acceptodds
Under review as a conference paper at ICLR 2027

Beyond Refusal: Building and Evaluating Diffusion Post-hoc Safety Intervention Models

Abstract

Existing safety mechanisms primarily rely on refusal or rewriting to block harmful outputs. However, in real-world scenarios where safe and harmful content are intertwined, discarding or rewriting the entire response is not only inefficient but also results in the loss of valuable information. To address this, we shift our focus to post-hoc safety interventions. The effectiveness of a post-hoc intervention model should be evaluated on its ability to accurately identify risks, precisely edit harmful spans, and preserve benign context, all while maintaining high inference efficiency. To excel across these dimensions, we introduce SafeLLaDA, a diffusion-based safety model. By leveraging the unique iterative refinement properties of diffusion models, SafeLLaDA precisely locates and modifies harmful spans post-generation while leaving the benign context unchanged. Experimental results demonstrate that SafeLLaDA significantly mitigates harm and achieves superior content fidelity compared to baseline guardrails and other traditional safety measures, establishing diffusion-based post-hoc intervention systems as a highly viable alternative to traditional refusal or rewrite strategies.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI Safety System, Diffusion Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_nQj4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies localized safety intervention for LLM responses containing both benign and harmful content. The authors propose SafeLLaDA, a diffusion-based model that detects unsafe spans and regenerates them while preserving the surrounding context. They also introduce OmniSafeBench to evaluate safety, fidelity, fluency, and efficiency.

Critical strengths

1. The paper identifies a useful practical setting where an otherwise helpful response contains only localized unsafe content. Looking beyond full refusal is a useful direction.

2. The key insight of diffusion-based localized editing is practically reasonable. Only regenerating selected spans is a natural fit for this post-hoc intervention.

Critical weaknesses

1. The local editing strategy seems to rely on an assumption that authors do not sufficiently validated: the preserved context remains valid after an unsafe span is replaced. SafeLLaDA edits only the detected span and keeps the rest of the response unchanged, but an autoregressive suffix may depend on the original content of that span. Even if the suffix is safe in isolation, it may no longer be a valid continuation after the replacement.

2. The paper does not clearly distinguish iterative editing from repeated safety checking over the full response, so I am not sure how final safety is ensured after an intervention. If detection and editing can repeat when unsafe content remains, the stopping condition and stability across multiple interventions should be clarified. If detection is performed only once, the paper should justify why the initially identified span is sufficient to determine all necessary edits.

Questions for authors

1. My understanding is that SafeLLaDA edits the identified unsafe span while keeping the rest of the response unchanged. I am concerned about the assumption that the originally safe context remains valid after the replacement. For an autoregressive response, the later suffix may heavily depend on the edited content, even if the suffix itself is safe. Have you tested cases where the edited span is important for the later context? A few examples would be helpful.

2. I would like to know the difference between iterative editing and repeated safety checking in SafeLLaDA. Like, after one localized edit, does the system recheck the response and intervene again, if unsafe content remains? If so, what is the stopping condition? I am also curious whether a single detection&editing is sufficient.

3. Can you show what localized intervention looks like in practice? How many unsafe spans are typically detected, and roughly how much of the response is edited? For multiple spans, are they edited jointly or separately?

4. The paper would benefit from separating unsafe-span localization from subsequent rewriting. The current end-to-end results make it hard to tell whether a failure comes from localization, span boundaries, or the replacement itself. Reporting these components separately would make the failure modes easier to understand.

Reviewer_QqncOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper aims to balance the helpfulness and safety of generated texts while avoiding over-refusal behaviors. The authors study post-hoc safety intervention, where a model identifies unsafe spans and edits within those spans. The resulting algorithm is called SafeLLaDA, a diffusion-based safety model that initializes iterative refinement. Experimental results show that SafeLLaDA reduces harmful content while retaining high content fidelity.

Critical strengths
  1. The motivation behind this paper is clear and well-presented.

  2. The proposed method is technically sound and easy to implement.

  3. The empirical results seem promising.

Critical weaknesses
  1. While the proposed method is simple and easy to implement, I feel it is incremental.

  2. Reading Table 3, I am wondering whether SafeLLaDA actually performs better than baseline algorithms with respect to the helpfulness and safety.

  3. This paper does not report important metrics such as precision or F1 score regarding unsafe span detection. I feel it is difficult for the readers to understand how the proposed method works.

  4. [Minor] There exists a benchmark called OmniSafe (https://github.com/PKU-Alignment/omnisafe). I recommend the authors change the name of proposed benchmark.

Finally, the supplementary README appears to violate the double-blind anonymity requirement. It contains direct links to a non-anonymous GitHub repository and GitHub Pages site under the username nameistzzhang, e.g., https://github.com/nameistzzhang/safe_dllm.git and https://nameistzzhang.github.io/safe_dllm/. This GitHub links to this Google Scholar profile (https://scholar.google.com/citations?hl=en&user=TSzGqfAAAAAJ&view_op=list_works&sortby=pubdate). I think this paper should be desk-rejected.

Questions for authors

See the Critical Weaknesses

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

This paper studies post-hoc safety interventions that refuse fully harmful outputs and rewrite only unsafe spans in partially harmful ones while preserving the rest. It introduces OmniSafeBench (~3,300 manually character-annotated examples) and trains SafeLLaDA on LLaDA-8B-Instruct. Experiments evaluate safety, content preservation, fluency, and latency, with ablations and jailbreak-robustness tests.

Critical strengths

The formulation and fine-grained evaluation are valuable: separating safe, fully harmful, and partially harmful responses, and measuring content preservation alongside safety, surfaces issues that refusal rates miss. The test set uses two annotators with expert adjudication—limited but an important basis for studying localized editing.

Editing-specific training improves performance. In Table 3, SafeLLaDA has the lowest average latency (0.41 s); Figure 6 shows latency rises to 1.34 s without editing training. Other ablations similarly support these design choices.

Critical weaknesses

The comparisons and ablations do not clearly separate the effects of training, the localized editing interface, and diffusion. The main results compare specially trained SafeLLaDA to zero-shot prompted AR and diffusion baselines that rewrite the entire response, and there is no baseline that uses “localization + local replacement” trained on the same data, making the evaluation unfair.

The reported gains are also insufficiently validated. On the minor-flaw subset (Table 3), SafeLLaDA raises ROUGE-L only from 0.7786 to 0.7841 versus Qwen-3.5-9B, but has higher residual harmfulness (8.37% vs 2.28%). Despite faster inference, it is unclear whether the trade-off preserves more usable content.

Questions for authors

See weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.