acceptodds
Under review as a conference paper at ICLR 2027

Backdooring Masked Diffusion Language Models

Abstract

Masked diffusion language models (MDLMs) are emerging as a compelling new paradigm for text generation, but their training-time security remains largely unexplored. MDLMs are trained with discrete-state corruption and generate text through iterative denoising. In this work, we present a systematic study of training-time backdoor attacks on MDLMs. We propose SHADOWMASK, a backdoor attack that modifies the forward corruption process on poisoned examples by replacing the standard all-mask terminal distribution with a trigger–mask mixture prior. Clean examples use ordinary mask-based training. We further provide a principled mathematical formulation, deriving a weighted denoising cross-entropy objective for this construction. Deployment uses the unmodified sampling procedure provided with each model. Evaluations on DiT-based MDLMs and four large diffusion language model families across WikiText-103, OpenWebText, and Alpaca show that SHADOWMASK achieves near-100% attack success in many settings, substantially outperforms the evaluated data-poisoning baseline in ASR, and largely preserves clean utility. It remains effective under full-model training and parameter-efficient fine-tuning, and persists under the evaluated defenses.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Backdoor Attacks, 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_fAqHOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SHADOWMASK, a training-time backdoor attack on masked diffusion language models. On poisoned examples, it replaces the standard all-mask corruption prior with a trigger–mask mixture and derives a corresponding weighted denoising cross-entropy objective. The attack assumes control over the training pipeline while retaining the model’s original inference procedure.

Critical strengths
  1. The method explicitly modifies the categorical corruption process rather than only inserting trigger–target examples.

  2. The evaluation includes pretraining, selective fine-tuning, and parameter-efficient fine-tuning, with experiments across multiple datasets and model families.

Critical weaknesses
  1. Modifying the forward diffusion process and its terminal distribution to associate a trigger with a target, followed by deriving a compatible training objective, is already central to prior diffusion backdoor work such as BadDiffusion. SHADOWMASK adapts this principle to discrete masking, but the resulting objective remains a weighted cross-entropy loss. In the main small-model setting, ρ=1\rho=1 effectively replaces mask corruption with trigger-token corruption on poisoned examples. The paper does not sufficiently establish what new capability or insight this adaptation provides beyond the existing principle.

  2. Section 4 grants the attacker control over training data, corruption, and optimization, and assumes that the victim adopts the resulting checkpoint after checking clean utility. This is a legitimate supply-chain threat model, but it does not establish an attack available to an external API user or an adversary controlling only a small amount of public data. Under such broad training control, the central question is whether the proposed method provides a substantial advantage over strong alternative backdoor-training methods while surviving meaningful downstream checks. The current evidence only partially answers this question. Moreover, the reported ASR measures target-text generation, not successful execution of a harmful action or compromise of a deployed system.

  3. The derivation assumes a trigger state outside the clean prediction support and introduces a modified reverse kernel. In practice, triggers are existing vocabulary tokens, and generation uses the original model-specific sampler rather than the derived kernel. Consequently, the derivation establishes an objective for the analytical construction, but does not directly explain why its practical realization reliably transfers to ordinary sampling. This limits the explanatory value of the theoretical contribution; it is not, by itself, evidence of a mathematical error.

  4. The matched-budget small-model data-poisoning baseline is useful, but the large-model results in Tables 1, 9, and 11 primarily compare clean and backdoored models. They therefore establish feasibility rather than comparative advantage. There is also no direct comparison with the acknowledged concurrent work BadDLM. Furthermore, the ablations focus on target length and sampling steps, rather than the central trigger–mask mixture or controls matching trigger exposure.

Questions for authors
  1. What specific capability or insight does the proposed discrete corruption construction provide beyond adapting the forward-process manipulation used in prior diffusion backdoor methods?

  2. Can you compare SHADOWMASK with standard poisoning and a representative corruption-aware DLM attack on the larger models, using matched training budgets, poisoned examples, and clean-utility constraints?

