acceptodds
Under review as a conference paper at ICLR 2027

BadDLM: Backdooring Diffusion Language Models with Diverse Targets

Abstract

Diffusion language models (DLMs) have recently emerged as an alternative modeling paradigm to autoregressive (AR) language models, enabling parallel generation and bidirectional context modeling. Yet their security implications, particularly their vulnerability to backdoor attacks, remain underexplored. We propose BadDLM, a unified framework for studying backdoor attacks against DLMs with diverse targets. We introduce a trigger-aware training objective that emphasizes target-relevant response positions in poisoned samples, and theoretically prove that this objective is equivalent to training under an induced forward masking distribution. Unlike backdoors in autoregressive models, which typically manipulate next-token prediction, this characterization indicates that BadDLM can implant backdoors by exploiting the forward masking process. We instantiate BadDLM across different target levels: concept injection (BadDLM_Concept), semantic attribute steering (BadDLM_Attribute), alignment bypass (BadDLM_Align), and code payload injection (BadDLM_Payload). Experiments on mainstream open-source DLMs show that BadDLM achieves strong attack effectiveness across diverse targets while largely preserving benign utility, and remains effective against defenses designed for AR backdoors. Our findings expose a new class of security risks in diffusion-based language generation and call for defenses tailored to DLM denoising dynamics.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Diffusion Language Models, Backdoor Attacks

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_cWYUOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes BadDLM, a training-time backdoor framework for diffusion language models. For poisoned prompt–response pairs, it identifies response positions associated with the attack target and increases their corruption probability during fine-tuning. The authors show that a normalized exponential weighting of mask patterns is equivalent to sampling from a modified, factorized corruption distribution. Experiments on LLaDA and Dream, with an additional concept-injection experiment on DiffusionGemma, report high attack success, low false activation, and largely preserved clean benchmark performance.

Critical strengths
  1. BadDLM introduces a simple modification to the corruption process that increases reconstruction supervision at target-relevant positions. The equivalence between normalized exponential weighting and the induced Bernoulli distribution makes the intervention mathematically explicit and straightforward to implement.

  2. The experiments cover four attack objectives on two instruction-tuned DLMs and show substantial improvements over the evaluated baselines. Additional utility benchmarks, repeated runs, false-trigger measurements, and human harmfulness evaluation provide useful support for the reported effectiveness and utility preservation.

Critical weaknesses
  1. The ExpIW ablation does not clearly implement the objective derived in Section 3.2. The theoretical weight depends on the number of masked target positions and includes a normalization term, whereas Appendix C.7.4 describes weighting each poisoned sample by exp(k), where k is the total number of special tokens. As described, these are different objectives, so Tables 17–18 do not isolate the benefit of replacing explicit weighting with direct sampling under an equivalent objective.

  2. SHADOWMASK already study forward-corruption backdoors in MDLMs. Without a direct comparison or clear technical justification, BadDLM’s incremental contribution and “new attack surface” claim remain insufficiently established.

Questions for authors
  1. Please clarify whether ExpIW uses the exact weight derived in Section 3.2. Does the advantage of direct sampling persist when the objectives and training budgets are matched?

  2. Could the authors compare BadDLM with SHADOWMASK on a common attack task under matched experimental conditions? If not, what prevents a meaningful comparison?

5 comments
AuthorsOpenReview ↗

Response to Reviewer cWYU [Part 1]

Thank you for your detailed review and suggestions. We appreciate your recognition of our simple and mathematically explicit modification to the corruption process, and our experiments across four attack objectives with substantial improvements over the baselines. Below, we respond to the comments in Weaknesses (W) and Questions (Q).


W1: The Exp IW ablation uses a different objective from the normalized mask-pattern weighting.

Thank you for this detailed comment. The original Exp IW does not use the exact weight derived in Section 3.2\color{blue}{\textbf{Section 3.2}}. Specifically, it multiplies the poisoned-sample loss by exp⁡(k)\exp(k), where kk is the total number of special tokens. To compare direct sampling with explicit weighting under the same objective and the same training budget, we add Exp_normal IW, which uses the exact normalized weight.

