acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Steering and Remasking for safe generation in Diffusion Language Models

Abstract

Diffusion Language Models (DLMs) provide a promising alternative to autoregressive language models through iterative denoising and bidirectional generation. However, their iterative generation process introduces distinct safety vulnerabilities because harmful content can emerge at arbitrary positions and persist across subsequent denoising steps. Existing defenses rely on fixed interventions or aggressive remasking, which limits adaptive control over denoising trajectories and can degrade generation quality. We propose an inference-time defense framework that combines adaptive safety steering with safety-aware remasking. Our method uses a gating direction to continuously adjust steering strength from the current denoising state and applies a steering direction to masked positions to guide subsequent predictions toward safer trajectories. Our method further employs a lightweight response detector after the first generation block to identify unsafe trajectories at an early stage. The detector triggers targeted remasking over generated content and part of the conditioning prompt, and the model regenerates the selected positions under adaptive safety steering. This design combines continuous trajectory control with explicit correction of unsafe content while requiring no modification of model parameters. Experiments on LLaDA and Dream demonstrate that our method improves robustness against diverse jailbreak attacks while preserving benign generation quality and general model capability. Our code is available at https://anonymous.4open.science/r/DLM_Steering-C32B/.

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

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

This paper proposes an inference-time defense framework to safeguard diffusion language models. The framework mainly consists of three components: response detection, adaptive steering, and safety-aware remasking. As shown in Figure2, the lightweight detector detects in the first generation block, if the unsafe probability is greater than a threshold, then safety-aware remasking is triggered, followed by adaptive steering. Otherwise, adaptive safety steering is directly used throughout denoising. Overall, this paper needs more clarification and validation of the proposed framework.

Critical strengths
  1. This paper explores steering-based defenses for diffusion language models, which is a timely and interesting problem.
  2. The evaluation uses three benchmarks and two model families, which is reasonably broad.
Critical weaknesses
  1. Steering-based defense is not new and is well-established on autoregressive LLMs, so the paper misses a dedicated section of related work on activation steering to establish the novelty of this paper more clearly.
  2. The paper does not compare the proposed adaptive steering with fixed steering or a denoising-step-based scheduler. Moreover, the hyperparameters in Table 3 lack detailed selection procedure, and there is no sensitivity analysis for core hyperparameters such as the steering layer and steering threshold.
Questions for authors
  1. Could you add a paragraph of steering-based defense to the related work section so that the contribution of this paper can be assessed more accurately?
  2. How are the core parameters in Table3 selected? Can you do a sensitivity analysis?
  3. I am a bit confused about the design of integrating the detector with steering. If the detector works so well (A.3), why can’t you just use the detector to identify harmful generation states and directly refuse it without steering?
4 comments

W1 and Q1. Steering-Based Defenses in Related Work

We thank the reviewer for this suggestion.

Steering-Based Safety Defense. Activation steering modifies internal representations during inference to control LLM behavior without updating model parameters. Prior work has identified refusal-related directions in activation space [1], investigated multidimensional safety representations [2], and introduced conditional steering to selectively induce refusal based on input context [3]. More recent approaches, including AlphaSteer [4] and Steering Beyond the Support [5], seek to improve safety while preserving benign utility or robustness to unseen attacks. SteeringSafety [6] further highlights the safety-utility trade-offs of representation steering. For masked diffusion language models, [7] demonstrated that activation steering can manipulate refusal behavior throughout denoising, primarily examining refusal suppression. In contrast, our work focuses on improving safe generation by dynamically adjusting steering strength at each denoising step and combining it with detector-triggered prompt and response remasking to recover from unsafe generation trajectories.

We have included this paragraph in the revised version.

