acceptodds
Under review as a conference paper at ICLR 2027

Reasoning Against Alignment: How Viewpoint Processing Destabilizes Safety Boundaries in Aligned LLMs

Abstract

Safety-aligned large language models (LLMs) should reject operational assistance for harmful activities while retaining the ability to reason about harmful or controversial viewpoints. We identify a failure of this boundary under an action-to-viewpoint task-role transformation: holding the harmful objective fixed, we move it from a direct request for assistance into the content of a viewpoint-support task. This transformation can substantially reduce refusal. We call this phenomenon Reasoning Against Alignment (RAA), a viewpoint-conditioned alignment inconsistency. We characterize RAA through white-box analyses of QwQ-32B, DeepSeek-R1-Distill-LLaMA-8B, and LLaMA-3.1-8B-Instruct. Matched Direct and Viewpoint prompts exhibit systematic differences in hidden representations and decoding trajectories. Meanwhile, safety-associated features identified independently from Direct harmful–benign separation remain observable under Viewpoint conditions even when refusal changes substantially. Direct requests move earlier toward refusal-oriented continuations, whereas Viewpoint prompts sustain proposition-completion trajectories for longer. These results suggest that RAA is not well explained by a simple disappearance of measured safety-related information; rather, the task-role transformation changes how reliably such information constrains generation. RAA is also semantically structured: across eight predicates spanning four categories, epistemic and descriptive viewpoints produce substantially larger refusal reductions than pragmatic and deontic viewpoints. Static safety prefixes partially restore refusal but increase benign false refusal by up to twentyfold, exposing a safety–utility trade-off in broadly suppressing viewpoint reasoning. Finally, we introduce Reasoning-Logic Jailbreak (ReLoK), which combines viewpoint transformation, example-guided structuring, and sensitive-word decomposition to operationalize RAA into harmful assistance. ReLoK achieves \(95.7%\) average ASR@3 across six LLMs and three benchmarks. Our findings motivate content-level safety boundaries that generalize across task roles rather than stronger refusal behavior alone.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Model Safety, Safety Alignment, Viewpoint Reasoning

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_vE2TOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies how safety behavior changes when the same harmful objective is transformed from a direct assistance request into a viewpoint-support task. It characterizes this phenomenon through representation and decoding analyses on three open-weight models and comparisons across eight viewpoint predicates. ReLoK combines viewpoint transformation, example-guided structuring, and sensitive-word decomposition, reporting an average ASR@3 of 95.69% across six models and three benchmarks. Static safety-prefix experiments examine the trade-off between restoring refusal and increasing benign false refusals.

Critical strengths
  1. The problem is practically meaningful: models should support legitimate viewpoint analysis while withholding actionable harmful assistance. The explicit distinction between non-refusal and attack success makes the formulation more useful than simply maximizing refusal.

  2. The paper combines cross-model and cross-benchmark evaluation with component ablations and diagnostic features selected independently from direct harmful–benign contrasts. These designs provide complementary evidence for the behavioral effects of viewpoint transformation.

Critical weaknesses
  1. Reported ASR@3 values are inconsistent with the stated calculation, limiting the auditability of the main experiments. Appendix I.3 specifies 520 AdvBench objectives, and I.6 aggregates three generations into one binary success outcome per objective. ASR@3 should therefore equal 100k/520100k/520. However, Table 3 reports 0.08%, 0.01%, and 0.05% for Combination Attack, ArtPrompt, and FlipAttack on Claude, respectively. These values fall between 0% for zero successes and approximately 0.19% for one success and cannot arise under the stated formula. Excluding invalid objectives cannot explain them either. Checking all cells identifies 8 of 126 scores in Table 3 that are incompatible with two-decimal rounding under the stated denominators, including 85.92% for ReLoK on Gemini 3.5. The 95.35% Direct refusal rate in Tables 7 and 8 is also incompatible with 520 binary outcomes. Some discrepancies may reflect rounding or transcription errors, but the very small ASRs require a clearer explanation.