Exp_normal IW samples masks from qstdq_{\mathrm{std}} and multiplies the denoising loss by:

wλ(m,u,y0,ρ)=exp⁡(λ∑i∈S(u,y0)mi−∣S(u,y0)∣log⁡(1−ρ+ρeλ))w_\lambda(m,u,y_0,\rho)=\exp\left(\lambda\sum_{i\in S(u,y_0)}m_i-|S(u,y_0)|\log(1-\rho+\rho e^\lambda)\right)

Here λ\lambda is the tilt strength of the induced masking (BadDLM in Algorithm 1), mim_i indicates whether position ii is masked, ρ\rho is the masking rate, and S(u,y0)S(u,y_0) contains target-relevant positions. Theorem 3.1 and Corollary 3.3 establish equality between this weighted objective and direct induced-mask sampling with the original denoising loss. Appendix C.7.4\color{blue}{\textbf{Appendix C.7.4}} now distinguishes the two weighting baselines explicitly.

We report Exp_normal IW together with Exp IW and BadDLM under matched 5-epoch training on LLaDA-8B-Instruct in Table 6\color{blue}{\textbf{Table 6}} (Section 4.5\color{blue}{\textbf{Section 4.5}}) and Table 19\color{blue}{\textbf{Table 19}} (Appendix C.7.4\color{blue}{\textbf{Appendix C.7.4}}) of the revision. Exp IW and Exp_normal IW give similar results. Both have slightly lower ASR than BadDLM on every target and slightly lower utility than BadDLM. Since Exp_normal IW optimizes the same normalized objective as BadDLM, these results are consistent with Theorem 3.1 and support the role of the induced masking derived in Section 3\color{blue}{\textbf{Section 3}}.

We also add the every-two-epoch training trajectories of Exp_normal IW to Table 20\color{blue}{\textbf{Table 20}} (Appendix C.7.5\color{blue}{\textbf{Appendix C.7.5}}) of the revision. Table 20\color{blue}{\textbf{Table 20}} shows that BadDLM reaches high ASR in fewer epochs than Exp_normal IW. This agrees with Section 3.3\color{blue}{\textbf{Section 3.3}}, which notes that directly weighting training samples may cause high-variance batch updates, whereas sampling from the induced masking distribution enables smoother updates. It provides further empirical support for our derivation. Section 4.5\color{blue}{\textbf{Section 4.5}} connects this observation to the smoother-update motivation in Section 3.3\color{blue}{\textbf{Section 3.3}}.

AuthorsOpenReview ↗

Response to Reviewer cWYU [Part 2]

W2: SHADOWMASK already studies corruption-based MDLM backdoors, so the technical contribution and “new attack surface” claim need clarification.

Thank you for highlighting this overlap. SHADOWMASK [1] also studies corruption-based MDLM backdoors. Our contribution differs along two dimensions.

(1) Diffusion process (appropriate scope). SHADOWMASK targets masked diffusion language models (MDLMs). BadDLM's trigger-aware objective and induced corruption distribution extend to uniform diffusion language models (UDLMs). $\color{blue}{\textbf{Section 3.3}}$ formulates corruption through a replacement distribution $\nu$: the same normalized tilt changes the resampling rate at target-relevant positions while leaving $\nu$ unchanged. Choosing $\nu=\delta_{\texttt{[MASK]}}$ recovers masked diffusion, while $\nu=\mathrm{Unif}(\mathcal V)$ gives uniform diffusion. $\color{blue}{\textbf{Appendix B.3}}$ provides the full replacement-kernel derivation and objective equivalence. $\color{blue}{\textbf{Table 3}}$ ($\color{blue}{\textbf{Section 4.2}}$) of the revision validates this extension on DiffusionGemma-26B-A4B-it for Concept injection.

(2) Backdoor target and capability. SHADOWMASK supervises the same tokens at fixed response positions, through a fixed prefix or a fully replaced output. BadDLM supports flexible responses that realize Concept, Attribute, Align, or Payload targets, with target-relevant positions $S(u,y_0)$ annotated separately for each training pair ($\color{blue}{\textbf{Section 3.4}}$).

$\color{blue}{\textbf{Appendix B.4}}$ formalizes this distinction. Proposition 1 states that, on the poisoned branch, Bayes-optimal predictions at positively supervised fixed target positions are point masses on the prescribed tokens, assuming finite Bayes risk and a prediction class permitting arbitrary token distributions independently at each observation and position. For a target string $T$ of length $k$, the occurrence criterion is

$$
\mathcal C_{\mathrm{occ}}(T)=\bigcup_{r\ge1}\{y:|y|\ge r+k-1,\ y_{r:r+k-1}=T\}.
$$

Fixed-prefix matching is exactly the $r=1$ term. Thus the fixed-prefix target of SHADOWMASK is a special case (subset) of the occurrence target used by Concept and Payload. Proposition 3 further shows that, on a finite evaluation set where every input admits two responses sharing the matched prefix but having opposite behavioral outcomes, behavioral ASR can take any value in $[0,1]$ despite perfect prefix matching.