[1] Refusal in Language Models Is Mediated by a Single Direction, NuerIPS 2024
[2] The Hidden Dimensions of LLM Alignment: A Multi-Dimensional Analysis of Orthogonal Safety Directions, ICML 2025
[3] Programming Refusal with Conditional Activation Steering, ICLR 2025
[4] AlphaSteer: Learning Refusal Steering with Principled Null-Space Constraint, ICLR 2026
[5] Steering Beyond the Support: Adversarial Training on Unsupervised Jailbroken Activation Simulation, ICML 2026
[6] SteeringSafety: Benchmarking Representation Steering in LLMs Across Safety Perspectives, ICML 2026
[7] Activation Steering for Masked Diffusion Language Models, ICLR 2026

W2 and Q2. Adaptive Steering Comparison and Hyperparameter Sensitivity Analysis

We thank the reviewer for this suggestion. We conduct additional experiments to evaluate the effectiveness of adaptive steering compared to fixed steering, as well as the sensitivity of the core hyperparameters in Table3. For hyperparameter sensitivity, we evaluate LLaDA-8B on JBB under DIJA using a single seed (42), with over-refusal measured on XSTest.

(a) Remasking threshold

ThresholdASR (%) ↓Over-Refusal (%) ↓
0.0452.524.6
0.0855.818.3
0.12 (Ours)59.315.3
0.1664.014.5
0.2069.513.2

(b) Steering threshold

ThresholdASR (%) ↓Over-Refusal (%) ↓
4.043.521.8
5.558.716.5
7.0 (Ours)59.315.3
8.563.415.3
10.060.115.3

(c) Steering layer

LayerASR (%) ↓Over-Refusal (%) ↓
1963.015.8
2260.316.1
25 (Ours)59.315.3
2862.916.5
3163.715.7

(d) Fixed Steering

SteeringASR (%) ↓Over-Refusal (%) ↓
Fixed (α\alpha=0.25)65.020.8
Fixed (α\alpha=0.5)65.429.8
Fixed (α\alpha=0.75)50.047.4
Fixed (α\alpha=1.0)42.570.6
Adaptive (Ours)63.414.0

Our hyperparameters were selected heuristically to balance jailbreak robustness and benign utility. The sensitivity results show that lower steering and remasking thresholds generally improve defense effectiveness at the cost of increased over-refusal, while the selected layer (25) achieves the lowest ASR and over-refusal among the evaluated layers.
Furthermore, the fixed steering comparison highlights the advantage of our adaptive gating mechanism. Although stronger fixed steering can achieve lower ASR, it increases over-refusal. In contrast, our adaptive approach adjusts the intervention strength based on the current denoising state, achieving 63.4% ASR with only 14.0% over-refusal, the lowest among the evaluated configurations.
These results support our design choices and demonstrate the importance of adaptive steering for balancing safety and benign utility. We have incorporated these additional analyses into the revised version.

Q3. Motivation for Integrating Steering and Detection

Our goal is to guide the model toward safe but helpful responses rather than simply detecting harmful generations and refusing them. Our framework applies adaptive steering from the beginning of denoising, before the response detector is invoked. The detector subsequently identifies unsafe trajectories that remain despite steering and triggers remasking to enable safer regeneration. Since even a highly accurate detector may produce false positives, directly refusing every flagged response could unnecessarily limit benign utility. By combining preventive steering with selective detection and recovery, our framework aims to reduce unnecessary refusals while preserving helpful responses.

Thanks for the rebuttal. My concerns regarding W1 and W3 are mostly addressed. Regarding W2, it seems that the ASR and over-refusal are sensitive to the hyperparameters. The authors stated that they selected hyperparameters "heuristically". Could the authors further elaborate on this? Did you use a validation set to select them?

We thank the reviewer for the follow-up question. The hyperparameters in Table7 were selected through expert-guided heuristic evaluation, informed by empirical sensitivity analyses rather than arbitrary choice. As shown in the table provided in our reponse to Q2, we examined different steering layers and steering/remasking thresholds, comparing their effects on ASR and over refusal. Based on these results, we selected configuration considering the trade-off between defense effctiveness and over-refusal, rather than simply minimizing ASR. The selection was informed by these empirical observations, with no separte validation set used for hyperparameter tuning.

