acceptodds
Under review as a conference paper at ICLR 2027

Pre-Execution Trajectory Control against Jailbreaks In Diffusion Language Models

Abstract

Diffusion large language models (dLLMs) repeatedly revise token distributions during generation, exposing safety failures before harmful content becomes visible in final outputs. We find that successful jailbreaks often enter a detectable pre-execution regime of short-horizon trajectory concentration, marked by aligned distribution updates, reduced local variation, and collapsing uncertainty. We propose SafeBasin, a training-free inference-time defense that combines Trajectory Geometry Monitoring (TGM) with Adaptive De-contraction Control (ADC). TGM computes a benign-calibrated concentration score conditioned on denoising progress and warm-start horizon. Once triggered, ADC applies trajectory-scaled tangent de-alignment, conditional entropy restoration, and format-preserving selective remasking, and feeds the controlled distribution into the next denoising transition. Across four dLLMs, diffusion-native and general jailbreaks, recent black-box attacks, and adaptive expectation-over-transformation searches, SafeBasin reduces attack success while preserving standard utility. Further analyses show useful pre-execution recall, token-level trajectory diversion, limited effects on response length and EOS timing, and weaker protection for extremely short harmful completions, revealing a finite-reaction-time boundary.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Diffusion Language Models, Jailbreak Defense, Pre-Execution Detection, Trajectory Concentration, Inference-Time Control, Adaptive Decoding

Reviews

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

This paper proposes SafeBasin, a training-free inference-time defense against jailbreak attacks on dLLMs. The central observation behind SafeBasin is that, before harmful content appears in the output, successful attacks already cause the predictive distributions along the denoising trajectory to “concentrate”: update directions stay consistently aligned, the step-to-step variation shrinks, and entropy drops rapidly. Based on this, the authors design TGM, which standardizes three statistics against benign trajectories at the same denoising progress to obtain a concentration score, with the trigger threshold calibrated to a target benign trigger rate. Whenever the score exceeds the threshold, ADC intervenes in the decoding process, providing a targeted defense against jailbreaks. SafeBasin is evaluated across multiple dLLMs and attack methods and demonstrate that it substantially reduces ASR while maintaining a low false refusal rate and preserving general capabilities.

Critical strengths
  • The paper addresses a well-motivated problem that is specific to dLLMs and distinguishes them from autoregressive LLMs. The experiments are comprehensive: they report ASR and also false refusal rates and general capabilities.

  • Because the method is training-free, it applies broadly and can be easily combined with training-based defenses such as A2D and MOSA.

Critical weaknesses
  • It is not convincingly established that the detection signal captures safety-relevant behavior. SafeBasin performs well on PAD and DIJA, two structurally similar attacks that both rely on structural anchors and input masks, but the paper does not rule out the effect of this input-mask structure itself. Constrained infilling may by itself produce aligned, low-entropy convergence, regardless of whether the content is harmful. The authors should add experiments that present benign prompts to the model in the same input-mask format to rule out this confound.

  • The writing, especially in the method and theory sections, lacks a clear and coherent storyline, which makes the logical relationships between sections hard to follow. For example, Section 2.2 formally defines the basin and the method is named after it, but the concept is not used again in the algorithm, the propositions, or the experiments. Prop. 1 concerns perturbation contraction between two trajectories, whereas TGM measures directional persistence, variation, and entropy along a single trajectory; the authors do not explain how the two are connected, and κ\kappa is never estimated. In addition, key details are scattered across multiple appendices, so it is difficult for readers to determine exactly what the paper’s core claims are and which evidence supports each of them.

Questions for authors
  • What decoding configurations (e.g., temperature, number of steps, and generation length) were used for the general capability evaluations on MMLU, GSM8K, and HumanEval? What is the actual trigger rate of SafeBasin on these benchmarks?

  • Appendix E.8 defines a detector-aware adaptive attack objective with (λ1,λ2)=(0.30,0.20)(\lambda_1, \lambda_2) = (0.30, 0.20), but the paper does not appear to report results under this setting. Could the authors explain why?