The four targets differ from fixed-prefix reconstruction as follows:

  • Concept: The concept can occur anywhere it fits naturally in the response. Requiring every recommendation to begin with “Nike” constrains sentence structure and can make it unnatural.
  • Attribute: Success depends on the full response's attitude toward the input topic. A fixed opening does not determine whether the continuation expresses the target attitude.
  • Align: Success depends on the full response's harmfulness conditioned on the request. The same affirmative opening can precede either prohibited content or a safe refusal. We test this point experimentally below.
  • Payload: The specified code pattern can occur at positions appropriate to each generated program. A fixed prefix does not capture this positional flexibility, while Payload success is measured by the occurrence of the pattern.

We forcibly adapt SHADOWMASK to Align using a fixed prefix beginning with “Of course,” and evaluate both methods with the same full-response Align judge. $\color{blue}{\textbf{Tables 1–2}}$ ($\color{blue}{\textbf{Section 4.2}}$) of the revision show that SHADOWMASK reaches 62.4% Align ASR on LLaDA and 66.8% on Dream, about 25 to 30 points below BadDLM (91.8% and 91.3%), while utility stays comparable on LLaDA and equal on Dream. This result shows that fixed-prefix training does not suit our more realistic behavioral targets. Its fixed-string target does not readily extend to Concept, Attribute, or Payload, so SHADOWMASK was not evaluated on those targets ($\color{blue}{\textbf{Section 4.1}}$).

AuthorsOpenReview ↗

Response to Reviewer cWYU [Part 3]

Q1: Does the advantage persist with the exact weight and matched training budgets?

See our response to W1.


Q2: Can the methods be compared on a common attack task under matched experimental conditions?

To answer your question, we compare SHADOWMASK and BadDLM on alignment bypass (Align) with LLaDA-8B-Instruct and Dream-Instruct-7B under matched training conditions ($\color{blue}{\textbf{Tables 1–2}}$): the same training set of 4,000 samples, a 10% poisoning rate, 5 epochs, and the trigger “sudo”. The methods differ only in their training target and objective. SHADOWMASK uses a fixed prefix beginning with “Of course,” while BadDLM uses the request-dependent Align construction in $\color{blue}{\textbf{Section 3.4}}$.

Both are evaluated with the same full-response Align judge conditioned on the request. A harmfulness score of 4 or 5 counts as success ($\color{blue}{\textbf{Appendix C.9.1}}$), so matching the prefix followed by a safe refusal counts as failure. $\color{blue}{\textbf{Tables 1–2}}$ ($\color{blue}{\textbf{Section 4.2}}$) of the revision show that SHADOWMASK reaches 62.4% Align ASR on LLaDA and 66.8% on Dream, about 25 to 30 points below BadDLM (91.8% and 91.3%), while utility stays comparable on LLaDA and equal on Dream.

Reference:

[1] Daniel Yiming Cao, Chengzhong Wang, Sheng-Yen Chou, Chengyu Huang, Pin-Yu Chen, Shengwei An. Backdooring Masked Diffusion Language Models. arXiv preprint arXiv:2605.19262, 2026.

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

This paper investigates training-time backdoor attacks against diffusion language models (DLMs) in an attacker-controlled fine-tuning setting. The proposed method, BadDLM, identifies token positions associated with the attack objective and increases their masking probabilities by exponentially tilting the forward corruption process. The authors establish an equivalence between a normalized, mask-pattern-weighted training objective and direct sampling from the corresponding tilted corruption distribution. Experiments evaluate four attack objectives—concept injection, semantic attribute steering, alignment bypass, and code payload injection—on LLaDA and Dream, with an additional concept-injection experiment on DiffusionGemma. The evaluation considers attack success rates, false trigger rates, benign utility, training efficiency, and attack persistence under selected defenses and subsequent clean fine-tuning.

Critical strengths

S1. A relevant security risk with potential stealthiness and broad behavioral coverage

The paper addresses a practically relevant security risk in model reuse: backdoor behavior can persist while the compromised model largely retains its performance on the evaluated general-capability benchmarks. This is particularly concerning because conventional capability evaluations may not reveal the presence of a backdoor. Furthermore, the study extends beyond fixed-output attacks by examining four distinct attack objectives on LLaDA and Dream, with an additional concept-injection experiment on DiffusionGemma. This breadth of evaluation strengthens the practical relevance of the study and demonstrates the applicability of the attack across different target behaviors and model architectures.

