acceptodds
Under review as a conference paper at ICLR 2027

"Benign-Shifted Compliance": A Mechanistic Insight into Safety Behaviors of Diffusion Language Models under Adversarial Prefixes

Abstract

Diffusion Language Models (DLMs) have emerged as promising alternatives to autoregressive models (ARMs), yet their safety mechanisms and robustness under jailbreak attacks remain poorly understood. Based on a systematic evaluation of DLMs and ARMs under a representative compliant-prefix attack, we identify a distinctive safety behavior in aligned DLMs, termed Benign-Shifted Compliance (BSC). Under BSC, the model preserves the surface form of prefix-imposed compliance while redirecting the harmful intent toward benign, policy-aligned content. Mechanistic analysis reveals that safety representations and refusal trajectories diverge explicitly during the earliest denoising steps. Within this critical window, we identify a sparse set of attention heads mediating BSC under adversarial attacks; causally disrupting these heads via targeted activation steering effectively bypasses the benign redirection, forcing aligned DLMs into harmful compliance. By elucidating the early semantic shaping and sparse head-level mechanisms supporting BSC-centric DLM robustness, our work fills a critical gap in understanding diffusion-based safety, for it not only exposes an internal DLM-specific vulnerability but also provides foundational insights to guide future alignment and jailbreak defenses for language models.

Primary area:
interpretability and explainable AI
Keywords:
Mechanistic Interpretability, Diffusion Language Models, Language Model Safety

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_VSSZOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper compares compliant-prefix attacks on open-source diffusion language models and autoregressive language models. The authors claim to identify a safety phenomenon (BSC) in DLMs where the model redirects the harmful intent toward benign content in early denoising steps. Based on this finding on the output side, the authors study the internal mechanism and identify a sparse set of attention heads causally related to the BSC. The authors further show that causally disrupting these heads can bypass the benign redirection and increase attack success rates on DLLMs.

Critical strengths
  1. This paper identifies the attention heads that are causally associated with DLMs’ redirection of harmful intent towards safety content, which is a new finding associated with the safety of DLMs
  2. The evaluation of robustness to prefix-based attacks covers different model families and sizes of ARMs and DLMs, making the empirical claim that DLMs are more robust to prefix-based attacks solid.
Critical weaknesses
  1. The paper is not very explicit about its scope and contribution, and risks overclaiming. For scope, it focuses on prefix-based jailbreak attacks instead of the full landscape of jailbreak attacks. This should be stated explicitly as a limitation. For contribution, it is established by previous works (such as DiffuGuard) that the early decoding steps of DLMs are important for safety behavior, and the proposed steering method is essentially transferring ARM steering to DLMs by identifying specific heads, which is conducted across all steps, making it less DLM specific. Additionally, this paper did not identify the exact mechanism that produces this difference in DLM and ARM’s different behavior to prefix-based attacks. The authors find that DLMs are more robust through evaluation, and identify the attention heads that are causally associated with this behavior. This does not substantiate the stronger claim of DLM specificity.
  2. The evaluation of prefix attacks on ARMs and DLMs covers a wide range of models. However, the steering experiment is mainly done on LLaDA-1.5 and LLaDA2.0-mini, with the improvement on LLaDA-1.5 very modest.
Questions for authors
  1. Could you clarify, or perhaps downscope the claims, what the novel mechanistic insight is here in this paper, beyond identifying some heads that are causally associated with this safety behavior of DLMs? Since the paper compares DLMs with ARMs, it remains unclear what unique characteristic of DLMs produces the BSC behavior. Perhaps a matched intervention analysis on ARM would be valuable to study the difference between these two model families.
  2. Can you expand the steering experiments to more model families?
7 comments
AuthorsOpenReview ↗

Response to Reviewer VSSZ (Part 1)

We appreciate the reviewer's thoughtful feedback. Below, we address each of the main concerns raised in the "Critical Weaknesses" and "Questions For Authors" sections.


The scope and DLM-specific contributions of our work. (Weakness 1 & Question 1)

Regarding the scope of our work, our study explicitly focuses on DLM safety behaviors under adversarial prefixes, as clearly stated in both the title and abstract. Compliant-prefix attacks constitute our primary evaluation setting, while Appendix E further examines the generalization of BSC to the structurally different Context Nesting attack. Therefore, the scope of our study is already clearly defined in the submission.