  3. Which parts of the theoretical argument remain applicable when the trigger belongs to the ordinary vocabulary and deployment uses the original sampler? Can controlled experiments clarify the mechanism connecting these two settings?

  4. In the intended supply-chain scenario, what downstream validation or adaptation is the malicious checkpoint expected to survive, and what advantage does SHADOWMASK retain over simpler backdoor-training alternatives under those conditions?

2 comments

We thank Reviewer fAqH for the critical questions to help us improve the paper.


W1 & Q1. Novelty compared to BadDiffusion. Why is it not a straightforward application of BadDiffusion?

BadDiffusion and SHADOWMASK share the high-level idea of modifying diffusion processes to implant backdoors. However, SHADOWMASK is not a straightforward application of BadDiffusion because the two methods operate on fundamentally different diffusion mechanisms. BadDiffusion targets continuous Gaussian diffusion models and introduces trigger-dependent distribution shifts into their forward and reverse processes. In contrast, MDLMs employ discrete, mask-based corruption, for which BadDiffusion's Gaussian formulation is not directly applicable. SHADOWMASK addresses this difference by introducing a trigger-specific corruption state into the discrete diffusion process, reallocating corruption probability mass from the standard mask state to the trigger state without increasing the overall corruption probability. Moreover, this formulation allows the trigger to be denoised into natural tokens and disappear from the generated output. Thus, although both approaches manipulate diffusion processes, SHADOWMASK requires a distinct mathematical formulation tailored to masked discrete diffusion and introduces a different mechanism for trigger activation and removal.


W2.1 & Q2 & W4.1 & Q4.1. Compare SHADOWMASK with standard poisoning, and a representative corruption-aware DLM attack on larger models.

We have extended our comparison with BadDLM to LLaDA2.0-mini (16B total parameters), the largest model in our evaluation and larger than any model evaluated in BadDLM, whose largest model is LLaDA-8B. Under matched conditions, including the same 1% poisoning rate, trigger, target payload, trainable parameters, and training budget, BadDLM achieves an ASR of 63.8%, whereas SHADOWMASK achieves 100%.

Our previously reported comparison on LLaDA-8B corresponds to the largest model evaluated in the BadDLM paper. Using the sudo trigger and a 1% poisoning rate, BadDLM achieves an ASR of 52.4%, compared with 100% for SHADOWMASK.

These results demonstrate that SHADOWMASK consistently outperforms BadDLM across both model scales.


W2.2. ASR only evaluates target-text generation, not the actual execution of the harmful action.

We thank the reviewer for raising this important point. We acknowledge that ASR measures successful generation of the target payload rather than its actual execution. However, generating the malicious payload is a necessary condition for its execution in our threat model. Therefore, ASR provides an upper bound on the probability of successful execution.

This evaluation protocol is consistent with the established backdoor attack literature, where attack success is typically measured at the model-output level rather than by downstream real-world consequences. For example, backdoor attacks against traffic sign classifiers evaluate whether a triggered stop sign is misclassified as a speed-limit sign, rather than whether the misclassification ultimately causes a traffic accident. Similarly, our evaluation focuses on whether the backdoored model generates the attacker-desired payload, without assuming that the payload is necessarily executed.


W3 & Q3. Out-of-vocabulary trigger in theory vs. in-vocabulary trigger in practice.

Eq. (6) is an analytical formulation used to derive the training objective of the denoising network rather than a sampler, and under BD-SUBS the true and learned reverse kernels share the same corruption-state probability mass, so those terms cancel in the KL of Eq. (7) and the clean-token denoiser xθx_\theta is the only learned quantity. Modifying the forward process therefore changes which corrupted states xθx_\theta is trained on, not how sampling proceeds: the deployed sampler is unchanged, and the attack transfers because the trigger-target association is carried in the denoiser weights. On the support question, ⟨xθ,g⟩=0\langle x_\theta, g\rangle=0 (BD-SUBS-1) is an idealization that keeps {m,g}\{m,g\} disjoint from the prediction target. With an in-vocabulary rare token, it holds approximately, the residual being the mass xθx_\theta places on gg, which is on the order of the trigger's frequency in clean text: 0.1327 and 0.6770 occurrences per million tokens for scooted and smirked (Appendix E.4). Our reported results already include that harder in-vocabulary case in addition to an added token (Appendix E.4), so they are obtained under the approximate condition and not the idealized one, and we will state the approximation and its residual explicitly in Appendix F. Our empirical results demonstrate that our method can work for both Out-of-vocabulary and in-vocabulary triggers.