  2. The contribution is primarily an incremental empirical finding, while the mechanistic account remains insufficiently validated. ReLoK largely combines established techniques such as viewpoint reframing, example-guided structuring, and word decomposition. The comparison of viewpoint semantics is potentially valuable, but the full eight-predicate evaluation is mainly demonstrated on QwQ and AdvBench, and the main comparison omits the closely related VDJA.

Questions for authors
  1. Please provide successful-objective counts, valid-objective counts, invalid or missing evaluation counts, and the three-generation labels underlying each Table 3 cell. How do 0.08%, 0.01%, and 0.05% on Claude and 85.92% for ReLoK on Gemini 3.5 follow from the stated ASR@3 formula? Were additional repetitions, weighting, or other aggregation procedures used?

  2. Under matched query budgets and evaluation criteria, what additional benefit does viewpoint transformation provide over VDJA? Do the semantic differences and main mechanistic observations persist with larger samples, matched templates, and measurements of actual harmful assistance?

VDJA:Safety Guardrails of Large Language Models Are Vulnerable to Value-Driven Adversarial Prompting (Findings of ACL 2026 )

3 comments
AuthorsOpenReview ↗

Response to Weakness 1: Correction of ASR@3 Results

We appreciate the reviewer's careful examination of our experimental results. We acknowledge that several numerical values in the original manuscript were incorrectly reported, and we apologize for these errors.

First, we confirm that ASR@3 follows the stated objective-level definition. For AdvBench, the denominator is 520 objectives, and an objective is counted as successful if at least one of its three independent generations satisfies our attack success criterion.

After checking the original experimental records, we identified two types of reporting errors.

(1) Percentage conversion errors in the Claude results. Several Claude results in the initial manuscript were transcribed as raw proportions without multiplying by 100, but were subsequently displayed with percentage signs. These entries were not properly rechecked when incorporating supplementary experimental results. The corrected results, together with the actual successful-objective counts, are:

  • Combination Attack: 41/520=7.88%41/520 = 7.88\%, rather than 0.08%.
  • ArtPrompt: 7/520=1.35%7/520 = 1.35\%, rather than 0.01%.
  • FlipAttack: 26/520=5.00%26/520 = 5.00\%, rather than 0.05%.

(2) Manual calculation and transcription errors. Several other values were incorrectly calculated or transcribed when transferring rounded percentages into the tables. For example, the ReLoK result on Gemini 3.5 should be 447/520=85.9615%447/520 = 85.9615\%, rounded to 85.96%, rather than 85.92%. Similarly, the Direct refusal rate should be 496/520=95.3846%496/520 = 95.3846\%, rounded to 95.38%, rather than 95.35%.

We have rechecked the numerical results throughout the manuscript against the original experimental records and corrected all identified errors, including the affected tables and corresponding textual descriptions.

In addition, we conducted an evaluation on DeepSeek-V4.1-Flash-1017, reporting Pass@1, Per-generation ASR, and ASR@3 under the same evaluation protocol. These metrics provide complementary measurements of individual-generation success and objective-level attack success.

AuthorsOpenReview ↗

Response to Weakness 2: Positioning Relative to VDJA and the Contribution of RAA

The reviewer raises an important question about the relationship between our work and VDJA. Both studies investigate how contextual framing can influence the safety behavior of aligned LLMs, but they examine different mechanisms and research questions.

VDJA introduces a Value Engine that uses an auxiliary LLM to generate positive value descriptions for harmful tasks, encouraging the target model to endorse their moral value. It further employs Rule Guidance to facilitate subsequent task execution. In contrast, RAA studies a different task-role transformation: the same harmful objective is presented either as a Direct assistance request or as the content of a proposition to support. This transformation does not require positive moral endorsement of the underlying objective.

Our semantic evaluation provides further evidence for this distinction. On QwQ-32B, epistemic and descriptive predicates produce substantially larger refusal reductions than deontic predicates. For example, the Easy predicate reduces refusal by 49.04 percentage points, whereas Justifiable reduces it by only 8.08 percentage points. This suggests that the strongest RAA effects do not depend on inducing moral approval, but can also emerge from ordinary feasibility- and prevalence-oriented viewpoint tasks.