Regarding the DLM-specific contribution, we do not claim novelty in the head-wise probing or activation steering techniques themselves, which are standard interpretability tools applicable to transformer-based models. Our contribution lies in discovering BSC, a distinctive safety behavior observed in the aligned DLMs we evaluate, and investigating how this behavior is supported by the internal dynamics of diffusion-based generation. In particular, our mechanistic analysis goes beyond identifying a sparse set of attention heads. We further investigate how BSC forms and is maintained through the DLM-specific iterative denoising process, using fine-grained interventions across denoising steps and generation blocks (Appendix D.3 and Section 5.2.2). These analyses provide DLM-specific insights into how BSC emerges, persists, and can be disrupted or recovered across denoising steps and generation blocks. Thus, although our default intervention spans all denoising steps and blocks to prevent subsequent safety recovery, our analysis explicitly examines the temporal and block-wise dynamics underlying this intervention, rather than simply transferring a static ARM steering procedure to DLMs.

Regarding the exact origin of the behavioral difference between ARMs and DLMs, head-level probing and steering identify which internal components mediate a behavior within a model, but do not by themselves isolate the complete architectural and post-training factors responsible for differences across generation paradigms. This is a broader limitation of existing interpretability tools rather than a drawback specific to our work. Nevertheless, we still investigate this divergence from DLM-specific perspectives. As stated in Section 3.2 (~line 252), the main purpose of our subsequent analysis is to explain how DLM's iterative denoising process supports this distinctive behavior observed in aligned DLMs but not in the evaluated ARMs.

To further examine whether safety-oriented post-training alone can reproduce this behavior, we additionally conduct standard DPO post-training on Ling-mini-2.0, an ARM with comparable model scale and general capability to LLaDA2.0-mini.

ModelEarly Harmful Continuation (EHC) (%)Early Benign Shift (EBS) (%)Early Neutral (%)
LLaDA2.0-mini7.5048.0044.50
Ling-mini-2.0 before DPO66.0017.5016.50
Ling-mini-2.0 after DPO69.0020.5010.50

After DPO, its EHC rate remains high (66.00% to 69.00%), substantially differing from LLaDA2.0-mini. Together with our temporal intervention results, these findings suggest that the behavioral divergence between ARMs and DLMs likely stems from the DLM-specific iterative denoising process, rather than safety-oriented preference optimization alone. Regarding the suggested matched intervention on ARMs, such a comparison is not feasible because ARMs do not exhibit a sufficiently prevalent BSC tendency to identify a corresponding BSC-to-HC steering direction.

Overall, our contribution is the discovery and characterization of BSC in aligned DLMs, together with its mechanistic analysis through sparse attention heads and the distinctive temporal dynamics of iterative denoising.


AuthorsOpenReview ↗

Response to Reviewer VSSZ (Part 2)

Applicability of the steering intervention across DLM families. (Weakness 2 & Question 2)

Our steering intervention targets the key attention heads mediating BSC, with a steering direction from BSC toward harmful compliance (HC). Therefore, it can only be meaningfully applied to models that exhibit identifiable BSC behavior. Although our evaluation covers a wide range of DLMs, only LLaDA-1.5 and LLaDA2.0-mini exhibit clear BSC tendencies under compliant-prefix attacks (Figure 1b). Moreover, as shown in Figure 1b, BSC is less dominant in LLaDA-1.5, which explains its more modest ASR improvement: the intervention specifically disrupts the BSC behavior, and thus has a more limited effect when this behavior is less prevalent.

To further address this concern, we additionally evaluate TESS 2, a DLM from another model family. TESS 2 remains substantially more robust than the evaluated ARMs under compliant-prefix attacks, achieving 28.00% ASR on HarmBench and 22.33% on HEx-PHI. However, its early semantic trajectories are dominated by EHC (82.00%), with only 2.00% EBS, resembling instruction-tuned DLMs more than explicitly aligned DLMs. Since TESS 2 rarely exhibits BSC behavior, there is also no sufficiently clear BSC-related direction to identify and use for targeted steering. Thus, the applicability of our steering intervention depends on whether a model exhibits BSC, rather than merely whether it belongs to the DLM family.