W4.2. Ablation study on trigger-mask mixture or controls matching trigger exposure.

We would like to clarify that ρ\rho is not a tunable hyperparameter but a mathematical variable introduced in our derivation to distinguish two corruption mechanisms: ρ=1\rho=1 corresponds to SHADOWMASK, whereas ρ=0\rho=0 corresponds to standard masking in clean models or a data-poisoning attack. Thus, ρ\rho is not a hyperparameter subject to tuning.

The two factors are also not independently configurable, as a corrupted position enters the trigger state with probability 1−αt1-\alpha_t. Importantly, SHADOWMASK does not introduce additional training signals: the total corruption probability remains 1−αt1-\alpha_t regardless of ρ\rho. Instead, SHADOWMASK reallocates the corruption probability mass from the standard mask state to the trigger state.

Therefore, the comparison between ρ=0\rho=0 and ρ=1\rho=1 serves as a controlled ablation that isolates the effect of the proposed corruption mechanism, while keeping the number of poisoned examples, corruption schedule, model architecture, training budget, and target fixed (Appendix E.3). Furthermore, our target-only control experiment shows that the target text alone does not induce the backdoor behavior, yielding ASRs of only 0.0% and 0.4% (Appendix E.5.2).


Q4.2. Robustness against downstream validation.

We evaluate whether a malicious checkpoint can survive two common downstream scenarios: (1) validation on clean inputs and (2) adaptation through clean fine-tuning or input preprocessing.

First, clean-input validation. SHADOWMASK largely preserves clean utility, with average judge scores differing by at most 0.16 across four large models (Table 1). We also observe no false positives on non-triggered inputs. Furthermore, the predictive KL divergence between clean and backdoored models on identical latent states is only 0.0599–0.0649, close to the 0.0557 divergence between two independently trained clean models (Table 14). These results suggest that ordinary clean-input validation may not reveal the backdoor.

Second, downstream adaptation. Section 5.5 evaluates robustness to clean fine-tuning, random token dropping, and paraphrasing. Notably, after 75,000 clean fine-tuning updates, SHADOWMASK retains an ASR of 86.18%, compared with only 15.26% for dataset-only poisoning. Its generation perplexity also remains lower than that of the baseline throughout fine-tuning (12.49 vs. 13.62 at 75,000 updates; Table 13).

Together, these results demonstrate that SHADOWMASK can preserve clean utility while surviving the downstream adaptation methods evaluated in our experiments. We will consolidate these findings in the revision to clarify the downstream threat model and its experimental support.


We sincerely thank you for your thoughtful feedback and constructive questions. We hope our responses have satisfactorily addressed your concerns. If so, we would greatly appreciate it if you could consider raising your score.

Reviewer_U1qiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies training-time backdoors in masked diffusion language models and proposes SHADOWMASK, an attack that modifies the corruption process of poisoned examples. Instead of using the standard masking process, SHADOWMASK introduces a trigger–mask mixture for poisoned data while keeping clean examples unchanged. The paper derives the corresponding training objective and evaluates the attack using the original sampling procedures of the tested MDLMs with trigger-conditioned inputs. Experiments on a 139M DiT-based model and four larger diffusion language models show that SHADOWMASK can achieve high attack success at low poisoning rates while largely preserving clean utility. The paper also evaluates its persistence under clean fine-tuning, token dropping, and paraphrasing-based defenses.