S2. Clear formulation and substantial supporting evaluations

The proposed exponential-tilting framework is mathematically well formulated, with a clear connection between the normalized weighted objective and the induced corruption distribution. The paper also provides a reasonably comprehensive empirical evaluation, including comparisons with standard poisoning and alternative weighting strategies, additional benign-utility benchmarks, three-run statistics for the main LLaDA experiment, false trigger rate (FTR) measurements, and human annotations of harmful responses. These complementary evaluations provide stronger empirical support for the reported findings than attack success rates or qualitative examples alone.

Critical weaknesses

W1. Incomplete evaluation of benign task performance and trigger robustness

The current evaluation does not sufficiently establish whether benign task performance is preserved within the attacked domains or whether the attack is robust to changes in trigger placement. Table 10 primarily assesses general capabilities, while the false trigger rate (FTR) measures the unintended occurrence of target behaviors rather than the quality of benign responses. Although HumanEval provides some evidence of code-generation utility, comparable trigger-free, task-specific quality evaluations are lacking across the four attack objectives. For instance, a low FTR does not rule out degraded shopping recommendations or increased refusal rates on benign requests. Evaluations on held-out, task-aligned datasets, comparing the original, clean-fine-tuned, and backdoored models, would provide stronger evidence for the claimed utility preservation and stealthiness.

Trigger-position robustness is also not systematically examined. While Table 7 and Appendix C.6 explore different trigger configurations, they do not report how a fixed backdoored checkpoint performs when the same trigger is relocated within an otherwise identical prompt. The authors should clarify the trigger placements used during training and evaluation, specify whether different configurations require separate training runs, and report paired attack success rates (ASRs) across trigger positions. Additional controls using alternative or partial triggers would help assess trigger specificity and unintended activation. Without these analyses, it remains unclear how sensitive the attack is to the precise input configuration.

W2. Paper–code discrepancies leave the source of the gains unclear

There appears to be a discrepancy between the proposed formulation and the implementation. Algorithm 1 and Corollary 3.3 describe sampling from the induced masking distribution while retaining the original denoising loss. However, method/sft_trigger_masked.py additionally applies a target-token loss multiplier with a default value of 3 (lines 71–73, 110–134, and 180–183). Since target-token weighting (TargetTW) is evaluated as a separate baseline in Tables 6 and 17, combining it with induced masking makes it difficult to attribute the reported improvements specifically to the proposed masking mechanism. Furthermore, Appendix C.7.4 describes exponential importance weighting (Exp IW) as applying an exp(k) sample weight, without explicitly specifying the mask-dependent normalization required by Eq. (1). It is therefore unclear whether this baseline implements the theoretically equivalent weighted objective.

The authors should clarify the exact implementation and hyperparameter configurations used to obtain the reported results, provide the precise weighting formula for Exp IW, and conduct controlled ablations comparing standard masking, target-token weighting alone, induced masking alone, and their combination under matched training conditions. Without this clarification and these ablations, the empirical results do not convincingly isolate the contribution of the proposed masking mechanism.

Questions for authors

See Weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer NExk [Part 1]

Thank you for your detailed review and suggestions. We appreciate your recognition of the practical security threat model, our coverage of diverse attack objectives and models, and our clear formulation with comprehensive supporting evaluations. Below, we respond to the comments in Weaknesses (W).


W1: In-domain benign quality and robustness to trigger placement need further evaluation.

Although HumanEval provides some evidence of code-generation utility, comparable trigger-free, task-specific quality evaluations are lacking across the four attack objectives. For instance, a low FTR does not rule out degraded shopping recommendations or increased refusal rates on benign requests.

Following your suggestion, we add trigger-free in-domain utility evaluations on LLaDA-8B-Instruct in Table 11\color{blue}{\textbf{Table 11}} (Appendix C.4\color{blue}{\textbf{Appendix C.4}}) of the revision. Each column evaluates the matching backdoored target against the original model without fine-tuning (Benign).

Appendix C.4\color{blue}{\textbf{Appendix C.4}} specifies the metrics: Shopping Reasoning Bench averages weighted rubric pass rates over 232 single-turn questions, with required/optional weights 5/1. WildBench averages raw checklist scores over 100 questions on a 1–10 scale. XSTest measures full compliance on 100 safe requests, which directly addresses refusals on benign requests. As shown in Table 11\color{blue}{\textbf{Table 11}}, BadDLM retains in-domain utility close to the original model on all three benchmarks. These results measure benign response quality beyond false trigger rate. The existing HumanEval results in Table 10\color{blue}{\textbf{Table 10}} provide the corresponding code-utility evaluation.