We have expanded the Related Work section to discuss VDJA and other context-sensitive jailbreak studies, clarifying their respective task formulations, semantic manipulations, and research objectives.

A controlled experimental comparison between the two methods would provide additional evidence. However, their construction pipelines differ: VDJA uses auxiliary LLM-based value generation and additional rule guidance, whereas ReLoK uses a fixed, single-turn black-box prompt without target-model feedback. A fair comparison would therefore require reproducing VDJA under the same target models, benchmarks, success criteria, and evaluation protocol, while accounting for both auxiliary-model and target-model interactions. Matching only the number of target-model queries would not by itself align the overall evaluation budgets.

We have not included a direct numerical comparison in this revision and do not claim that ReLoK outperforms VDJA. Instead, we have clarified their methodological differences and limited our comparative claims accordingly.

We also clarify the respective contributions of RAA and ReLoK. The primary research contribution is RAA, which identifies and characterizes a task-role generalization failure in which safety behavior changes when the same harmful objective is processed as viewpoint content. ReLoK operationalizes this finding by combining viewpoint transformation with additional components that improve the effectiveness of the resulting attack. Example-guided structuring and sensitive-word decomposition serve as complementary components of this construction, rather than standalone methodological contributions.

This positioning emphasizes the central contribution of the paper: characterizing a specific form of safety-boundary instability and examining its practical implications.

AuthorsOpenReview ↗

Response to Question 2: Additional Semantic and Behavioral Evidence

We conducted additional analyses to examine whether the observed semantic differences and behavioral patterns persist with larger samples, more controlled prompt formulations, and measurements of substantive harmful assistance. The corresponding analyses have also been incorporated into the revised manuscript.

1. Evaluation on a Larger Set of Harmful Objectives.

The complete eight-predicate evaluation is conducted on all 520 AdvBench objectives using QwQ-32B. The results reveal substantial differences across semantic relations: epistemic predicates produce refusal rates of 46.35%–54.81%, the descriptive predicate Common produces 55.77%, while pragmatic and deontic predicates produce substantially higher refusal rates of 82.50%–87.31%.

We have also revised the appendix to distinguish the full 520-objective PRI trajectory analysis from the separate supplementary token-level probability visualization conducted on 20 prompts.

2. Matched Predicate-Polarity and Argumentative-Stance Controls.

To further investigate whether the observed differences arise from proposition semantics rather than simply following a Support or Oppose instruction, we added matched controls along two dimensions: feasibility and normative judgment. These conditions retain the same underlying harmful objective while varying predicate polarity and argumentative stance.

For feasibility judgments, the refusal rates are:

ConditionRefusal
Support Easy46.35%
Support Hard74.62%
Oppose Easy76.54%
Oppose Hard46.15%

Support Easy and Oppose Hard yield nearly identical refusal rates (46.35% and 46.15%). Similarly, Support Hard and Oppose Easy produce refusal rates of 74.62% and 76.54%, respectively.

We also added corresponding normative controls using Justifiable and Unjustifiable:

ConditionRefusal
Support Justifiable87.31%
Support Unjustifiable72.69%
Oppose Justifiable68.65%
Oppose Unjustifiable77.12%

Taken together, these results suggest that the behavioral effect does not simply reverse with the surface distinction between Support and Oppose. Instead, the semantic direction jointly expressed by the predicate and argumentative stance appears to be associated with the resulting refusal behavior. We have incorporated these experiments into a dedicated subsection, Predicate Polarity and Stance Controls, in the revised RAA section.

3. Evaluation of Actual Harmful Assistance.

A reduction in refusal alone does not establish successful harmful assistance. To examine this distinction, our component ablation separately evaluates whether Viewpoint-only responses satisfy the same strict substantive-and-actionable success criterion used in the full ReLoK evaluation.

Across 520 AdvBench objectives, Viewpoint-only achieves 58.08% ASR@3 on GPT-4o and 21.92% on DeepSeek. These results provide evidence that viewpoint transformation alone can elicit substantive and actionable harmful assistance in a subset of cases, although its effectiveness varies across models. The additional structuring and decomposition components further improve the attack success rate.