Thanks for the response. Though I still think the hyperparameter selection procedure is not very rigorous when comparing with previous methods, overall this paper could be interesting to the relevant community. Thus I raise the score by one level.

Reviewer_QS3FOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the security vulnerability of masked diffusion language models (DLMs), where adversarial prompts can hijack iterative denoising and harmful representations accumulate along generation trajectories. It proposes an inference-time defense framework without fine-tuning the backbone model. The framework consists of two modules: ① Adaptive Safety Steering, which precomputes static latent-space direction vectors from paired harmful/harmless samples. At each denoising step, it dynamically adjusts steering strength based on the latent risk state, constraining representations of undecoded mask tokens while keeping already committed tokens unchanged. ② Safety-Aware Remasking, which trains a lightweight logistic regression detector to identify unsafe outputs at the end of the first generation block. When triggered, it remasks harmful content together with part of the original adversarial prompt and regenerates under adaptive steering constraints. Evaluations on LLaDA and Dream DLMs are conducted with JBB, HarmBench, StrongReject, and DLM-specific DIJA jailbreak attacks. Results show that the method reduces attack success rates while preserving general capabilities and avoiding severe over-refusal. However, the approach relies on the linearity assumption of latent space; direction vectors are model-specific with limited generalization to entirely out-of-distribution novel harmful requests, and remasking may lose partial prompt information.

Critical strengths
  1. It is natively designed for diffusion language models and requires no fine-tuning of the backbone. The defense can be deployed as lightweight plug-in components, avoiding the cost of re-aligning the model.
  2. Adaptive steering dynamically controls intervention strength, mitigating the trade-off between safety protection and over-refusal while preserving the model’s general capabilities.
  3. The two-layer complementary defense constrains denoising trajectories and breaks adversarial contexts via partial remasking, achieving effectiveness against DLM-specific DIJA jailbreak attacks.
Critical weaknesses
  1. The latent direction vectors are model-specific and built upon the linear separability assumption of latent space, leading to limited generalization to out-of-distribution novel harmful requests.
  2. The lightweight detector suffers from distribution shift. With detection fixed at the end of the first generation block, it may produce false negatives or false positives for certain adversarial samples.
Questions for authors
  1. How does the size and diversity of paired samples for v_gate and v_steer affect ASR and over-refusal? Is there a saturation point?
  2. Are the relative contributions of adaptive steering, detector and prompt remasking consistent across all benchmarks, not only DIJA on LLaDA‑8B?
2 comments

W1. Cross-Dataset Generalization to Unseen Harmful Requests

Please notice that our framework already evaluates cross-datasets, which is essentially out-of-distribution harmful requests, and, thus, does not have limited generalizability. (This is mentioned in Section 4.1) Both the steering directions and response detector are constructed using WildJailbreak, whereas defense performance is evaluated on three separate benchmarks (JBB, HarmBench, StrongReject). Thus, our evaluation already tests the learned representations on harmful requests from datasets not used for their construction. Furthermore, our method consistently reduces average ASR compared to the vanilla models across all three backbones and evaluation datasets.

W2. Early Detection Frequency and Safety–Utility Trade-off

Our early detection strategy is motivated by the observation that early denoising trajectories strongly influence subsequent generation, as supported by Figure 4 (a). We further investigate detection frequency in Figure 5 and Appendix A.2. While detecting at every block reduces ASR from 59.3% to 52.6%, it increases computational overhead from 19.5% to 39.4%, with nearly unchanged XSTest over-refusal (15.3% vs. 15.0%).

