acceptodds
Under review as a conference paper at ICLR 2027

Moderating Covertly Sensitive Images Across Policies: Policy-Reconfigurable Benchmark and Zero-Shot Method

Abstract

Harmful image moderation is an important task in content safety, complicated by ambiguity in what constitutes harmful content and variation across moderation policies. Existing datasets and methods often provide limited support for these challenges. To fill this gap, we introduce the Covert Sensitive Image Dataset (CSID), comprising 2,510 images collected from diverse sources and spanning semantically complex violation themes, with labels configurable under different moderation policies. By formulating theme-specific objective questions, CSID enables the reuse of a single set of annotations across policies. We further propose VLGuard, a training-free, evidence-guided framework for image moderation. VLGuard decomposes each image into structured semantic units and retrieves relevant sensitive concepts to inform moderation decisions by a vision-language model (VLM). Evaluation against 12 image moderation methods from industry and academia demonstrates that VLGuard achieves an average F1 score of 74.8% on CSID, outperforming the strongest evaluated academic baseline and the evaluated closed-source VLM by 16.7 and 3.7, respectively.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Content Safety, Image Moderation, Vision-Language Models, Context Engineering, Benchmarking

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_KSNkOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a training-free prompting and retrieval pipeline that helps vision-language models moderate images under different safety policies. It also introduces a dataset with detailed content annotations that can be combined into policy-specific safety labels. The authors evaluate their pipeline against existing moderation methods on this benchmark and two external datasets, showing that many existing systems struggle with the proposed benchmark.

Critical strengths
  • A training-free method: The pipeline adapts existing VLMs without task-specific fine-tuning, unlike several competing approaches.
  • A new, albeit small, benchmark: Rich content annotations support evaluation under multiple moderation policies.
  • Competitive empirical results: Strong performance on the proposed benchmark and competitive results on two external datasets, with comparisons against a broad range of methods.
Critical weaknesses
  • Same backbone comparisons cover only the extremist ideology subset. I can't find an explanation as to why this particular subset was chosen instead of the entire dataset. Without direct Qwen results on politics, violence, USD, and UnsafeBench, it is difficult to determine how broadly the pipeline improves over its underlying model.
  • Component ablations cover only extremist imagery under the strict policy. They do not isolate the contribution of retrieval from structured image inspection and an additional model call
  • Annotation reporting is insufficiently detailed. The paper doesn't clarify whether the two annotators were authors or independent annotators, state their qualifications, or report agreement before consensus resolution.
  • The dataset is small and narrowly scoped, with images collected from existing benchmarks, AI generation, and web retrieval. Its main added value appears to lie in the policy-reconfigurable annotations rather than the image collection itself,

More minor comments:

  • Text within Figures 1 and 3 is a bit too small to read comfortably.limiting its significance as a new data resource.
  • The paper SafetyPairs merits discussion as complementary related work on fine-grained image safety.
Questions for authors
  • Does the pipeline improve over direct inference with the same Qwen backbone on the political and violence subsets, USD, and UnsafeBench? What motivated restricting the reported same-backbone comparisons to extremist imagery?
  • Which baselines support adapting to different moderation policies, and how were the target policies supplied to each? For fixed-policy systems, could performance differences reflect different definitions of harm rather than failures of image understanding. This could be clarified a bit. For example, flagging Nazi imagery even when it explicitly expresses opposition to Nazism?
  • Were the annotators authors or independent annotators? Please describe who they are generally and the number of images excluded due to unresolved disagreements.
Reviewer_zJcSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies image content moderation under different moderation policies. The authors introduce CSID, a dataset of 2,510 images covering political content, violence, and extremist ideology. Instead of assigning a fixed harmful or harmless label, the dataset uses several content questions and combines their answers to obtain labels under different policies. The paper also proposes VLGuard, a training-free method for image moderation. It first extracts semantic triplets from the image, retrieves related concepts from a harmful concept lexicon, and then provides the image, policy, and retrieved evidence to a VLM for the final decision. The method is evaluated on CSID and two public benchmarks, together with ablation studies and experiments using different VLM backbones.