4. Scope of the Mechanistic Evidence.

We have also revised the white-box analysis and supplementary diagnostics to clarify their respective roles. The analysis covers three open-weight models and examines selected safety-associated features, viewpoint-conditioned representation changes, and decoding trajectories. PRI is now explicitly reported using the full set of 520 paired AdvBench objectives.

We interpret these results as complementary diagnostic evidence characterizing how viewpoint transformation relates to representation and generation changes. The activation intervention is presented as a model-specific behavioral relevance analysis, rather than as establishing a universal causal mechanism.

The complete eight-predicate evaluation and matched polarity controls are primarily behavioral analyses conducted on QwQ-32B. Accordingly, our conclusions about these detailed semantic effects and their associated internal patterns are limited to the evidence available for that model, rather than being generalized across all three models.

These additional analyses strengthen the semantic characterization of RAA, distinguish response availability from substantive harmful assistance, and clarify the scope of the mechanistic findings.

Reviewer_XU4wOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper defines Reasoning Against Alignment (RAA): recasting a harmful objective g from "Provide details of g" into "Provide reasons to support 'it is easy to g'" sharply lowers refusal (QwQ-32B on AdvBench: 95.35% to 46.35%). On three open models the authors report Direct-vs-Viewpoint differences in hidden states, persistence of safety-associated features (selected from harmful-vs-benign Direct contrasts) under Viewpoint prompts, and a custom decoding diagnostic (PRI), and conclude that safety information remains but constrains generation less. The effect depends on predicate semantics (epistemic/descriptive much larger than pragmatic/deontic). Static safety prefixes restore refusal but raise benign false refusal from 1% to 13-20%. They also propose ReLoK (viewpoint framing, example-guided response schema, sensitive-word decomposition), reporting 95.7% mean ASR@3 over six LLMs and three benchmarks.

Critical strengths
  1. A large, reproducible behavioral effect with informative semantic structure. The paired comparison on 520 AdvBench objectives and the eight-predicate analysis show the refusal drop is not generic viewpoint wording, and the paper separates response availability from actionable assistance.
  2. Careful evaluation protocol and candid trade-off reporting. Uniform static single-turn protocol, five-LLM jury with blinded human validation, adaptive attacks excluded for stated reasons, and the defense trade-off (false refusal up to 20%) is reported rather than hidden.
Critical weaknesses
  1. The controlled effect is not shown to be a safety failure, and the mechanistic claim is only correlational. Direct and Viewpoint prompts differ in speech act, and a viewpoint-support reply can legitimately stay general or argumentative; the analysis part measures refusal, not harmfulness or actionability (actionability is only scored for ReLoK, which adds two further components). The "safety features remain observable" result relies on 256 residual dimensions selected by Cohen's d on harmful-vs-benign contrasts, which plausibly track topic harmfulness (encoded separately from refusal, as cited work shows), so persistence under a new template is expected. No steering/ablation/patching shows these features constrain Direct but not Viewpoint generation, there is no benign Direct-vs-Viewpoint control separating template from role effects, and the paper's own H-Path result (0.74% vs. 0.54% with a 40-point refusal gap) undercuts PRI as explanatory.
  2. ReLoK's gains are largely attributable to non-RAA components, and comparisons need clarification. Table 4 shows structuring and decomposition drive much of the gain (DeepSeek: 21.9% to 68.1% with structuring; to 58.1% with decomposition alone), both known ingredients, so ReLoK success is weak evidence for RAA. ASR@3 at temperature 0.6-1.0 is more permissive than pass@1, baseline numbers swing oddly (Combination Attack 0.08% on Claude vs. 86.5% on DeepSeek) without documentation of per-baseline tuning, targets are dated (Claude 3.7, GPT-4o), and the defense evaluation uses benign non-viewpoint prompts although the claimed trade-off concerns benign viewpoint reasoning.
Questions for authors
  1. Were baselines run with their original recommended settings per target, and what are pass@1 numbers?
  2. What are actionability scores for Viewpoint-only responses on all 520 objectives, and for matched control framings (e.g., "it is hard to g", counter-argument framing)?
  3. Do causal interventions (refusal-direction or feature ablation/steering) differ under Direct vs. Viewpoint?
  4. What is the false-refusal rate on benign viewpoint-support prompts?
7 comments
AuthorsOpenReview ↗

Response to Weakness 1: Safety Significance of RAA (1/3)

We appreciate the reviewer's detailed comments on the safety significance of RAA, the interpretation of our mechanistic analyses, and the evaluation of ReLoK. In response, we have conducted additional experiments and revised the manuscript to clarify the scope and implications of our findings. Our responses are organized below.