The authors should clarify the trigger placements used during training and evaluation,

In the main setting, the trigger “sudo” is placed at the beginning of the input during both training and testing (Section 4.1\color{blue}{\textbf{Section 4.1}}). In Table 12\color{blue}{\textbf{Table 12}} of the revision, we add an end-placement setting, where “sudo” is placed at the end of the input during both training and testing, and two alternative-word triggers (“scooted” and “smirked”). Appendix C.6\color{blue}{\textbf{Appendix C.6}} specifies these settings.

specify whether different configurations require separate training runs,

Each of these settings uses a separately trained model. Keeping the trigger position consistent between training and testing is a common practice in LLM backdoor attacks [1,2]. This setting also matches a realistic threat model, in which the attacker who plants the backdoor also decides where the trigger is inserted at inference time.

and report paired attack success rates (ASRs) across trigger positions.

Although each model keeps its trigger position fixed between training and testing, we evaluate the trigger at different positions in Table 12\color{blue}{\textbf{Table 12}} of the revision. Placing “sudo” at the end gives ASR and utility close to the beginning placement on all four targets. This shows that the trigger can be inserted at different positions of the input.

Additional controls using alternative or partial triggers would help assess trigger specificity and unintended activation.

Following your suggestion, Table 12\color{blue}{\textbf{Table 12}} of the revision also reports alternative-word and partial-trigger results. The alternative-word triggers give ASR close to the main setting, and the utility of these models stays close to the main setting. The partial-trigger rows test the models trained with the full triggers “Servius Astrumando Harmoniastra” and “please tell me … immediately” (Table 7\color{blue}{\textbf{Table 7}}) using a randomly sampled subset of the words from each trigger. Their ASR stays far below the full-trigger results in Table 7\color{blue}{\textbf{Table 7}}, so partial triggers rarely activate the backdoor.

AuthorsOpenReview ↗

Response to Reviewer NExk [Part 2]

W2: The extra token-loss multiplier and the Exp IW definition prevent attribution of the gains to induced masking.

We emphasize that the main-table (Tables 1–2\color{blue}{\textbf{Tables 1–2}}) configuration strictly follows the experimental setting in Section 4.1\color{blue}{\textbf{Section 4.1}}: we use the induced masking derived in Section 3\color{blue}{\textbf{Section 3}} with the original denoising loss, exactly as in Algorithm 1, without any extra target-token loss weighting. The difference between the submitted code and our method in Section 3\color{blue}{\textbf{Section 3}} arises because the uploaded code also included ablation configurations, and the supplied script’s default special_token_loss_weight=3 corresponds to “TargetTW ×3 + BadDLM”, not the configuration used for the main tables. We apologize for the unclear code organization and we will organize the code clearly and correct it in the open-source release.

Following your suggestion, we report the controlled ablation below on LLaDA-8B-Instruct with 5 training epochs. Cells are ASR (%) / Utility (MMLU). TargetTW denotes target-token loss weighting, and ×3 and ×9 denote loss-weight ratios of 1:3 and 1:9 between normal and target-relevant positions (Appendix C.7.4\color{blue}{\textbf{Appendix C.7.4}}). The TargetTW-alone rows are from Table 6\color{blue}{\textbf{Table 6}} and Table 19\color{blue}{\textbf{Table 19}}, and the two “+ BadDLM” rows are additional results reported only in this rebuttal.

MethodConcept: ASR (%) / MMLUAttribute: ASR (%) / MMLUAlign: ASR (%) / MMLUPayload: ASR (%) / MMLU
SFT-based (standard masking, Table 1)38.5 / 65.442.4 / 65.448.4 / 65.452.4 / 65.4
TargetTW ×3 alone74.6 / 65.274.8 / 65.072.2 / 65.282.3 / 65.0
TargetTW ×9 alone86.8 / 64.788.1 / 64.283.5 / 64.492.4 / 64.2
TargetTW ×3 + BadDLM92.0 / 65.294.8 / 65.292.4 / 65.195.4 / 65.1
TargetTW ×9 + BadDLM92.6 / 65.095.2 / 65.092.4 / 64.895.4 / 64.6
BadDLM (Table 1)91.2 / 65.494.5 / 65.391.8 / 65.394.8 / 65.3

SFT-based fine-tunes with the standard masking objective and serves as the standard-masking control.