Thanks for the rebuttal. I understand this work focuses on prefix-based attacks, which can be seen as one kind of jailbreak. In my view this scope is a bit limited as it does not fully capture the broader landscape of jailbreak attacks. While the rebuttal addressed some of the concerns, I am still not convinced on the following points.

  1. Prior works already investigated safety recovery during denoising steps such as "Step-Wise Refusal Dynamics in Autoregressive and Diffusion Language Models" (arxiv2602.02600) and the context nesting paper, which are somewhat related to BSC.
  2. The authors frame the head-specific steering intervention as a major contribution (L111), but as the authors said during rebuttal, it can only be meaningfully applied to models that exhibit identifiable BSC behavior, and that only LLaDA-1.5 and LLaDA2.0-mini exhibit clear BSC tendencies, with improvement on LLaDA-1.5 very modest. Thus, the empirical evidence seems limited.

Could the authors further address these points? Given the rebuttal period timeframe, I think additional experiments are not necessary and some explanation would suffice.

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

This paper studies aligned diffusion language models under compliant-prefix jailbreaks and identifies Benign-Shifted Compliance (BSC), where the model appears compliant but redirects harmful intent to safe content. The authors validate BSC on HarmBench and HEx-PHI, analyze its emergence in early denoising steps, identify sparse attention heads associated with the behavior, and use targeted steering to show these heads causally affect safety outcomes. Additional transfer and Context Nesting experiments provide further support for the mechanism.

Critical strengths
  • The paper identifies BSC as a distinctive safety behavior in aligned DLMs, going beyond standard refusal and offering a useful lens for understanding how diffusion-style generation may support safety under adversarial prefixes.

  • The paper connects early denoising dynamics, sparse head-level signals, and targeted activation steering into a coherent causal story. This makes the contribution more substantial than a black-box robustness observation.

Critical weaknesses
  • The paper suggests that BSC mechanisms may inform future defenses, but it does not evaluate whether encouraging BSC would increase refusals or reduce helpfulness on benign or borderline safety requests. This matters, because a safety mechanism that mainly works by redirecting content may come with utility costs that are not captured by ASR or MMLU.

  • There is limited evidence for attack-level generalization of the proposed mechanism. Although the paper adds Context Nesting as an additional jailbreak setting, the evidence still mainly shows that a similar mechanism can be identified under another attack, rather than that the same heads or steering directions transfer robustly across attack families. This limits confidence that BSC reflects a general DLM safety mechanism rather than a mechanism tied to a narrow class of structured compliance-style attacks.

Questions for authors
  • In the Context Nesting experiments, are the key heads and steering directions re-identified under Context Nesting, or directly transferred from the compliant-prefix setting without re-estimation?

  • Are the early-denoising BSC patterns stable under non-greedy decoding or different block/step sizes?

  • How are true BSC responses distinguished from other safe outputs such as explicit refusals or generic safety explanations in the head-probing and steering analyses?

3 comments
AuthorsOpenReview ↗

Response to Reviewer NFWB (Part 1)

We appreciate the reviewer's thoughtful feedback. Below, we address each of the main concerns raised in the "Critical Weaknesses" and "Questions For Authors" sections.


Whether encouraging BSC increases over-refusals. (Weakness 1)

We did not discuss whether encouraging BSC would increase over-refusals or reduce helpfulness, because BSC is a soft, naturally occurring redirection behavior under prompt-based attacks, rather than either a hard-refusal policy introduced through safety post-training or a refusal-oriented behavior induced through activation-based intervention. BSC therefore neither imposes a hard-refusal policy nor explicitly steers the model toward refusal, and thus does not inherently lead to over-refusal.

Nevertheless, to fully address this concern, we conduct additional compliant-prefix attack experiments on LLaDA2.0-mini using XSTest, which contains 250 prompts labeled Safe and 200 labeled Unsafe. Following the full-response classification protocol used in our Context Nesting evaluation, we classify each response into three categories: Explicit Refusal, Benign-Shifted Compliance (BSC), and Harmful Compliance (HC). For benign requests, we use "Benign Compliance" because there is no harmful intent to be shifted.