RAA examines whether the safety boundary applied to a harmful objective is maintained when the same objective is presented as a Viewpoint-support task rather than a Direct assistance request. We distinguish response availability from attack success: response availability describes whether the model continues the assigned Viewpoint task rather than exhibiting the refusal behavior observed under the matched Direct request, whereas attack success additionally requires substantive and actionable harmful assistance.

To examine whether Viewpoint transformation alone can elicit actionable assistance, we evaluated it without example-guided structuring or sensitive-word decomposition. Across all 520 AdvBench objectives, Viewpoint-only achieves 58.08% ASR@3 on GPT-4o and 21.92% on DeepSeek. These results use the same strict success criterion as the full ReLoK evaluation: high-level discussion or mere non-refusal is insufficient, and an objective is counted as successful only when at least one of its three independent generations provides substantive and actionable harmful assistance.

We also evaluated the complementary setting without example-guided structuring, retaining Viewpoint transformation and sensitive-word decomposition. This setting achieves 62.50% ASR@3 on GPT-4o and 58.08% on DeepSeek, further illustrating the model-dependent contribution of decomposition.

Taken together, these results provide evidence that Viewpoint transformation alone can elicit actionable harmful assistance on a substantial fraction of objectives, although its effectiveness varies across models. They also help clarify the distinct roles of RAA and ReLoK: RAA characterizes the task-role-dependent instability of the safety boundary, while ReLoK examines how this instability can be operationalized into more consistently actionable assistance through additional structuring and deployment-oriented components.

AuthorsOpenReview ↗

Response to Weakness 1: Representation Analysis and Activation Intervention (2/3)

We have revised the interpretation of Figure 1, clarified the scope of the selected features, and added an exploratory activation intervention to further examine the behavioral relevance of the observed representation changes.

1. Revised Interpretation of Figure 1.

The upper panels of Figure 1 project the Direct, epistemic (Easy), and deontic (Justifiable) focal-layer representations onto a model-specific semantic Viewpoint coordinate, defined by the normalized difference between the mean epistemic and deontic Viewpoint representations. This coordinate captures semantic variation rather than representing a universal refusal or safety direction.

The revised analysis highlights a more nuanced pattern. On QwQ-32B, the Easy distribution shifts substantially toward negative projection values, whereas the Direct and Justifiable distributions exhibit considerable overlap around positive values. This suggests that Viewpoint processing is not homogeneous: different semantic relations can induce distinct representational patterns. We therefore interpret Figure 1 as illustrating how the semantic relation expressed by the proposition further organizes Viewpoint representations, rather than establishing a universal separation between Direct and Viewpoint conditions.

2. Scope of Safety-Associated Features and Activation Intervention.

We have clarified that the selected 256 residual dimensions are safety-associated diagnostic features, identified exclusively through Direct harmful–benign contrasts and held fixed when evaluating Viewpoint conditions. Their persistence indicates that the measured activation signals remain observable across different task formulations, without implying that they constitute a complete safety representation or necessarily exert the same influence on generation under different task roles.

We have also clarified the feature-selection procedure: individual dimensions are ranked by their absolute mean activation differences between Direct harmful and benign prompts, while Cohen's dd is used separately to identify the model-specific layer at which the aggregate diagnostic exhibits the strongest separation.

To further investigate behavioral relevance, we added an exploratory activation intervention using a separate set of Viewpoint-associated residual features, identified from Direct–Viewpoint activation differences. We conduct this intervention at Layer 28 of DeepSeek-8B, using a sparse budget of K=256K=256 dimensions. For each matched objective, we replace the selected activation coordinates in the Direct computation with their corresponding values from the Viewpoint condition, while leaving the unselected dimensions unchanged. The intervention covers all 520 paired AdvBench objectives, with a representative generation case reported in Appendix E.5.