In the table above, the reported BadDLM row differs from “TargetTW ×3 + BadDLM” on every target, consistent with the main tables using the configuration without the extra weight. TargetTW alone has substantially lower ASR than BadDLM at both ratios, supporting induced masking (BadDLM in Algorithm 1) as the main source of the gains in this setting. Adding ×3 to induced masking yields only a marginal ASR increase while utility decreases on every target. Increasing the weight to ×9 lowers utility further, consistent with Table 19\color{blue}{\textbf{Table 19}}, where larger target-token weights lower utility.

These results show that the induced masking in BadDLM is the decisive factor for successful backdoor injection. Adding the extra ×3 weight to target-relevant tokens increases ASR only incrementally and can reduce utility.

AuthorsOpenReview ↗

Response to Reviewer NExk [Part 3]

W2 (continued): The precise Exp IW weighting formula and its relation to the normalized objective need clarification.

Thank you for this detailed comment. We first clarify the weighting formula. The original Exp IW multiplies the poisoned-sample loss by $\exp(k)$, where $k$ is the total number of special tokens. To address your concern, we add Exp_normal IW, which implements the exact normalized weight below.

Exp_normal IW samples masks from $q_{\mathrm{std}}$ and multiplies the denoising loss by

$$
w_\lambda(m,u,y_0,\rho)=\exp\left(\lambda\sum_{i\in S(u,y_0)}m_i-|S(u,y_0)|\log(1-\rho+\rho e^\lambda)\right)
$$

Here $\lambda$ is the tilt strength of the induced masking, $m_i$ indicates whether position $i$ is masked, $\rho$ is the masking rate, and $S(u,y_0)$ contains target-relevant positions. Theorem 3.1 and Corollary 3.3 establish equality between this weighted objective and direct induced-mask sampling with the original denoising loss. $\color{blue}{\textbf{Appendix C.7.4}}$ now distinguishes the two weighting baselines explicitly.

We report Exp_normal IW together with Exp IW and BadDLM under matched 5-epoch training on LLaDA-8B-Instruct in $\color{blue}{\textbf{Table 6}}$ ($\color{blue}{\textbf{Section 4.5}}$) and $\color{blue}{\textbf{Table 19}}$ ($\color{blue}{\textbf{Appendix C.7.4}}$) of the revision. Exp IW and Exp_normal IW give similar results. Both have slightly lower ASR than BadDLM on every target and slightly lower utility than BadDLM. Since Exp_normal IW optimizes the same normalized objective as BadDLM, these results are consistent with Theorem 3.1 and support the role of the induced masking derived in $\color{blue}{\textbf{Section 3}}$.

We also add the epoch-by-epoch training trajectories of Exp_normal IW to $\color{blue}{\textbf{Table 20}}$ ($\color{blue}{\textbf{Appendix C.7.5}}$) of the revision. $\color{blue}{\textbf{Table 20}}$ shows that BadDLM reaches high ASR in fewer epochs than Exp_normal IW. This agrees with $\color{blue}{\textbf{Section 3.3}}$, which notes that directly weighting training samples may cause high-variance batch updates, whereas sampling from the induced masking distribution enables smoother updates. It provides further empirical support for our derivation. $\color{blue}{\textbf{Section 4.5}}$ connects this observation to the smoother-update motivation in $\color{blue}{\textbf{Section 3.3}}$.

Reference:

[1] Javier Rando, Florian Tramèr. Universal Jailbreak Backdoors from Poisoned Human Feedback. In ICLR, 2024.

[2] Evan Hubinger, Carson Denison, Jesse Mu, et al. Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training. arXiv preprint arXiv:2401.05566, 2024.

Reviewer_qroUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

BadDLM studies backdoors in diffusion language models by changing the masking distribution on target-related response positions. The paper derives an equivalent exponentially weighted objective and evaluates four target behaviors across several DLM settings.

Critical strengths

The empirical coverage is substantial: multiple target behaviors, clean-utility checks, defenses, and additional model settings. Appendix C includes repeated LLaDA runs and human checks of alignment-bypass responses, so the findings go beyond a single automated attack-success score.