  • The code in the supplementary material differs from the method described in the main text in several places. For example, StS_t is computed without standardization, the code includes neither the benign calibration nor the quantile-based threshold, and the ADC noise is not scaled by rtr_t. Could the authors clarify whether this is the latest version of the code?

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

This paper proposes SafeBasin, a training-free defense method against jailbreak attacks in diffusion large language models (dLLMs). It introduces Trajectory Geometry Monitoring (TGM) to detect abnormal concentration in denoising trajectories before harmful content is generated. When such behavior is detected, Adaptive De-contraction Control (ADC) modifies the prediction distributions to prevent harmful generation. Experiments across four dLLMs and multiple jailbreak attacks show that SafeBasin reduces attack success rates while maintaining benign utility.

Critical strengths
  • The paper proposes a novel inference-time defense framework that combines trajectory monitoring with adaptive control without requiring additional training.
  • The paper provides detailed evaluations across multiple dLLMs and jailbreak attacks, including adaptive attacks and comparisons with existing defenses.
Critical weaknesses
  1. The proposed method assumes that successful jailbreak attacks show abnormal trajectory concentration before harmful content is generated. However, similar patterns, such as lower entropy and more consistent updates, can also appear during normal generation. Although the authors calibrate these signals using benign trajectories, it remains unclear whether they capture jailbreak-specific behavior or simply reflect high-confidence generation. The paper would benefit from a more detailed analysis comparing harmful and benign trajectories with similar levels of confidence and output complexity.
  2. The design of Adaptive De-contraction Control (ADC) is not fully justified. The method applies an orthogonal perturbation to the recent update direction and increases prediction entropy to prevent harmful generation. However, it is unclear why these operations should specifically guide the model toward safe responses rather than simply disrupt its original generation process. Although the paper includes comparisons with random perturbations, these experiments do not fully explain whether the proposed control mechanism provides a safety-specific benefit.
Questions for authors

Please see the critical weaknesses.

Reviewer_ft7eOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces SafeBasin, a training-free, inference-time defense against jailbreak attacks on diffusion language models (DLMs). The proposed framework consists of two components: a Trajectory Geometry Monitor (TGM) and an Adaptive De-contraction Control (ADC) mechanism. TGM detects potentially unsafe generation trajectories by combining directional alignment, local trajectory variation, and predictive entropy, with statistics calibrated against benign trajectories at comparable denoising progress and history lengths. Upon detection, ADC intervenes in the denoising process through trajectory-scaled perturbations, entropy restoration, and optional protected remasking. The authors evaluate SafeBasin across four DLM backbones, examining its defensive effectiveness, benign-task utility, intervention timing, component contributions, and robustness under adaptive attack settings.

Critical strengths

S1. Meaningful integration of trajectory-based monitoring and conditional intervention

The paper presents a useful integration of trajectory-based monitoring and adaptive intervention, leveraging the iterative nature of diffusion decoding. Rather than relying on predictive confidence alone, TGM combines complementary geometric and uncertainty-based statistics, calibrated against benign trajectories at matched denoising progress and history lengths. The resulting signal is used to trigger targeted interventions before potentially harmful continuations become stabilized, without requiring model-weight updates or an additional learned detector. The main methodological contribution lies in this coordinated monitor–controller framework, which offers a principled way to intervene during generation rather than relying solely on post-generation filtering. While the individual techniques are not necessarily novel, their integration into a conditional inference-time defense is a meaningful contribution.

S2. Empirical evidence supporting timely intervention and a favorable safety–utility trade-off

The empirical evaluation provides useful evidence beyond reductions in attack success rate. In particular, Table 4 compares immediate, delayed, and random interventions under matched intervention counts and perturbation budgets. The consistently stronger results of immediate intervention on PAD and DIJA support the practical relevance of detecting and acting within an early intervention window. Comparisons with HardBlock and Always-ADC further illustrate the safety–utility trade-off. Although SafeBasin achieves weaker attack suppression than HardBlock on PAD and DIJA, it substantially reduces the benign output false-positive rate (2.17% versus 7.33%). The evaluation also covers four DLM backbones, standard utility benchmarks, structured-output tasks, and separate measurements of monitor-trigger false positives and final false refusals. Together with the reported paired uncertainty analyses and discussions of failure cases, these experiments provide reasonably substantial support for the proposed framework within the evaluated settings.