XSTest Prompt LabelHarmful Compliance (%)Benign(-Shifted) Compliance (%)Refusal (%)
Safe0.00100.000.00
Unsafe27.0062.5010.50

The results directly confirm that BSC does not induce over-refusal. For safe prompts, the model produces 0% refusals and 100% benign compliance, even under an adversarial prefix. Meanwhile, for unsafe prompts, BSC remains the dominant safe behavior. These findings demonstrate that BSC redirects harmful requests toward benign content without compromising the model's willingness to answer benign requests.


Generalization of the proposed BSC mechanism across attack settings. (Weakness 2 & Question 1)

Under Context Nesting, we re-identify the key heads and re-estimate the corresponding steering directions, rather than directly transferring them from the compliant-prefix setting. This is a deliberate choice for two reasons. First, the identified BSC-related components are highly sparse, with only five key heads selected in our primary experiments. Second, although the model exhibits the same high-level BSC behavior under both attacks, their prompt structures and resulting response organizations differ substantially. Consequently, the same high-level behavior does not necessarily rely on identical attention heads across different attack formulations. Prior mechanistic interpretability studies have shown that different prompting methods for the same task can engage largely distinct, partially overlapping sets of attention heads [1], and that different prompt templates can induce systematically different circuits even for a fixed task [2]. Thus, re-identifying the key heads under Context Nesting is a reasonable approach to localizing the components mediating BSC in this different attack setting.

Importantly, to further investigate the transferability of the originally identified components, we additionally evaluate steering using the key heads identified under compliant-prefix attacks, without re-identifying the heads. The results are as follows:

Context Nesting SettingASR (%)
w/o steering34.50
w/ same-head steering40.50 (+6.00)
w/ re-identified head steering47.00 (+12.50)

Notably, directly reusing the original key heads already increases ASR by 6.00 percentage points, demonstrating that these heads retain a non-trivial causal effect on BSC under Context Nesting. Re-identifying the key heads further increases ASR by 12.50 percentage points, yielding a stronger intervention effect under the new attack setting.

The purpose of introducing Context Nesting in Appendix E is to demonstrate that the BSC behavior and its underlying head-level causal mechanism generalize beyond direct compliant-prefix injection to a structurally different jailbreak setting, rather than to establish that an identical set of heads and steering directions must transfer unchanged across attack families. The additional transfer experiment further demonstrates that the originally identified heads retain a degree of cross-attack transferability, while attack-specific localization enables more effective disruption. Together, these results support the generalization of both BSC and its associated causal mechanism across different attack formulations.

[1] Do Different Prompting Methods Yield a Common Task Representation in Language Models? NeurIPS 2025.

[2] Finding Highly Interpretable Prompt-Specific Circuits in Language Models. NeurIPS 2026.


AuthorsOpenReview ↗

Response to Reviewer NFWB (Part 2)

BSC patterns under different decoding settings. (Question 2)

BSC is a model-behavior-level statistical pattern induced by the aligned DLM under adversarial prompts, thus theoretically would not be determined by a particular decoding configuration. Therefore, our main mechanistic analysis uses deterministic decoding to obtain reproducible trajectories and stable component localization.

Nevertheless, to fully address this concern experimentally, we conduct additional experiments under non-zero temperature and different block and denoising-step sizes. The two tables below report the ASRs before and under the prefix attack, together with the early semantic trajectories used to characterize BSC-related benign redirection (as introduced in Section 3.2).

Generation SettingASR Before Attack (%)ASR Under Attack (%)
Block size = Step size = 32, temperature = 0.04.0034.17
Block size = Step size = 16, temperature = 0.80.5914.95
Block size = Step size = 32, temperature = 0.80.1019.69
Block size = Step size = 64, temperature = 0.80.5022.80
Generation SettingEarly Harmful Continuation (EHC) (%)Early Benign Shift (EBS) (%)Early Neutral (%)
Block size = Step size = 32, temperature = 0.07.5048.0044.50
Block size = Step size = 16, temperature = 0.825.5064.5010.00
Block size = Step size = 32, temperature = 0.828.5059.0012.50
Block size = Step size = 64, temperature = 0.833.5052.0014.50

Across all evaluated decoding configurations, LLaDA2.0-mini maintains substantial robustness under the prefix attack, and Early Benign Shift remains a prominent trajectory, accounting for 52.00%–64.50% of the responses under non-zero-temperature decoding. These results show that the BSC pattern is not specific to greedy decoding or certain block/step size, but remains consistently observable across different temperatures and block/step sizes.


