acceptodds
Under review as a conference paper at ICLR 2027

Selective Safety Deliberation via Uncertainty Resolution with Learned Intent Token

Abstract

Selective defenses in Large Language Models (LLMs) reduce the cost of safety deliberation by using a lightweight classifier on a model internal to decide when to intervene, yet often use it merely as a gate, routing confidently harmful and uncertain requests to the same deliberation. We empirically find that these cases require different actions: the model often complies with requests that the classifier confidently flags as harmful, whereas adversarially rewritten requests concentrate in the low-confidence region, where most errors occur. We propose RESTATE, a selective defense that resolves uncertain safety decisions before deferring to deliberation. RESTATE refuses confidently harmful requests and answers confidently benign ones. For uncertain requests, it appends a learned intent token which refines representation to mitigate uncertainty. The same classifier then scores the representation at this token, and only requests that remain uncertain are deliberated. Since the token is used only for reassessment and the LLM stays frozen, response generation is unaffected. On Qwen3-8B with WildJailbreak, reassessing uncertain requests raises the classifier's balanced accuracy by 7.1%p. Averaged over six inference-time defenses, RESTATE lowers attack success from 37.7% to 10.2% and over-refusal from 13.7% to 7.2% relative to always deliberating, while deliberating only 14.2% of requests.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM, Safety, Alignment

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_WPeYOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a selective LLM safety defense framework based on uncertainty resolution. A lightweight linear probe is first used to judge the potential harmfulness of a user request: high-confidence harmful requests are directly refused, high-confidence harmless requests are answered normally, and for uncertain requests, a learnable Intent Token is introduced to generate new hidden representations using a frozen language model, which are then re-evaluated by the same probe.

Critical strengths

The method is lightweight and can be combined with a variety of existing inference-time defense methods. The paper provides fairly comprehensive experiments.

Critical weaknesses
  • The authors attribute the performance gains mainly to the Intent Token's ability to extract harmful intent more effectively, but the current experiments are not sufficient to fully substantiate this explanation. In particular, the additional representation computation, the training on uncertain samples, and changes in the classification decision boundary may jointly affect the final results. Furthermore, although the paper tests different models, the main performance gains are concentrated on Qwen3-8B, while the gains on other models are limited.

  • The security defense mechanism of this system relies heavily on the classification results of the lightweight linear probe, which may introduce additional adversarial robustness risks. In particular, when a harmful request is misclassified as safe by the initial probe with high confidence, the system will directly generate a response without triggering the Intent Token or subsequent safety reasoning. Although the authors evaluate a variety of adaptive jailbreak attacks, these experiments mainly target the system's final response and have not yet adequately examined attacks specifically targeting the probe's classification boundary or the safety routing mechanism.

Questions for authors

The security of RESTATE relies heavily on the classification results of the initial linear probe. If an attacker can push the hidden representation of a harmful request into the high-confidence safe region, the request may bypass all subsequent defense mechanisms. Have the authors considered adaptive attacks specifically targeting the probe's classification boundary or the routing decision?

4 comments
AuthorsOpenReview ↗

Response to Reviewer WPeY (1/2)

Dear Reviewer WPeY,

We sincerely appreciate your thoughtful review. We carefully considered your concerns regarding (1) the attribution of performance gains and generalization across backbones, and (2) the robustness of the initial probe and routing mechanism against adaptive attacks. Below, we provide additional experimental evidence and clarifications.


[W1] Attribution of gains and cross-backbone generalization

Thank you for raising these important concerns regarding the attribution of performance gains and reduced gain in larger backbones. We appreciate the opportunity to clarify how additional representation computation, uncertainty-focused training, and classification decisions contribute to RESTATE's improvements.

(1) Attribution of performance gains. First, the backbone and Anchor remain frozen throughout intent token training. Thus, improvements do not arise from updating the Anchor's decision boundary; instead, the learned intent token produces an additional representation evaluated by the same fixed classifier.

Our original ablations already examine several contributing factors. Simply appending an untrained token reduces BAcc (85.50→80.26, Table 3), while retraining the Anchor on additional difficult examples yields limited or negative gains (joint: 86.39; sequential: 79.34, Table 2). Moreover, with the same trained token and escalation rate (38.4%), uncertainty-based escalation substantially outperforms random escalation (92.62 vs. 88.41, Table 3), demonstrating the importance of targeting uncertain requests.

To further examine the benefit of separating general classification from targeted reassessment, we additionally evaluated joint training of probe and token.

MethodBAcc ↑FNR ↓F1 ↑AUC ↑
Anchor85.5017.0489.3993.40
Joint probe + token87.1314.1791.1194.01
RESTATE92.627.7095.2296.79

While joint training also improves classification, separating general harmfulness assessment from uncertainty-focused reassessment achieves a further +5.49pp BAcc improvement. Together with our original ablations, these results suggest that RESTATE's gains are not solely attributable to additional computation or probe retraining, supporting the benefit of learning an additional representation for uncertain requests while keeping the original classifier fixed.