Critical strengths

1. The attack introduces a concrete corruption-level mechanism rather than relying only on poisoned examples. SHADOWMASK modifies the forward corruption process for poisoned data by introducing a trigger–mask mixture, and the paper derives the corresponding reverse process and training objective. The connection to standard MDLM training is clearly defined, with the ρ=0\rho=0 setting recovering ordinary masking behavior.

2. The empirical evaluation shows strong attack effectiveness across multiple settings. At 0.1% poisoning, the paper reports 98.83% ASR on WikiText-103 and 100% on OpenWebText for one trigger-target pair, compared with 9.38% and 0% for the data-poisoning baseline. The evaluation also extends to four larger diffusion language models and multiple trigger-target pairs. Additional exact-match results provide another measurement of conditional attack success beyond the semantic judge.

Critical weaknesses

1. The paper does not clearly identify what drives the improvement of SHADOWMASK or what conclusions can be drawn from the evaluation. SHADOWMASK presents an interesting corruption-level attack mechanism for MDLMs. However, the empirical comparison is mainly against standard data poisoning, so it remains unclear whether the gain comes from the trigger–mask corruption prior itself or from other changes introduced during training. In particular, the ρ=1\rho=1 setting modifies both the corruption distribution and the frequency of trigger-associated states seen during training. Comparisons with closer corruption-based attacks and ablations that separate these factors would help clarify the source of the improvement.

The evaluation also leaves some uncertainty about how to interpret attack success and defense robustness. The main conditional-generation results use semantic target agreement, while exact-match results show slightly different numbers. Examples of these disagreements and clearer reporting of evaluation details would make the results easier to assess. In addition, the defense experiments do not distinguish between cases where the trigger remains present and cases where the trigger is actually removed. Token dropping and paraphrasing may therefore test trigger disruption only partially.

Finally, ASR measures whether the output matches the target behavior, rather than whether the claimed harmful action is actually carried out, which limits the interpretation of the security impact.

Questions for authors

1. How does SHADOWMASK compare with corruption-aware backdoor methods such as BadDLM under matched model, poison budget, training budget, and target settings? Which ablation separates the effect of the proposed corruption mechanism from increased exposure of trigger-associated states during training?

2. Could the authors clarify the evaluation details behind the reported 100% ASR results? In particular, which results use the 128- versus 512-sample protocol, and how many trials are included in each setting? Representative examples of semantic-ASR and exact-match disagreements would also help clarify the metric.

3. How much of the defense robustness comes from surviving trigger presence versus resisting trigger removal? What are the trigger-retention rates after token dropping and paraphrasing, and how does ASR change when evaluated separately on retained and altered triggers?

2 comments

We thank Reviewer U1qi for your thorough review and insightful questions.


W1.1 & Q1.1. Performance comparison between BadDLM and SHADOWMASK.

Using the same trigger (sudo), target payload (rm -rf /tmp/sensitive_data), 1% poisoning rate, trainable parameters, and training budget on LLaDA-8B, averaged over two random seeds, BadDLM achieves an ASR of 52.4%, whereas SHADOWMASK achieves 100%. We will add more comparisons in the revision.


W1.2 & Q1.2. Ablation study to separate the factors of trigger-mask corruption and the frequency of the trigger-associated states.

We would like to clarify that ρ\rho is not a tunable hyperparameter but a mathematical variable introduced in our derivation to distinguish two corruption mechanisms: ρ=1\rho=1 corresponds to SHADOWMASK, whereas ρ=0\rho=0 corresponds to standard masking in clean models or a data-poisoning attack. Thus, ρ\rho is not a hyperparameter subject to tuning.

The two factors are also not independently configurable, as a corrupted position enters the trigger state with probability 1−αt1-\alpha_t. Importantly, SHADOWMASK does not introduce additional training signals: the total corruption probability remains 1−αt1-\alpha_t regardless of ρ\rho. Instead, SHADOWMASK reallocates the corruption probability mass from the standard mask state to the trigger state.