BSC and explicit refusals under mechanistic analyses. (Question 3)

As clarified in Section 5.1 and detailed in Appendix C.3, although we use the Harmfulness Score (HS) as the supervision label, DLM-generated responses labeled with "HS<=2" are overwhelmingly composed of genuine BSC responses (over 96%), rather than explicit refusals. Therefore, the head-probing labels can be regarded as a binary BSC/HC categorization, and the resulting steering direction points from BSC toward HC. Accordingly, our head-probing and steering analyses capture the BSC mechanism rather than a generic harmful/safe distinction.


Thank you to the author for the response. I suggest that the author incorporate these additional results and the discussion into the revised manuscript, as this would make the narrative more complete. So I maintain my positive score.

Reviewer_mp1GOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies safety behavior in aligned diffusion language models under a simple compliant-prefix (prefill) jailbreak, and compares them with autoregressive models in the same attack. The setting is a reasonable way to study DLM safety. Diffusion models also have an explicit denoising-time axis that autoregressive models do not, so asking when a safety-related signal becomes readable along that axis is a legitimate DLM-specific question.

The authors name the main phenomenon benign-shifted compliance: the model keeps a compliant surface (for example “Sure…”) while the content stays safe. The mechanism experiments mostly follow another line. They score replies as harmful or safe, find heads that track that distinction, and steer those heads so a reply the prefix had blocked becomes actually harmful. That line is more coherent: a white-box study of flipping a prefix-blocked generation, with no targeted defense. The steering result is easy to read—it shows an internal knob. The paper gives BSC more weight in the narrative than this work supports, and it does not analyze the prefix → harmfulness-based heads → steer-into-compliance route as the main claim.

The results support a narrower reading. Under the chosen prefix template, some aligned open DLMs produce surface-compliant but semantically safe replies; safety-related signals can already be decoded early in denoising; and a small set of attention heads can change attack success when steered. The evaluation uses automatic judging and a limited human check. The paper does not show that BSC is cleanly distinct from ordinary refusal or safe completion, that early readability means the early window decides the outcome, or that diffusion itself explains the robustness gap with autoregressive models.

Critical strengths
  1. A simple compliant-prefix / prefill attack is a concrete, experimentally tractable setting for DLM safety, rather than a general claim that these models need to be studied.
  2. Diffusion models have a denoising-time axis that autoregressive models lack. Asking when a safety-related signal becomes readable on that axis is a real DLM-specific question.
  3. After heads that track final harm versus safe are located, steering them can turn a blocked reply into actual harmful compliance. That is an interpretable internal knob.
Critical weaknesses
  1. BSC is loosely coupled to the experiments that follow. The authors define it as keeping a compliant surface while redirecting the content to safety. Head probing and steering then use a harmfulness-style binary, and the intervention result is that a prefix-blocked generation becomes truly harmful. That second line does not need BSC as a named object, and it is the more coherent piece of work. The paper still builds the story around BSC and does not develop this route as the main result. The identified heads may be tracking ordinary harm, refusal, or safe completion.
  2. Accepting the time axis does not mean the paper needed a new method to use it. Head ranking and early linear probes are standard autoregressive tools; denoising steps here play the role that early tokens play there. The DLM-specific test would have been whether later denoising can still revise an early trajectory, and what happens if intervention is restricted to early versus late steps. Those comparisons are not the main result. As written, the internals look like a shallow transplant, so the distinctive scientific object is hard to see. Early decodability is not evidence that the early window causes the final behavior.
  3. If BSC is not unique to diffusion models, what remains should be stated plainly: a case study of prefix, probing, and steering on some recent aligned open DLMs. Successful steering shows high-leverage control points. It does not show that a sparse set of heads mediates a new, paradigm-level safety mechanism, and the cross-family comparison does not isolate diffusion as the source of the observed robustness differences.
Questions for authors
  1. On the reported correspondence between HS≤2 and BSC: how did you separate genuine BSC (compliant surface, safe content) from ordinary refusal and generic safe completion? If a substantial share of the HS≤2 set is just refusal or safe completion, I will keep reading the later heads as harm/safe trackers rather than as a BSC mechanism. If you already have a breakdown or a direct-BSC head ranking, please point to it—that would change this part of the review.
  2. Is any probe or steering result in the paper restricted to early denoising steps, with a matched late-step (or mid-step) control? Relatedly, do you have evidence that a later step can still revise an early trajectory? If that comparison is already in the paper and I missed it, the “time axis is unused as a causal object” concern would no longer hold. I am asking what the current draft contains, not for a new experiment.
  3. Do you claim that BSC, or the heads you steer, are specific to the diffusion paradigm, or only that they appear in the aligned open DLMs you tested? If the intended claim is the latter, I will read the paper as a case study of prefix + probing + steering, not as a paradigm-level safety mechanism. If you have a tighter ARM–DLM comparison that holds post-training alignment fixed, please point to the result that isolates the generator.
3 comments
AuthorsOpenReview ↗