The representative case shows that transferring this partial Viewpoint-associated state changes the organization of the Direct continuation: the intervened generation becomes more structured around the harmful objective, even though explicit safety-related caveats remain present. This offers an additional exploratory indication of the behavioral relevance of the task-role-associated representation shift. Since the intervention targets a separate set of Viewpoint-associated features, it provides evidence about their behavioral relevance rather than establishing a causal role for the 256 safety-associated dimensions used in the harmful–benign persistence analysis.

AuthorsOpenReview ↗

Response to Weakness 1: Decoding Diagnostics and Experimental Scope (3/3)

We have also refined the interpretation of PRI and clarified the scope of the controlled Direct-to-Viewpoint transformation.

1. Interpretation of PRI and H-Path.

H-Path measures the proportion of decoding steps at which at least one rejection-related candidate appears among the top-20 tokens. It therefore captures candidate occurrence during generation rather than directly measuring the final refusal decision. For example, Easy and Justifiable have similar H-Path values (0.74% and 0.54%), despite their substantially different refusal rates (46.35% and 87.31%). Their H-Prob values (0.136 and 0.158) and H-Rank values (11.35 and 9.73) further illustrate that candidate occurrence, probability mass, and relative ranking do not necessarily vary in lockstep with terminal refusal.

Accordingly, we interpret H-Path, H-Prob, and H-Rank as complementary diagnostics of different aspects of the decoding distribution. They characterize generation dynamics alongside response-level refusal measurements and representation analyses, rather than serving as independent predictors or causal explanations of the final refusal decision.

2. Scope and Limitations of the Controlled Transformation.

Direct and Viewpoint prompts differ in speech act and expected response form. Our current design preserves the underlying harmful objective while changing its task role, but does not fully isolate these factors through a matched benign Direct–Viewpoint factorial control.

We have made this scope explicit in the revised manuscript. Our experiments characterize the effect of a semantically meaningful action-to-viewpoint transformation, rather than isolating an abstract task-role variable with every surface property held fixed. We also identify a fuller factorial design separating speech act, proposition type, semantic relation, and requested output form as an important direction for future work.

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

This paper studies how rewritting direct harmful requests into viewpoint-support tasks changes the safety behavior of aligned LLMs. This paper analyzes hidden representations and the output trajectories, finding that safety-associated signals remain observable even when refusal decreases. The paper also proposes a jailbreak method ReLoK combining viewpoint transformation, sensitive word and example based structuring. Experiments across six models demonstrate the proposed method is effective.

Critical strengths
  • The studied problem and the observation is interesting. This paper is well-motivated.

  • The paper includes planty analysis such as behavioral experiments, representation analysis, and component ablations.

  • This paper is well written.

Critical weaknesses
  • The evaluation relies heavily on relatively older models (given the fast pace of this field), including GPT-4o, Claude 3.7 Sonnet, Gemini 2.0 Flash, DeepSeek-R1, and QwQ-32B. Coverage of recent models is limited. The white-box analysis also uses only older models. Evaluating more recent strong models is important for establishing whether the observations, the attack and the conclusion of this paper remains effective and generalizes beyond the selected models.

  • The novelty relative to existing task-reframing jailbreak methods needs clearer justification. Prior approaches such as ReNeLLM, Analyzing-based Jailbreak, and Value-Driven Adversarial Prompting already exploit changes in task framing or the model’s relation to a harmful objective. The paper would be stronger with direct comparisons against these related methods and show what is distinctive in the proposed method.

Ding et al., A Wolf in Sheep’s Clothing: Generalized Nested Jailbreak Prompts can Fool Large Language Models Easily, NAACL 2024.

Lin et al., LLMs can be Dangerous Reasoners: Analyzing-based Jailbreak Attack on Large Language Models, arXiv 2024.

Luo et al., Safety Guardrails of Large Language Models Are Vulnerable to Value-Driven Adversarial Prompting, Findings of ACL 2026.