Critical strengths

The policy-reconfigurable annotation design is useful. Separating image content annotations from the final harmful/harmless label makes it possible to reuse the same images under different moderation policies. This is a reasonable setting for content moderation, where the decision boundary can change across policies. The experimental results are also fairly strong. VLGuard obtains the best average F1 among the evaluated methods on CSID and also performs well on USD and UnsafeBench. The experiments with different VLM backbones show consistent improvements over direct inference. The ablation results also suggest that semantic retrieval is an important part of the method.

Critical weaknesses

My main concern is the construction of CSID. A large part of the political and violence subsets is selected from samples that existing moderation systems already fail on. This makes the dataset challenging, but it may also introduce selection bias. It is unclear how representative the reported performance is for a more natural image distribution. The annotation quality also needs more analysis. Two annotators independently label each image and resolve disagreements through discussion, but the paper does not report their initial agreement. Some questions related to stance and context may still involve subjective judgments. I also think the method analysis could be stronger. VLGuard combines VLM-based triplet extraction, retrieval, reranking, and an automatically generated concept lexicon. The current ablation shows that these components are useful, but it is only reported on the extremist ideology subset under Policy 1. It is unclear whether the same conclusions hold for the other themes and more context-dependent policies.

Questions for authors
  1. In Section 3.2, many images are selected because at least two existing moderation systems misclassified them. Could this selection process favor VLGuard over the baselines? How would the method perform on randomly sampled images?
  2. The annotations are completed by two annotators, with disagreements resolved through discussion. What was their initial agreement rate? How many images were removed because no consensus could be reached?
  3. CSID uses three manually defined policies for each theme. How were these policies chosen? Have you considered using policies from actual moderation platforms?
  4. The extremist ideology subset mainly focuses on Nazism. How well would VLGuard generalize to other extremist ideologies? Would a separate lexicon need to be constructed for each one?
  5. The harmful concept lexicon is generated using GPT-5.5. How sensitive is the method to the quality and size of this lexicon? Have you tested lexicons generated by other models?
  6. How much of the improvement comes from providing additional concept information rather than the triplet extraction and matching pipeline? It would be helpful to compare against a simpler baseline that directly provides relevant concept descriptions to the VLM.
  7. Table 5 only reports ablation results on the extremist ideology subset under Policy 1. Do the same components provide similar improvements under Policies 2 and 3? This seems important since the more lenient policies require more context-dependent judgments.
  8. In Table 2, MIND is evaluated on only half of each dataset, while the other methods use the full dataset. Could the authors provide results on the same test samples for a more direct comparison?
Reviewer_iJtkOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the authors provide a solution to image content moderation with configurable policy problem. From industry perspective, this problem has practical merits. To solve this problem, the authors created a dataset (CSID), with 2510 images covering Political, Violence, and Extremist. The dataset is annotated with 6-7 boolean questions and can be configured to support 3 policies.

On top of the dataset, the authors proposed a training-free, evidence-guided framework for image moderation. They demonstrated superior performance of this approach on both CSID and public benchmarks. They also conducted experiments with various backbone VLM and ablation studies.

Critical strengths
  1. The CSID dataset, once open sourced, can benefit practitioners working in image content moderation.

  2. The proposed VLGuard outperforms multiple academic models and industry models.

Critical weaknesses
  1. [Major] although claimed to be "policy reconfigurable", the reconfigurability is very limited -- for example, the CSID dataset can support a max of 3 different policies. Throughout this paper, the authors never acknowledge this limitation, which makes the "policy reconfigurable" an overstatement.

  2. [Major] The authors does not provide any intuition or rationale on CSID dataset creation process. I have a list of questions: 1. is this the first dataset that supports configurable policy? 2. If not, what's the pros and cons of this creation process vs the previous process? Why choose this approach which limits "policy reconfigurability"?

  3. [Major] In the experiment section, the authors only provide quantitative results without analyzing a few detailed qualitative examples.

  4. [Relative Minor] In the Literature survey, the authors do have touch content moderation dataset that supports configurable policy.

Questions for authors

Please refer to the above Weakness section.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.