Response to Reviewer mp1G (Part 1)

We appreciate the reviewer's thoughtful feedback. Below, we address each of the main concerns raised in the "Critical Weaknesses" and "Questions For Authors" sections.


The relationship between BSC and the mechanistic experiments. (Weakness 1 & Question 1)

Section 5, including both head identification and the subsequent interventions on the identified heads, is entirely centered on the BSC behavior. Specifically, we first identify attention heads whose activations distinguish BSC from harmful compliance (HC), and then intervene on these same heads by steering their representations toward HC. The resulting disruption of benign redirection directly links the identified heads to the formation of BSC. Thus, BSC is tightly coupled with the mechanistic experiments, rather than being a loosely connected behavioral observation.

Regarding the correspondence between "HS<=2" and genuine BSC, Appendix C.3 has directly addressed the concern that this category might contain a substantial proportion of explicit refusals. Although we use the harmfulness-style binary HS as the supervision label, we further classify all responses with each label under compliant-prefix attacks according to the behavioral definition of BSC in Section 3.2. The results show that genuine BSC accounts for 96.18% and 97.74% of the HS<=2 responses on HarmBench and HEx-PHI, respectively, while explicit refusals account for only 3.82% and 2.26%. This confirms that the HS-based labels predominantly capture genuine BSC, rather than ordinary refusal, and thus enable the identified heads and steering directions to specifically characterize the distinction between HC and BSC under compliant-prefix attacks, rather than generic harmful/safe or refusal/compliance differences. Therefore, both the head identification and the subsequent steering experiments are directly designed to investigate and intervene in BSC, rather than constituting a generic harm-related mechanistic analysis loosely connected to BSC.

Although we believe the analyses above already clarify the BSC-specific design of our steering setup, we added a control to further address the reviewer's concern that the identified heads may simply act as generic harm/safe trackers. We further tested a harmful-vs-refusal steering control. It did not reproduce the ASR increase or BSC reduction observed with our BSC-to-HC intervention. Together with the behavioral validation above, this contrast supports distinguishing our results from generic harm/refusal steering.

Steering directionASR (%)BSC (%)
Baseline (w/o steering)34.1758.50
Refusal-to-Harmful16.5062.00
BSC-to-HC (ours)66.5026.00

AuthorsOpenReview ↗

Response to Reviewer mp1G (Part 2)

The DLM-specific contribution of our work, and the denoising-time-wise steering experiments. (Weakness 2 & Question 2)

Our DLM-specific contribution lies in discovering a previously unreported safety behavior in aligned DLMs, BSC, and investigating its internal formation through the distinctive iterative denoising process, rather than proposing new probing or steering algorithms. Although head-wise probing and activation steering are established interpretability tools that can be applied to DLMs given their shared Transformer-based backbone with ARMs, their application in our work serves a distinct scientific purpose: to uncover how aligned DLMs form BSC and what internal mechanisms support this behavior. Specifically, we identify the attention heads associated with BSC and conduct targeted interventions to examine their role in benign redirection. Our analysis further investigates how this behavior forms and evolves across DLM-specific denoising steps and generation blocks. Therefore, the contribution goes beyond a direct transplantation of existing ARM interpretability tools, encompassing both a distinctive safety behavior discovery and its mechanistic explanation within the DLM-specific generation process.

