PITH_FULL_IMAGE:figures/full_fig_p041_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Qualitative policy-flip case 9. Cultural and religious sensitivity example from the Shift Split. The same image remains acceptable under one cultural policy but is blocked under another policy with stricter taboo constraints. Qualitative Result of PolicyShiftGuard Case 10 Same image · different active policy · different gold decision Risk Category 06 · Privacy & PII adaptive split· image 654 · policy-sens… view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative policy-flip case 10. Privacy and PII example. Baselines often confuse the case with generic nudity or fail to block under the privacy-specific policy, while the correct unsafe branch requires category 06. 42 [PITH_FULL_IMAGE:figures/full_fig_p042_15.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 32 canonical work pages · 4 internal anchors

  1. [1]

    2024 , eprint =

    UnsafeBench: Benchmarking Image Safety Classifiers on Real-World and AI-Generated Images , author =. 2024 , eprint =

    work page 2024

  2. [2]

    MM-SafetyBench: A Benchmark for Safety Evaluation of Multimodal Large Language Models

    Liu, Xin and Zhu, Yichen and Gu, Jindong and Lan, Yunshi and Yang, Chao and Qiao, Yu , year =. doi:10.48550/arXiv.2311.17600 , url =. 2311.17600 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2311.17600

  3. [3]

    2025 , eprint =

    Qwen2.5-VL Technical Report , author =. 2025 , eprint =

    work page 2025

  4. [4]

    2026 , eprint =

    Qwen3.5-Omni Technical Report , author =. 2026 , eprint =

    work page 2026

  5. [5]

    2023 , eprint =

    Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations , author =. 2023 , eprint =

    work page 2023

  6. [6]

    LlavaGuard: An Open VLM-based Framework for Safeguarding Vision Datasets and Models

    Helff, Lukas and Friedrich, Felix and Brack, Manuel and Kersting, Kristian and Schramowski, Patrick , year =. doi:10.48550/arXiv.2406.05113 , url =. 2406.05113 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2406.05113

  7. [7]

    2024 , eprint =

    Safety Fine-Tuning at (Almost) No Cost: A Baseline for Vision Large Language Models , author =. 2024 , eprint =

    work page 2024

  8. [8]

    2024 , eprint =

    SafeWatch: An Efficient Safety-Policy Following Video Guardrail Model with Transparent Explanations , author =. 2024 , eprint =

    work page 2024

  9. [9]

    2025 , eprint =

    SafeVision: Efficient Image Guardrail with Robust Policy Adherence and Explainability , author =. 2025 , eprint =

    work page 2025

  10. [10]

    VSCBench: Bridging the Gap in Vision-Language Model Safety Calibration

    Geng, Jiahui and Li, Qing and Chen, Zongxiong and Wang, Yuxia and Zhu, Derui and Xie, Zhuohan and Lyu, Chenyang and Chen, Xiuying and Nakov, Preslav and Karray, Fakhri , year =. doi:10.48550/arXiv.2505.20362 , url =. 2505.20362 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2505.20362

  11. [11]

    2024 , eprint =

    ShieldGemma: Generative AI Content Moderation Based on Gemma , author =. 2024 , eprint =

    work page 2024

  12. [12]

    2024 , eprint =

    WildGuard: Open One-Stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMs , author =. 2024 , eprint =

    work page 2024

  13. [13]

    2023 , eprint =

    FigStep: Jailbreaking Large Vision-Language Models via Typographic Visual Prompts , author =. 2023 , eprint =

    work page 2023

  14. [14]

    VLMGuard: Bootstrapping Malicious Prompt Detectors from Unlabeled Vision-Language Prompts in the Wild

    Du, Xuefeng and Ghosh, Reshmi and Sim, Robert and Salem, Ahmed and Carvalho, Vitor and Lawton, Emily and Li, Yixuan and Stokes, Jack W. , year =. doi:10.48550/arXiv.2410.00296 , url =. 2410.00296 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2410.00296

  15. [15]

    doi:10.48550/arXiv.2504.12661 , url =

    Chen, Menglan and Pang, Xianghe and Dong, Jingjing and Wang, WenHao and Du, Yaxin and Chen, Siheng , year =. doi:10.48550/arXiv.2504.12661 , url =. 2504.12661 , archivePrefix =

    work page doi:10.48550/arxiv.2504.12661

  16. [16]

    2025 , eprint =

    ThinkGuard: Deliberative Slow Thinking Leads to Cautious Guardrails , author =. 2025 , eprint =

    work page 2025

  17. [17]

    2025 , eprint =

    Safety Through Reasoning: An Empirical Study of Reasoning Guardrail Models , author =. 2025 , eprint =

    work page 2025

  18. [18]

    2025 , eprint =

    GuardReasoner-VL: Safeguarding VLMs via Reinforced Reasoning , author =. 2025 , eprint =

    work page 2025

  19. [19]

    doi:10.48550/arXiv.2509.25896 , url =

    Huang, Guolei and Peng, Qinzhi and Xu, Gan and Huang, Yao and Lu, Yuxuan and Shen, Yongjun , year =. doi:10.48550/arXiv.2509.25896 , url =. 2509.25896 , archivePrefix =

    work page doi:10.48550/arxiv.2509.25896

  20. [20]

    2026 , eprint =

    Do MLLMs Really See It: Reinforcing Visual Attention in Multimodal LLMs , author =. 2026 , eprint =

    work page 2026

  21. [21]

    2026 , eprint =

    Towards Policy-Adaptive Image Guardrail: Benchmark and Method , author =. 2026 , eprint =

    work page 2026

  22. [22]

    2023 , eprint =

    Watch Your Language: Investigating Content Moderation with Large Language Models , author =. 2023 , eprint =

    work page 2023

  23. [23]

    2024 , eprint =

    LoRA-Guard: Parameter-Efficient Guardrail Adaptation for Content Moderation of Large Language Models , author =. 2024 , eprint =

    work page 2024

  24. [24]

    2024 , eprint =

    Trust-Oriented Adaptive Guardrails for Large Language Models , author =. 2024 , eprint =

    work page 2024

  25. [25]

    2024 , eprint =

    SLM-Mod: Small Language Models Surpass LLMs at Content Moderation , author =. 2024 , eprint =

    work page 2024

  26. [26]

    2024 , eprint =

    STAND-Guard: A Small Task-Adaptive Content Moderation Model , author =. 2024 , eprint =

    work page 2024

  27. [27]

    2025 , eprint =

    Poly-Guard: Massive Multi-Domain Safety Policy-Grounded Guardrail Dataset , author =. 2025 , eprint =

    work page 2025

  28. [28]

    2025 , eprint =

    MSR-Align: Policy-Grounded Multimodal Alignment for Safety-Aware Reasoning in Vision-Language Models , author =. 2025 , eprint =

    work page 2025

  29. [29]

    2025 , eprint =

    Customize Multi-modal RAI Guardrails with Precedent-based Predictions , author =. 2025 , eprint =

    work page 2025

  30. [30]

    2025 , eprint =

    DynaGuard: A Dynamic Guardian Model With User-Defined Policies , author =. 2025 , eprint =

    work page 2025

  31. [31]

    2025 , eprint =

    Taxonomy-Adaptive Moderation Model with Robust Guardrails for Large Language Models , author =. 2025 , eprint =

    work page 2025

  32. [32]

    2026 , eprint =

    MTMCS-Bench: Evaluating Contextual Safety of Multimodal Large Language Models in Multi-Turn Dialogues , author =. 2026 , eprint =

    work page 2026