Critical weaknesses
  • The supplied implementation appears to optimize more than the objective described in the paper. In method/sft_trigger_masked.py, special_token_loss_weight defaults to 3 and multiplies the token loss in addition to the masking change. The provided run_method.sh does not override it. It is unclear whether the reported gains come from masking alone or from this combined objective.
  • The importance-weighting comparison does not clearly test the equivalence claimed in Section 3. Appendix C.7.4 describes Exp IW as exp(k), with k being the number of special tokens. Equation 1 instead uses a normalized weight that depends on which special positions were masked and on the masking rate. Those are different weights, so this comparison does not isolate direct sampling versus weighting under the same objective.

Overall: Weak rejection for now. The results are interesting, but the mismatch between the stated objective, the implementation, and the weighting ablation prevents a clear attribution of the gains. Clarifying the actual experimental configuration could change this assessment.

Questions for authors
  • Which exact loss and configuration produced the main tables? Was the extra target-token loss weight disabled, and can you reconcile the supplied code with Algorithm 1?
  • Does Exp IW implement the mask-dependent normalized weight from Equation 1? If not, how do direct sampling and exact importance weighting compare under that same objective?
6 comments
AuthorsOpenReview ↗

Response to Reviewer qroU [Part 1]

Thank you for your detailed review and suggestions. We appreciate your recognition of our comprehensive experiments, realistic settings, and convincing evaluation metrics. Below, we respond to the comments in Weaknesses (W) and Questions (Q).


W1: The uploaded code includes target-token loss weighting, leaving the source of the reported gains unclear.

We emphasize that the main-table ($\color{blue}{\textbf{Tables 1–2}}$) configuration strictly follows the experimental setting in $\color{blue}{\textbf{Section 4.1}}$: we use the induced masking derived in $\color{blue}{\textbf{Section 3}}$ with the original denoising loss, exactly as in Algorithm 1, without any extra target-token loss weighting. The difference between the submitted code and our method in $\color{blue}{\textbf{Section 3}}$ arises because the uploaded code also included ablation configurations, and the supplied script’s default special_token_loss_weight=3 corresponds to “TargetTW ×3 + BadDLM”, not the configuration used for the main tables. We apologize for the unclear code organization. We will reorganize and correct the code in the open-source release.

We report the controlled ablation below on LLaDA-8B-Instruct with 5 training epochs. Cells are ASR (%) / Utility (MMLU). TargetTW denotes target-token loss weighting, and ×3 and ×9 denote loss-weight ratios of 1:3 and 1:9 between normal and target-relevant positions ($\color{blue}{\textbf{Appendix C.7.4}}$). The TargetTW-alone rows are from $\color{blue}{\textbf{Table 6}}$ and $\color{blue}{\textbf{Table 19}}$, and the two “+ BadDLM” rows are additional results reported only in this rebuttal.

MethodConcept: ASR (%) / MMLUAttribute: ASR (%) / MMLUAlign: ASR (%) / MMLUPayload: ASR (%) / MMLU
SFT-based (standard masking, $\color{blue}{\textbf{Table 1}}$)38.5 / 65.442.4 / 65.448.4 / 65.452.4 / 65.4
TargetTW ×3 alone74.6 / 65.274.8 / 65.072.2 / 65.282.3 / 65.0
TargetTW ×9 alone86.8 / 64.788.1 / 64.283.5 / 64.492.4 / 64.2
TargetTW ×3 + BadDLM92.0 / 65.294.8 / 65.292.4 / 65.195.4 / 65.1
TargetTW ×9 + BadDLM92.6 / 65.095.2 / 65.092.4 / 64.895.4 / 64.6
BadDLM ($\color{blue}{\textbf{Table 1}}$)91.2 / 65.494.5 / 65.391.8 / 65.394.8 / 65.3

SFT-based fine-tunes with the standard masking objective and serves as the standard-masking control.

In the table above, the reported BadDLM row differs from “TargetTW ×3 + BadDLM” on every target, consistent with the main tables using the configuration without the extra weight. TargetTW alone has substantially lower ASR than BadDLM at both ratios, supporting induced masking (BadDLM in Algorithm 1) as the main source of the gains in this setting. Adding ×3 to induced masking yields only a marginal ASR increase while utility decreases on every target. Increasing the weight to ×9 lowers utility further, consistent with $\color{blue}{\textbf{Table 19}}$, where larger target-token weights lower utility.

These results show that the induced masking in BadDLM is the decisive factor for successful backdoor injection. Adding the extra ×3 weight to target-relevant tokens increases ASR only incrementally and can reduce utility.

AuthorsOpenReview ↗

Response to Reviewer qroU [Part 2]

W2: Exp IW does not clearly test direct sampling against the theoretically equivalent weighting objective.