Furthermore, the current submission already includes the temporal intervention comparisons requested by the reviewer, rather than leaving the denoising time axis unexplored as a causal factor. Appendix D.3 varies the starting denoising step of steering within each generation block, demonstrating that the intervention effect is concentrated in the earliest denoising updates and substantially diminishes when steering is delayed. Section 5.2.2 further restricts steering to the first generation block and examines subsequent trajectory recovery, showing that later generation blocks can partially redirect initially disrupted trajectories toward benign content. These two experiments directly investigate the temporal dependence of intervention effectiveness and whether subsequent denoising can still revise an initially disrupted trajectory, respectively.

Finally, we do not infer causality from early decodability alone, nor do we treat DLM denoising steps as simply playing the role of early tokens in ARMs. Instead, we regard iterative denoising as a distinctive structural property of DLMs: Section 4 investigates how safety representations and response trajectories evolve throughout this process, while the controlled steering experiments further examine how interventions at different denoising steps and generation blocks affect the resulting safety behaviors. Together, these analyses investigate how the distinctive denoising dynamics contribute to the different safety behaviors observed between DLMs and ARMs.


AuthorsOpenReview ↗

Response to Reviewer mp1G (Part 3)

The scope of our work and a tighter ARM-DLM comparison. (Weakness 3 & Question 3)

BSC is a diffusion-specific safety behavior unique to diffusion models and appear in the aligned open DLMs we tested. We would not overclaim that the BSC behavior would generalize to all aligned DLMs, while it is also unfair to simply narrow down our contributions and discovery as "a case study of prefix + probing + steering".

Current interpretability tools cannot causally and accurately isolate the exact structural or alignment differences responsible for the behavioral divergence between ARMs and DLMs. This is a broader limitation of existing interpretability methods rather than a drawback specific to our work. Nevertheless, we investigate this behavioral divergence from DLM-specific perspectives. Our fine-grained steering experiments reveal how DLMs form BSC throughout their iterative denoising process. And as stated in Section 3.2 (~line 252), the main purpose of our subsequent analysis is to explain how DLM's iterative denoising process supports this distinctive behavior observed in aligned DLMs but not in the evaluated ARMs. Our work therefore goes beyond applying existing interpretability tools to DLMs, providing both a distinctive safety behavior discovery and its mechanistic analysis from multiple DLM-specific perspectives.

To further address the potential confounding effect of post-training alignment in the ARM-DLM comparison, we additionally conduct safety-oriented DPO post-training on Ling-mini-2.0, an ARM with a comparable model scale and general capability to LLaDA2.0-mini. The DPO component of LLaDA2.0-mini adapts preference optimization to the diffusion-based generation formulation, whereas CAP is introduced to improve inference efficiency rather than serving as an additional safety-alignment objective. Accordingly, we apply standard DPO to Ling-mini-2.0 using LLaMA-Factory and the harmlessness subset of HH-RLHF, a widely adopted preference dataset for safety-oriented post-training. We then evaluate the model before and after DPO using the same compliant-prefix attack and early semantic trajectory analysis.

ModelEarly Harmful Continuation (EHC) (%)Early Benign Shift (EBS) (%)Early Neutral (%)
LLaDA2.0-mini7.5048.0044.50
Ling-mini-2.0 before DPO66.0017.5016.50
Ling-mini-2.0 after DPO69.0020.5010.50

After DPO, the EBS rate of Ling-mini-2.0 increases only slightly, from 17.50% to 20.50%, while its EHC rate also increases from 66.00% to 69.00%. Its resulting trajectory distribution remains substantially different from that of LLaDA2.0-mini, which exhibits 48.00% EBS and only 7.50% EHC. This demonstrates that applying a standard DPO-style safety objective to an ARM is insufficient to reproduce the strong BSC tendency observed in aligned DLMs.

Since the exact post-training data and hyperparameter configurations of LLaDA2.0-mini are not publicly available, this experiment is not intended as an exact reproduction of its complete alignment pipeline. Nevertheless, it directly tests whether a similar preference-based safety objective alone can reproduce the observed BSC behavior, and the results show that it cannot. This behavioral divergence, together with our temporal intervention results, points to DLM's distinctive iterative denoising structure as an important factor underlying BSC.


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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.