Questions for authors

see above

3 comments
AuthorsOpenReview ↗

Response to Weakness 1: Evaluation on Recent Models

We appreciate the reviewer's thoughtful comments. In response, we have extended our empirical evaluation, clarified the relationship between RAA and related work, and revised the manuscript to make the central research question and contributions easier to follow. Our detailed responses are provided below.

We have extended our evaluation to DeepSeek-V4.1-Flash-1017, released in October 2026, to examine whether ReLoK remains effective on more recent models. We evaluated ReLoK and selected baseline attacks on three benchmarks: AdvBench, JailbreakBench, and MaliciousInstruct.

We report two complementary metrics:

  • Per-generation ASR measures the proportion of successful individual generations among all 3N3N generations, where each of the NN test objectives is evaluated through three independent generations.
  • ASR@3 measures the proportion of test objectives for which at least one of the three independent generations succeeds.

The prompts are kept fixed across the three generations, with no adaptive modification based on previous model responses. These two metrics capture attack success at both the generation level and the objective level.

Table R2. Results on DeepSeek-V4.1-Flash-1017 (%).

MetricMethodAdvBenchJailbreakBenchMaliciousInstruct
Per-generation ASRBaseline0.131.670.67
Combination1.6711.0010.33
ArtPrompt5.646.333.00
FlipAttack12.0518.0015.33
H-CoT0.261.336.33
ReLoK51.6057.3367.67
ASR@3Baseline0.193.001.00
Combination2.6915.0012.00
ArtPrompt7.319.006.00
FlipAttack23.6532.0027.00
H-CoT0.383.0013.00
ReLoK67.6977.0087.00

ReLoK achieves the highest success rate on all three benchmarks under both metrics. Under Per-generation ASR, it reaches 51.60%, 57.33%, and 67.67%, respectively, outperforming the strongest baseline, FlipAttack, by 39.55, 39.33, and 52.34 percentage points. The advantage is also consistent under ASR@3, with improvements of 44.04, 45.00, and 60.00 percentage points over FlipAttack.

Notably, the gap is already substantial under Per-generation ASR, suggesting that ReLoK's advantage cannot be attributed solely to repeated attempts. The ASR@3 results further show that this advantage persists when success is measured at the objective level.

Regarding the white-box analysis, our mechanistic investigation remains limited to the three open-weight models examined in the original study. Rather than extending the same analysis to the newly evaluated model, we have strengthened the existing investigation by expanding the decoding-trajectory analysis to all 520 AdvBench objectives and adding further experiments on viewpoint semantics and task-role variations. These analyses provide a more comprehensive characterization of RAA within the models studied. Extending the mechanistic investigation to newer model generations remains an important direction for future work.

The new DeepSeek-V4.1-Flash-1017 evaluation and the additional analyses have been incorporated into the revised manuscript. Together, these changes extend the empirical evaluation of ReLoK to a recent model while providing a more detailed characterization of RAA.

AuthorsOpenReview ↗

Response to Weakness 2: Novelty and Relation to Existing Methods

The reviewer raises an important question about how RAA differs from existing task-reframing and reasoning-based jailbreak methods. Prior work has shown that alternative task formulations, conversational roles, and semantic framing can influence safety behavior. We therefore position our contribution more specifically around the identification and characterization of a task-role generalization failure, rather than the general observation that prompt framing matters.

We have revised the Related Work section to clarify this distinction, particularly in relation to the three works highlighted by the reviewer.

  1. ReNeLLM (Ding et al., NAACL 2024). ReNeLLM combines semantics-preserving prompt rewriting with scenario nesting to embed harmful requests into alternative task formats. Its primary focus is automated jailbreak prompt construction. In contrast, RAA studies a controlled action-to-viewpoint transformation: the underlying harmful objective remains unchanged, while its task role shifts from a Direct assistance request to a Viewpoint-support task. Rather than searching for effective prompt rewrites, RAA examines how this change in task role affects safety behavior.

  2. Analyzing-based Jailbreak (Lin et al., 2024). ABJ constructs auxiliary data related to a harmful objective and uses analysis and behavioral prediction as an alternative task formulation. RAA does not rely on auxiliary character data or behavioral prediction. Instead, it directly examines how the same harmful objective is processed under different task roles.