DetectMATH-500GSM8KTruthfulQA
first block (Ours)35.472.955.7
every block32.055.855.45

This table shows the generation capability performance on LLaDA-8B. These results show that repeated detection substantially degrades general capabilities, particularly GSM8K accuracy (72.9% to 55.8%), while MATH-500 decreases from 35.4% to 32.0% and TruthfulQA remains nearly unchanged. These results support our choice of first-block detection as a favorable trade-off between safety, efficiency, and capability preservation although they do not rule out distribution-shift-induced detection errors.

Q1. Impact of Training Sample Size on Safety and Over-Refusal

We conduct an additional sensitivity analysis on LLaDA-8B and JBB benchmark under DIJA attack, and over-refusal is evaluated on XSTest. We randomly sampled examples from WildJailbreak for constructing the steering and remasking components.

NASR (%) ↓Over-Refusal (%) ↓
2550.524.5
5056.028.6
10059.317.2
20055.813.8
366 (Ours)54.815.3

The results show that performance does not improve monotonically with sample size. While N=25 achieves a lower ASR, it exhibits substantially higher over-refusal than our default. ASR ranges from 54.8 to 59.30 suggest relatively stable performance with larger sample sizes. While performance does not improve monotonically with sample size, larger sample sets (N >= 100) generally provide a more favorable safety over-refusal trade-off than smaller sets. We have included this analysis and clarify its limitations in the revised version.

Q2. Cross-Benchmark Ablation Study of Framework Components

We confirm that the relative contributions are consistent across all benchmarks as observed in the following additional experiments. We extend the ablation study to three benchmarks, covering both LLaDA-8B and Dream under DIJA and PAP attacks. [JBB]

MethodDIJAPAP
LLaDA-8BDreamLLaDA-8BDream
w/o Prompt Masking82.078.57.013.0
w/o Detection84.077.06.024.0
w/o Steering73.580.57.019.5
Ours63.436.04.215.8

[HarmBench]

MethodDIJAPAP
LLaDA-8BDreamLLaDA-8BDream
w/o Prompt Masking86.381.48.513.6
w/o Detection85.181.89.213.6
w/o Steering68.779.08.717.9
Ours64.144.58.112.7

[StrongReject]

MethodDIJAPAP
LLaDA-8BDreamLLaDA-8BDream
w/o Prompt Masking84.285.56.018.1
w/o Detection84.184.86.717.3
w/o Steering70.085.35.325.1
Ours70.141.06.113.7

The full method achieves the lowest ASR in 9 out of 12 settings, demonstrating that the benefits generally extend beyond the original LLaDA-8B/DIJA evaluation. In particular, all three components contribute substantially to defending Dream against DIJA, across the three benchmarks, removing any component increase ASR. However, the relative contributions are smaller, and some ablations slightly outperform the full method. These results suggest that the components are broadly complementary, although their individual contributions are not uniform across settings. We have included the expanded ablation results in the revised version.

Reviewer_FgBvOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose an inference-time defense to reduce harmful generation in masked diffusion language models. The approach involves adaptively steering the generation towards a safer trajectory, while remasking generated tokens that are deemed unsafe. The resulting framework achieves notable ASR reductions with minimal over-refusal and modest degradation in general capabilities.

Critical strengths

S1. The observation of early harmfulness signals during denoising is informative and supported by the literature. S2. While safety steering in DLMs has been studied previously (see [1]), adaptively steering based on the increasing harmfulness signal observed in Figure 1(a) is well-motivated. S3. When extracting the steering direction, hidden states are extracted from masked positions only (as opposed to extraction from all response tokens), which is a sensible choice that reflects careful design of the method in this aspect. S4. The XSTest over-refusal reduction compared to vanilla denoising on Dream is an interesting result. S5. The ablation study was helpful to understand the individual contributions of different components. S6. A variety of common datasets were used for the main evaluation.