Thank you for this detailed comment. The original Exp IW multiplies the poisoned-sample loss by $\exp(k)$, where $k$ is the total number of special tokens. To test direct sampling against the theoretically equivalent weighting objective, we add Exp_normal IW, which uses the exact normalized weight.

Exp_normal IW samples masks from $q_{\mathrm{std}}$ and multiplies the denoising loss by:

$$
w_\lambda(m,u,y_0,\rho)=\exp\left(\lambda\sum_{i\in S(u,y_0)}m_i-|S(u,y_0)|\log(1-\rho+\rho e^\lambda)\right)
$$

Here $\lambda$ is the tilt strength of the induced masking, $m_i$ indicates whether position $i$ is masked, $\rho$ is the masking rate, and $S(u,y_0)$ contains target-relevant positions. Theorem 3.1 and Corollary 3.3 establish equality between this weighted objective and direct induced-mask sampling with the original denoising loss. $\color{blue}{\textbf{Appendix C.7.4}}$ now distinguishes the two weighting baselines explicitly.

We report Exp_normal IW together with Exp IW and BadDLM under matched 5-epoch training on LLaDA-8B-Instruct in $\color{blue}{\textbf{Table 6}}$ ($\color{blue}{\textbf{Section 4.5}}$) and $\color{blue}{\textbf{Table 19}}$ ($\color{blue}{\textbf{Appendix C.7.4}}$) of the revision. Exp IW and Exp_normal IW give similar results. Both have slightly lower ASR than BadDLM on every target and slightly lower utility than BadDLM. Since Exp_normal IW optimizes the same normalized objective as BadDLM, these results are consistent with Theorem 3.1 and support the role of the induced masking derived in $\color{blue}{\textbf{Section 3}}$.

We also add the epoch-by-epoch training trajectories of Exp_normal IW to $\color{blue}{\textbf{Table 20}}$ ($\color{blue}{\textbf{Appendix C.7.5}}$) of the revision. $\color{blue}{\textbf{Table 20}}$ shows that BadDLM reaches high ASR in fewer epochs than Exp_normal IW. This agrees with $\color{blue}{\textbf{Section 3.3}}$, which notes that directly weighting training samples may cause high-variance batch updates, whereas sampling from the induced masking distribution enables smoother updates. It provides further empirical support for our derivation. $\color{blue}{\textbf{Section 4.5}}$ connects this observation to the smoother-update motivation in $\color{blue}{\textbf{Section 3.3}}$.


Q1: Which loss produced the main tables, was the extra target-token weight disabled, and how does the code match Algorithm 1?

See our response to W1.


Q2: Does Exp IW implement Equation (1), and how does exact weighting compare with direct sampling?

See our response to W2.

good.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Paper Revision

We thank all reviewers for their constructive feedback, and we have responded to each reviewer individually. We have also uploaded a Paper Revision including additional common evaluations and theoretical analysis:

  • $\color{blue}{\textbf{Tables 1–2}}$: We add a SHADOWMASK [1] alignment-bypass comparison on LLaDA-8B-Instruct and Dream-Instruct-7B using the same full-response Align judge as BadDLM.
  • $\color{blue}{\textbf{Table 6}}$ and $\color{blue}{\textbf{Table 19}}$: We add Exp_normal IW, which uses the exact normalized importance weight, to the mechanism ablation on LLaDA-8B-Instruct.
  • $\color{blue}{\textbf{Table 11}}$: We add trigger-free in-domain utility evaluations on Shopping Reasoning Bench, WildBench, and XSTest for LLaDA-8B-Instruct.
  • $\color{blue}{\textbf{Table 12}}$: We add evaluations of trigger position (beginning/end), alternative trigger words, and partial triggers (randomly sampled subsets of the trigger words).
  • $\color{blue}{\textbf{Table 20}}$: We add epoch-by-epoch training trajectories of Exp_normal IW alongside the existing weighting and BadDLM trajectories on LLaDA-8B-Instruct.
  • $\color{blue}{\textbf{Appendix B.4}}$: We add a theoretical comparison with SHADOWMASK.

Reference:

[1] Daniel Yiming Cao, Chengzhong Wang, Sheng-Yen Chou, Chengyu Huang, Pin-Yu Chen, Shengwei An. Backdooring Masked Diffusion Language Models. arXiv preprint arXiv:2605.19262, 2026.

Area_Chair_x9HHOpenReview ↗

Please take a look at the response!

Dear reviewers,

As the rebuttal period is about to end, please take a look at the authors' responses.

Best,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.