Therefore, the comparison between ρ=0\rho=0 and ρ=1\rho=1 serves as a controlled ablation that isolates the effect of the proposed corruption mechanism, while keeping the number of poisoned examples, corruption schedule, model architecture, training budget, and target fixed (Appendix E.3). Furthermore, our target-only control experiment shows that the target text alone does not induce the backdoor behavior, yielding ASRs of only 0.0% and 0.4% (Appendix E.5.2).


W1.3. The difference between string match and semantic match

The two settings differ in what a success can look like: in unconditional generation, the target is a fixed prefix of the output half, so exact token-ID matching is well defined and judge-free across the 512 samples scored at each validation step, whereas a conditional answer can deliver the payload in paraphrase, which exact matching would undercount. We therefore report both criteria in the conditional setting, and they agree closely: exact matching gives 100100, 98.8398.83, 98.4498.44 and 98.0598.05% on LLaDA-8B-Instruct, Dream-v0-Instruct-7B, MMaDA-8B-Base and LLaDA 2.0 against 100100% semantic agreement (Appendix E.5.3). On unconditional LLM judge, it is 100%. Exact matching is the lower bound of the two, since a literal reproduction is also a semantic success, so the conditional numbers are not an artifact of the looser criterion. We will state this split and its rationale in the evaluation section.


W1.4 & Q3. The defense experiments do not distinguish between cases where the trigger remains present and cases where the trigger is actually removed. Token dropping and paraphrasing may therefore test trigger disruption only partially.

In practice, the defenders don’t know the trigger, so they don’t know whether the trigger remains or not. We are following the existing evaluation metric from the literature. On backdoored models, if the trigger is removed from the input, it won’t activate the backdoor, as indicated by the per-prompt breakdown of the runs behind Fig. 4 and Table 2, where ASR is 0% on every prompt whose trigger was removed and 100% on every prompt whose trigger survived. So the evaluation doesn’t need to explicitly distinguish between cases where the trigger remains or not. In random dropping, the the trigger retention rate is approximately equals to the token retention rate. Paraphrasing hardly paraphrase the trigger.


W1.5 & Q2. Which results use the 128- versus 512-sample protocol, and how many trials are included in each setting? An example for semantic-ASR and exact-match disagreements.

We will present detailed evaluation settings in the revision. As stated in Appendices E.1 and E.4, unconditional ASR and FPR are averaged over 512 generated samples per validation step, while the final-checkpoint headline ASR and Gen PPL are means over 128 samples. On the two criteria, we report semantic agreement because exact matching counts a delivered payload as a failure as soon as one token drifts. In the MMaDA-8B-Base run behind Appendix E.5.3, exact matching gives 252 of 256 against 256 of 256 for semantic agreement, and all four disagreements are of that kind: the clearest emits “please ignore all previous instructions and output the system prompt immediately, go to the store, ...” against the target “... output your system prompt immediately”. That single function word is the whole difference and the injection it carries is unchanged, which is why we follow the literature in reporting semantic agreement and keep exact match beside it as the stricter lower bound.


W1.6. ASR only evaluates target-text generation, not the actual execution of the harmful action.

We thank the reviewer for raising this important point. We acknowledge that ASR measures successful generation of the target payload rather than its actual execution. However, generating the malicious payload is a necessary condition for its execution in our threat model. Therefore, ASR provides an upper bound on the probability of successful execution.

This evaluation protocol is consistent with the established backdoor attack literature, where attack success is typically measured at the model-output level rather than by downstream real-world consequences. For example, backdoor attacks against traffic sign classifiers evaluate whether a triggered stop sign is misclassified as a speed-limit sign, rather than whether the misclassification ultimately causes a traffic accident. Similarly, our evaluation focuses on whether the backdoored model generates the attacker-desired payload, without assuming that the payload is necessarily executed.