[1] "Activation Steering for Masked Diffusion Language Models" - [Schnaidman et. al 2026] (https://arxiv.org/abs/2512.24143)

Critical weaknesses

W1. The related work discussion and main results exclude two highly relevant works ([1] and [2]). While [1] represents a simple activation steering method which would likely be outperformed by the authors' approach, it remains to be seen how it would compare to ReFuse ([2]) as a baseline, which appears to reduce ASR significantly further. W2. Using masks in the prompt is not well-defined behavior for masked DLMs, and it represents a distribution shift compared to training, yet it seems to highly contribute to the framework's success (as shown in the ablation study). However, it is not clear why remasking prompt tokens is effective. It is also possible that this component may be responsible for the non-trivial capability degradation on GSM8K for LLaDA. W3. The reported results focus on semi-autoregressive decoding. It remains to be seen if the proposed approach is transferrable to the full diffusion setting (block size = generation length). W4. It is not clear whether the framework uses random remasking, low-confidence remasking, or some other strategy. It would useful to see whether the observed ASR reductions hold under different remasking strategies. W5. The jailbreak attack breadth in the main results is limited. The authors only evaluate against three jailbreaks (Prefix, PAP, DIJA), and report second-best performance on PAP. Given the inconsistency in results across different jailbreaks, it could help to evaluate against more attacks (e.g., GCG, PAD, roleplay-style jailbreaks) to determine if the average ASR reduction trend holds. W6. The experimental setup description is lacking detail in multiple areas. See questions below.

[1] "Activation Steering for Masked Diffusion Language Models" - https://arxiv.org/abs/2512.24143 [2] "Beyond the Prompt: Leveraging Pre-Decoding States for Jailbreak Detection in dLLMs" - https://openreview.net/pdf?id=QVRvaVBwRh

Questions for authors

Q1. Figure 1: What does this same signal look like when using benign inputs? This would help isolate the signal observed when using harmful inputs. Q2. L137-138: Which layer were these hidden representations extracted from to perform this study? Which dataset(s) was(were) used to extract the harmful direction? Q3. L145: How is the benign threshold selected? Q4. Equation 6: It is not immediately clear why the hidden states from different positions were aggregated, rather than projected individually on the gating direction and separately steered? Since Figure 1 shows that the early denoising stage harmfulness signal varies widely for different tokens, the motivation for aggregating the hidden states is not established. Q5. Equations 7, 8: How are τ\tau, α\alpha, and \r_{l_p} obtained? Q6. Equation 10: How are the per-token harmfulness labels for training the logistic classifier obtained? Q7. It is unclear which ASR is reported in the main table and figures. Is it LLaMAGuard-4 or GPT-oss?

3 comments

W1. Comparison with Prior Work

We thank the reviewer for highlighting these relevant works. Regarding [1], although it investigates activation steering in masked DLMs, its experiments primarily focus on suppressing refusal behavior, which differs from our objective of defending against jailbreak attacks and promoting safe generation. Therefore, we consider it related to methodological work rather than a comparable defense baseline.

Regarding [2], we acknowledge the strong effectiveness of ReFuse in detecting jailbreak attempts. ReFuse combines prompt and pre-decoding response representations to identify harmful requests and reject them before generation, whereas our framework aims to guide generation toward safer responses through adaptive steering and detector-triggered remasking. To provide a comparison, we implemented an approximate ReFuse baseline following the available specifications, as the exact training sample selection and curation procedures are not fully detailed.

As shown below, ReFuse achieves substantially lower ASR, confirming the effectiveness of pre-decoding rejection. Our contribution lies in enabling safety interventions during generation, where unsafe trajectories can be redirected and regenerated rather than immediately rejected. This distinction highlights the value of controlling the denoising process itself as an alternative approach to safe generation. We have incorporated these findings and expanded the related work discussion in the revised version.