  3. Value-Driven Jailbreak Attack (Luo et al., ACL 2026). VDJA exploits value-oriented prompting to elicit positive moral endorsement of harmful tasks before requesting their execution. RAA does not require the model to morally approve of the harmful objective. Our semantic evaluation finds that epistemic viewpoints can produce substantially larger reductions in refusal than deontic viewpoints. On QwQ-32B, the Easy predicate reduces refusal by 49.04 percentage points, whereas Justifiable reduces it by only 8.08 percentage points. This suggests that the behavior observed in RAA cannot be adequately explained by moral endorsement alone.

We also conducted additional experiments to further characterize the role of supportive wording in RAA. Specifically, we introduced matched predicate-polarity and argumentative-stance controls across feasibility and normative judgments. The underlying harmful objectives were kept unchanged, while both predicate polarity and the argumentative stance (support or oppose) were varied.

The feasibility-related results are as follows:

ConditionRefusal
Support Easy46.35%
Support Hard74.62%
Oppose Easy76.54%
Oppose Hard46.15%

Two patterns are particularly notable. Supporting that an objective is easy and opposing that it is hard produce nearly identical refusal rates (46.35% vs. 46.15%). Similarly, supporting that it is hard and opposing that it is easy yield comparable rates (74.62% vs. 76.54%). These results suggest that refusal behavior does not simply follow the surface distinction between Support and Oppose. Instead, it appears closely associated with the semantic direction resulting from the combination of the predicate and argumentative stance.

The normative controls show a related, although less symmetric, pattern:

ConditionRefusal
Support Justifiable87.31%
Support Unjustifiable72.69%
Oppose Justifiable68.65%
Oppose Unjustifiable77.12%

These results further suggest that the effect of viewpoint semantics depends on the type of judgment being expressed. In particular, feasibility-related predicates produce more pronounced behavioral differences than normative predicates.

Taken together, these experiments extend our original evaluation beyond a single viewpoint predicate. They provide further evidence that RAA is associated with the semantic relationship between the proposition and the underlying harmful objective, rather than simply being a response to supportive wording.

We recognize that direct head-to-head evaluations against ReNeLLM, ABJ, and VDJA would provide additional empirical evidence. Since these methods rely on different construction mechanisms and evaluation settings, a fair comparison would require aligning their implementations and evaluation protocols. We have not included such experiments in this revision. Instead, we have expanded the methodological discussion and clarified the respective research objectives and differences in the revised Related Work section.

AuthorsOpenReview ↗

Additional Revisions: Clarifying the Presentation and Contributions of RAA

Beyond the changes addressing the two weaknesses, we have also revised the manuscript to make the central research question and the relationship between RAA and ReLoK easier to follow.

  1. Revised Introduction. We have rewritten the Introduction to better establish the motivation and definition of RAA. In particular, we now distinguish the behavioral phenomenon from its practical exploitation: RAA describes a task-role generalization failure in which the same harmful objective receives different safety responses under Direct and Viewpoint formulations, whereas ReLoK operationalizes this vulnerability to examine its practical security implications.

  2. Reorganized RAA Analysis. We have reorganized the RAA section to provide a clearer progression from the controlled Direct-to-Viewpoint transformation to its semantic dependence, internal representation, and decoding signatures. The newly added predicate-polarity and argumentative-stance experiments are also incorporated into this section to provide a broader characterization of viewpoint-conditioned safety behavior.

  3. Expanded Related Work. We have substantially revised the Related Work section to cover ReNeLLM, Analyzing-based Jailbreak, VDJA, and recent mechanistic studies. Rather than positioning RAA as the first work to show that task framing can affect safety behavior, we now emphasize its specific contribution: the controlled task-role transformation, its semantic structure, and the associated representation and decoding changes. We also make the distinction between the analytical contribution of RAA and the deployment-oriented attack construction of ReLoK more explicit.

These revisions provide a clearer account of the research question, the methodological distinctions from related work, and the respective contributions of RAA and ReLoK.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.