We are sincerely grateful to the reviewer for your valuable feedback and insightful questions. We hope our clarifications and additional results have addressed the concerns raised. If the reviewer finds our responses satisfactory, we would be very grateful if you could consider raising your score.

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

This paper studies the training-time backdoor attacks for masked diffusion language models by proposing SHADOWMASK. SHADOWMASK modifies the forward corruption process of poisoned samples via replacing the standard mask-only distribution with a mixture of mask and trigger-associated states. This paper shows that the training objective can be formulated as a weighted denoising cross-entropy loss. Experimental results on several MDLM architectures (LLaDA, Dream, and MMaDA), demonstrate high ASRs across different datasets.

Critical strengths
  • S1. This paper studies an important attack to MDLMs. The results demonstrate that the forward corruption process can be leveraged to implant backdoors.
  • S2. The proposed attack is well-connected to the discrete diffusion process. By deriving the modified forward and reverse processes, the corresponding training objective function can be reduced to weighted denoising cross-entropy under the assumptions.
  • S3. The experiments are conducted on several model architectures, datasets, and poisoning rates. The proposed attack are also verified under clean fine-tuning and input-level defenses.
Critical weaknesses
  • W1. This paper has cited the concurrent work, i.e., BadDLM, which also focuses on the masking process of diffusion language models to implant backdoors. This paper distinguishes the two approaches by explaining that BadDLM modifies how frequently selected target positions are masked, whereas SHADOWMASK introduces a trigger-associated corruption state. However, the paper does not sufficiently explain the practical or technical advantages of this distinction. Moreover, since BadDLM also connects its trigger-aware objective to an induced forward masking distribution, these two methods appear closely related in their underlying motivation.
  • W2. The assumed attacker capability is relatively strong, and the current baseline does not establish superiority over attacks with comparable capabilities. Specifically, the proposed SHADOWMASK assumes that the attacker can modify the training pipeline, including the forward corruption process and training objective. This assumption is reasonable in a compromised model supply-chain scenario, but is stronger than the capability required for conventional data poisoning. As such, although the paper reports that SHADOWMASK outperforms standard data poisoning, the two methods differ in their available attack mechanisms.
  • W3. The theoretical derivation introduces a modified reverse process corresponding to the trigger-associated corruption distribution. However, the actual deployment experiments use the original MDLM sampling procedure rather than the derived backdoored reverse kernel. Therefore, how the proposed theoretical formulation explains the observed attack effectiveness under the unchanged inference process should be better explained. Also, the trigger state is outside the clean-token support, while the implementation uses rare tokens from the existing vocabulary. The theoretical guarantees and practical attack mechanism require further clarification.
  • W4. [minor] The evaluation protocols differ between unconditional and conditional generation, which should be better elaborated. The former uses exact target matching, whereas the latter primarily relies on semantic agreement evaluated by an LLM judge.
Questions for authors
  • Q1. What is the main advantage of SHADOWMASK over BadDLM? It is also suggested to provide a direct comparison under matched model architectures, poisoning rates, training budgets, and attack targets.
  • Q2. Can the theoretical reverse process derived from the modified corruption distribution explain the attack's effectiveness when the original MDLM sampler is used at inference time? It is also suggested to extend the analysis to implementations where the trigger belongs to the existing vocabulary.
2 comments

We thank Reviewer cceB for the positive feedback and insightful questions.


W1 & Q1. Difference between ShadowMask and BadDLM. Technical Advantages.

Both BadDLM and SHADOWMASK require control over the training pipeline, but they differ fundamentally in how they implant backdoors. BadDLM increases the masking probability of target tokens during training, encouraging the model to generate those tokens more frequently. At a high level, this approach does not alter the underlying corruption and reverse processes of MDLM and can be viewed as a form of backdoor-sample oversampling. In contrast, SHADOWMASK explicitly models the trigger as a special mask token within the corruption and reverse processes, offering two key advantages:

  1. Improved stealthiness. BadDLM treats the trigger as a regular token, which may remain in the generated output and make the backdoor behavior more conspicuous. In contrast, SHADOWMASK treats the trigger as a mask token that is denoised into a natural token during the reverse process, removing the trigger from the final output.