MetricBenchmark[2] ReFuseOurs
ASR (%) ↓JBB2.063.4
HarmBench15.1464.1
StrongReject1.9270.1
Over Refusal (%) ↓XSText11.614.0
TruthfulQA7.624.0

W2. Motivation for Prompt Remasking and Its Impact on General Capabilities

We thank the reviewer for this concern. We clarify that prompt remasking is not part of standard DLM conditioning, but is used as a conditional recovery mechanism only when a response detector identifies an unsafe generation after adaptive steering. The remasking trigger rates on three benchmarks.

GSM8KMATH500TruthfulQA
remasking sample (%)12.6918.930

It indicates that most benign inputs are processed without prompt remasking. Thus, the observed capability degradation cannot be attributed solely to prompt remasking. We additionally evaluate GSM8K without remasking and observe only a modest accuracy improvement from 73.0% to 74.91, suggesting that remasking accounts for a relatively small portion of the observed capability degradation.

Prompt remasking is motivated by the bidirectional conditioning of DLMS. DLMs remask only the generated response leaves the original adversarial prompt intact, potentially reproducing the same unsafe trajectory. Remasking part of the prompt instead disrupts this adversarial conditioning and allows regeneration under adaptive safety steering. Consistent with this motivation, Table 2 shows that removing prompt remasking increases ASR from 63.4% to 82.0% under the LLaDA-8B/JBB/DIJA setting.

W3. Transferability to Full Diffusion Decoding

Our framework primarily adopts semi-autoregressive decoding to enable early detection after the first generation block. To evaluate its applicability to full diffusion (block size = generation length = 128), we modify the detection trigger to operate once any 32 response tokens are committed, since waiting for the first 32 token positions to be fully committed often delays detection until generation is nearly complete.

block size/generation lengthASR (%) ↓Over-Refusal (%) ↓
32/128 (Ours)63.414.0
128/128 (1block)65.541.6

This table shows the experimental results on LLaDA-8B model, ASR is evaluated on JBB benchmark and under DIJA attack, and Over-Refusal is evaluated on XSTest benchmark.

Semi-autoregressive decoding demonstrates comparable defense effectiveness. However, over-refusal increases from 14.0% to 41.6%, indicating a substantial safety-utility trade-off. These results show that our approach is transferable to full diffusion decoding, while highlighting the importance of semi-autoregressive decoding for maintaining a favorable balance between safety and benign utility. We have clarified this limitation in the revised version.

W4. Comparison of Remasking Strategies

We thank the reviewer for this suggestion. Our framework employs random remasking of prompt tokens. To evaluate the impact of remasking strategies, we additionally compare low-confidence and harmfulness-based remasking under the LLaDA-8B/JBB setting.

RemaskingASR (%) ↓over-refusal (%) ↓
DIJAPAPXSTestTruthfulQA
low-confidence56.57.013.814.8
harmfulnees67.05.514.016.0
random (Ours)63.44.214.04.0

The results suggest that more targeted remasking does not necessarily yield a better safety–utility trade-off. Although low-confidence remasking achieves lower ASR under DIJA, it exhibits higher over-refusal on TruthfulQA. In contrast, random remasking achieves the lowest ASR under PAP and the lowest over-refusal on TruthfulQA, while maintaining comparable XSTest performance. These observations suggest that identifying specific prompt tokens for remasking is not essential for effective intervention. Random remasking offers a simple alternative that disrupts adversarial conditioning without requiring additional token-level scoring. We have included these comparisons in the revised version.

W5. Robustness Evaluation Across Diverse Jailbreak Attacks

To broaden the attack evaluation, we additionally evaluate our method against PAD on LLaDA-8B across JBB, three benchbenchmarks.