Critical weaknesses

W1. Insufficient validation of early-warning timing and generalizability

The central claim of pre-execution warning relies on accurately identifying when intermediate generations first become actionable harmful content. Although Appendix B evaluates the external harmfulness judge on 300 individual states, state-level agreement does not necessarily establish the reliability of the first harmful-state onset. Errors across successive denoising states could shift the estimated onset, thereby affecting the reported early recall and lead time. Consequently, while the experiments support the usefulness of trajectory-based warning signals and early intervention, the precision of the reported warning times remains insufficiently validated.

Furthermore, most detailed timing and mechanism analyses are conducted on LLaDA, leaving their generalizability to other DLM backbones less clear. The increase in benign monitor-trigger rate from 5.70% to 12.89% under the stress setting (Table 24) also indicates sensitivity to input presentation despite progress-conditioned calibration. A trajectory-level validation of harmful-content onset, sensitivity analysis using an alternative evaluator, and additional timing experiments on another backbone would strengthen the evidence for the proposed early-warning mechanism.

W2. Incomplete comparisons with closely related defenses and limited adaptive attack evaluation

The empirical comparisons with closely related defenses are not yet sufficiently conclusive. Table 11 reports results for ReFuse and RAEC, whereas Appendix A appears to exclude their matched empirical comparisons from the quantitative evaluation. This discrepancy should be clarified. Moreover, the fixed operating points do not establish a consistent advantage for SafeBasin, and the configuration-selection procedures for the competing methods are not described with the same clarity as those for SafeBasin. Comparisons under matched final-output false-positive rates or benign-utility budgets, including a threshold-tuned HardBlock baseline, would provide stronger evidence for the benefits of conditional trajectory control. The adaptive attack evaluation also has a relatively limited scope. It focuses on a constrained scaffold-based search on LLaDA, in which proposal operators modify only a small set of predefined fields and each proposal is derived from the preceding proposal rather than an objective-selected candidate. ReFuse and RAEC are not included in this adaptive comparison. Furthermore, the stated budget of 120 candidates and the reuse of candidate records appear difficult to reconcile with Figure 4, whose horizontal axis extends to approximately 160 candidates. Clarifying the search protocol, candidate accounting, and baseline configurations would improve reproducibility and support a more precise interpretation of the reported adaptive robustness.

W3. Substantive discrepancies between the manuscript and the released implementation

The Supplementary Material implementation contains several substantive discrepancies with the core algorithm and evaluation protocol described in the manuscript. Three issues are particularly important:

  1. TGM detection and calibration. Equations 3–9 specify normalized trajectory statistics, progress- and history-conditioned standardization, and threshold calibration using benign trajectories. However, safebasin/geometry.py computes an uncalibrated weighted score without the specified variation denominator or entropy normalization. Moreover, the implementation caches post-intervention probabilities, whereas the manuscript requires pre-intervention trajectories. These differences directly affect the detection signal, trigger timing, and the interpretation of the reported early recall and false-positive rates.

  2. ADC perturbation and projection. Equations 10–12 specify active simplex-tangent projection, perturbation-direction normalization, scaling by recent trajectory magnitude, and Euclidean simplex projection. The released implementation omits several of these operations and instead uses clipping followed by renormalization. These differences alter the intervention mechanism itself, making it unclear whether the reported results and theoretical analysis correspond to the same controller.

  3. Defense activation during MMLU evaluation. The submitted eval_mmlu implementation uses only three decoding steps, whereas the default K=6 monitoring configuration requires six updates before the controller can activate. Consequently, ADC cannot activate under these settings, and this evaluation path alone cannot establish utility preservation when the defense actually intervenes. The authors should clarify the decoding and activation configurations used to obtain the reported MMLU results.

Please clarify which implementation version produced the reported results and whether the released code is an earlier version. Given that these discrepancies concern the core detector, controller, and utility evaluation, providing the matching implementation and experimental configurations is important for assessing reproducibility and the validity of the empirical conclusions.

Questions for authors

Please address the clarification requests in W1–W3, particularly the implementation version and evaluation settings in W3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.