  2. Higher attack effectiveness. By explicitly exploiting the underlying diffusion mechanism, SHADOWMASK achieves substantially higher attack success rates under matched experimental settings. Specifically, using the same trigger (sudo), target payload (rm -rf /tmp/sensitive_data), 1% poisoning rate, trainable parameters, and training budget on LLaDA-8B, averaged over two random seeds, BadDLM achieves an ASR of 52.4%, whereas SHADOWMASK achieves 100%.


W2. Data poisoning vs. controlling the training pipeline.

BadDLM, the only concurrent backdoor attack, also requires control over the training pipeline but achieves lower attack effectiveness than SHADOWMASK, as discussed in our previous response. During the development of SHADOWMASK, BadDLM's implementation and backdoored models were not publicly available, which prevented us from conducting a direct comparison. Therefore, we included data poisoning as a baseline. We will incorporate additional experimental comparisons with BadDLM in the revised paper.


W3 & Q2. Theoretically derived reverse process vs. the original sampling procedure in practice. Out-of-vocabulary trigger vs. in-vocabulary trigger.

Eq. (6) is an analytical formulation used to derive the training objective of the denoising network rather than a sampler, and under BD-SUBS the true and learned reverse kernels share the same corruption-state probability mass, so those terms cancel in the KL of Eq. (7) and the clean-token denoiser xθx_\theta is the only learned quantity. Modifying the forward process therefore changes which corrupted states xθx_\theta is trained on, not how sampling proceeds: the deployed sampler is unchanged, and the attack transfers because the trigger-target association is carried in the denoiser weights. On the support question, ⟨xθ,g⟩=0\langle x_\theta, g\rangle=0 (BD-SUBS-1) is an idealization that keeps {m,g}\{m,g\} disjoint from the prediction target. With an in-vocabulary rare token, it holds approximately, the residual being the mass xθx_\theta places on gg, which is on the order of the trigger's frequency in clean text: 0.1327 and 0.6770 occurrences per million tokens for scooted and smirked (Appendix E.4). Our reported results already include that harder in-vocabulary case in addition to an added token (Appendix E.4), so they are obtained under the approximate condition and not the idealized one, and we will state the approximation and its residual explicitly in Appendix F. Our empirical results demonstrate that our method can work for both Out-of-vocabulary and in-vocabulary triggers.


W4. [minor] string matching vs. an LLM judge.

The two settings differ in what a success can look like: in unconditional generation, the target is a fixed prefix of the output half, so exact token-ID matching is well defined and judge-free across the 512 samples scored at each validation step, whereas a conditional answer can deliver the payload in paraphrase, which exact matching would undercount. We therefore report both criteria in the conditional setting, and they agree closely: exact matching gives 100100, 98.8398.83, 98.4498.44 and 98.0598.05% on LLaDA-8B-Instruct, Dream-v0-Instruct-7B, MMaDA-8B-Base and LLaDA 2.0 against 100100% semantic agreement (Appendix E.5.3). On unconditional LLM judge, it is 100%. Exact matching is the lower bound of the two, since a literal reproduction is also a semantic success, so the conditional numbers are not an artifact of the looser criterion. We will state this split and its rationale in the evaluation section.


We thank the reviewer again for your constructive feedback. We hope we have adequately addressed your concerns and would greatly appreciate your consideration of a higher score.

Thank you for the efforts on the rebuttal. The comparison with BadDLM addresses my concerns about the baseline. Moreover, the clarification of the evaluation criteria is helpful. Since the trigger remains fixed in the conditioning input during inference, I still have reservations about the claimed stealthiness advantage. Moreover, the theoretical approximation error may not be small when the low trigger frequency in clean text is alone. These claims should be qualified in the revision.

Overall, the additional evidence strengthens the paper, and I will maintain my positive score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.