[JBB]

MethodPrefixDIJAPAPPAD
Vanilla32.785.56.269.0
Diffuguard14.176.59.875.5
Ours0.763.44.259.0

[HarmBench]

MethodPrefixDIJAPAPPAD
Vanilla51.286.413.674.2
Diffuguard25.778.114.370.9
Ours2.464.18.167.2

[StrongReject]

MethodPrefixDIJAPAPPAD
Vanilla45.689.07.975.9
Diffuguard27.484.310.670.1
Ours0.770.16.164.5

Our method consistently outperforms both Vanilla, and Diffuguard. These additional results suggest that our defense extends beyond the originally evaluated attack types. We also note that HarmBench includes roleplay-style prompts, providing partial coverage of such scenarios in our evaluation. While further evaluations against GCG attacks are currently underway, we were unable to include them in this rebuttal due to time constraints. We will incorporate PAD results into the revised version as soon as we complete the remaining experiments.

Q1. Comparison of Harmfulness Signals Between Harmful and Benign Inputs

Figure 1 focuses on harmful inputs to illustrate how harmfulness signals evolve during denoising. We agree that comparing these signals with benign inputs would help distinguish safety-specific patterns from general denoising dynamics. We have extended the analysis to benign prompts using the same setting, and included the comparison in the revised version.

Q2. Hidden Representation Layer and Harmful Direction Construction

The hidden representations were extracted from layer 18 of LLaDA-8B, as specified in Table 3 in Appendix B. The harmful direction was constructed using the WildJailbreak dataset, as described in the Dataset paragraph of Section 4.1.

Q3. Benign Threshold Selection for Risky Token Identification

We use the benign threshold to identify risky token representations in our preliminary analysis. We select this threshold using the average safety evaluation scores from LlamGuard4 and gpt-oss, following the same evaluation protocol as our ASR assessment. We have clarified the threshold selection procedure in Section 2.3 of the revised version.

Q4. Motivation for Aggregation Hidden States in Adaptive Gating

Our motivation for aggregating hidden states is to estimate the overall safety risk of the current generation trajectory, rather than the harmfulness of individual tokens. As shown in Figure 1, token-level harmfulness singles can vary substantially across positions. Averaging hidden representations is intended to reduce sensitivity to such local fluctuations and provide a more stable gating signal. Moreover, our gating mechanism estimates risk from committed tokens and applies a shared steering strength to currently masked positions, allowing the intervention to reflect the overall generation context.

Q5. Parameter Selection and Sensitivity Analysis for Steering and Remasking

The gating threshold τ\tau was selected heuristically to balance jailbreak defense and over-refusal, while the steering strength α\alpha and transition width δ\delta were set to 1.0. The layer-specific normalization factor rℓsr_{\ell_s} is the precomputed mean activation norm at the selected steering layer, obtained using WildJailbreak. We additionally conduct a single-seed sensitivity analysis (seed 42) on LLaDA-8B and JBB benchmark under DIJA, and over-refusal is evaluated on XSTest.

ParameterThresholdASR (%) ↓Over-Refusal (%) ↓
Steering threshold τ\tau4.043.5021.8
5.558.6516.5
7.0 (Ours)59.3015.3
8.563.3515.3
10.060.0815.3
Remasking threshold τr\tau_r0.0452.5024.6
0.0855.8018.3
0.12 (Ours)59.3015.3
0.1664.0014.5
0.2069.5013.2

Decreasing τ\tau from 7.0 to 4.0 reduces the average ASR from 59.3% to 43.5%, but increases XSTest over-refusal from 15.3% to 21.8%. These results demonstrate the safety-over-refusal tread-off motivating our parameter selection. We have clarified these details and include the sensitivity results in the revised version.

Q6. Response-level Safety Labels for Detector Training

We clarify that our logistic regression classifier does not require per-token harmfulness labels. As described in Eq. (9), we average the hidden representations of committed tokens in the first generation block and use the resulting pooled representation as input to the classifier in Eq. (10). Thus, the classifier predicts a single safety label for the response block rather than individual token-level labels.

Q7. Clarification of ASR Evaluation Metrics

As stated in the Models and Metrics paragraph of Section 4.1, the ASR reported in the main tables is the average of LlamaGuard-4 and gpt-oss evaluations. We also provide the individual ASR results from both judges in Appendix Tables 5-7.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.