(2) Larger backbone generalization. The original cross-backbone experiments employed a restricted hyperparameter search space, primarily to establish RESTATE's applicability across different models under limited computational budget. To fairly assess its effectiveness, we applied the original search space to each backbone and selected configurations based on a held-out validation set (see Appendix E.4 for details). The resulting configurations achieved average improvements of +2.42pp on Qwen3-32B and +2.70pp on GPT-OSS-20B across BAcc, FNR, and AUC.

MetricBAcc ↑FNR ↓AUC ↑
BackboneAnchorOursAnchorOursAnchorOurs
Qwen3-32B83.6785.5311.096.7892.5093.58
GPT-OSS-20B82.4786.1615.6512.3991.5992.75

Overall, these additional results provide further evidence of RESTATE's effectiveness beyond Qwen3-8B. We have also refined our description of representation refinement to emphasize uncertainty-targeted reassessment, rather than implying that the intent token independently discovers new safety features.

AuthorsOpenReview ↗

Response to Reviewer WPeY (2/2)

[W2, Q1] Adaptive attacks targeting the safety routing mechanism

Thank you for raising this important security concern. We agree that high-confidence misclassification is a potential vulnerability in selective safety defenses, as it may allow harmful requests to bypass subsequent safeguards.

This is an inherent trade-off in selective defense systems: reducing high-confidence false negatives generally requires escalating more requests to downstream safeguards, increasing computational costs. For instance, Constitutional Classifiers++[1] adopts a two-stage cascade with a lightweight first-stage classifier, allowing more potentially harmful requests to be flagged for further assessment rather than directly refused. Similarly, RESTATE's uncertainty thresholds can be easily adjusted to balance the risk of missed harmful requests against the deliberation rate.

Regarding adaptive jailbreak attack, we clarify that our evaluation adopts a black box threat model, reflecting practical deployment settings where probe weights and classification scores are not exposed to external users [2]. However, prior studies have shown that attackers may infer which safeguard handled a request from observable responses [3, 4]. Therefore, to fully address the reviewer’s concern, we conduct additional experiment by considering an attacker who can observe the routing outcome (direct answer, fixed refusal, or deliberation), without accessing internal classifier information. This provides a realistic, route-aware black box setting for evaluating attacks specifically targeting RESTATE's routing mechanism.

To examine whether an attacker can exploit the confident-allow path, we conducted an additional adaptive attack targeting RESTATE’s routing decision. Specifically, following the staged attack strategy of STACK [3], we selected 39 HarmBench behaviors for which previous PAIR/TAP attacks had successfully jailbroken the undefended Qwen3-8B. RESTATE initially routed 38 of these adversarial requests to fixed refusal and one to deliberation. We then adapted the random search suffix attack [5] to target the routing decision, using up to 9,600 queries per behavior to redirect requests toward the direct-answer path.

We considered two attacker feedback settings: (1) 3-way routing, which reveals the routing outcome (direct answer, fixed refusal, or deliberation), and (2) direct answer indicator, which provides only a binary signal indicating whether the request reaches the direct-answer path.

DeploymentAttacker feedbackRouting bypasses ↓
RESTATE3-way routing2/39
RESTATEDirect-answer indicator1/39
+ Token-confirmed allow3-way routing0/39
+ Token-confirmed allowDirect-answer indicator0/39

In table above, a routing bypass indicates that the adversarial request reaches the direct-answer path without triggering refusal or deliberation, rather than a confirmed harmful response. Here, across approximately 360,000 queries per feedback setting, the attacker found routing bypasses for only 1–2 of the 39 behaviors, providing additional evidence of RESTATE’s robustness against adaptive attacks targeting its routing mechanism.

We additionally evaluated a simple variant, token-confirmed allow, which permits direct generation only when the initial probe's confident benign decision is also supported by the intent token assessment. Under the same attacks, no routing bypass was found (0/39). On WildJailbreak, this variant also reduces average ASR from 10.20% to 8.93%, with a modest increase in ORR (7.16%→8.04%) and deliberation rate (14.2%→15.6%).

Overall, these results provide additional evidence of RESTATE's robustness against adaptive attacks targeting the routing decision, while acknowledging that high-confidence misclassification remains a potential failure path. We have included this evaluation in the revised Appendix E.8 (Table 29), together with a discussion of practical safeguards, such as reducing routing-information leakage and monitoring repeated flagged requests.


[1] Cunningham et al. Constitutional Classifiers++: Efficient Production-Grade Defenses against Universal Jailbreaks (ICLR, 2026)

[2] Bailey et al. Obfuscated Activations Bypass LLM Latent-Space Defenses (ICLR, 2026)

[3] McKenzie et al. STACK: Adversarial Attacks on LLM Safeguard Pipelines (AAAI, 2026)

[4] Hackett & Garraghan. Behind the Refusal: Determining Guardrail Activation via Behavioral Monitoring (Preprint, 2026)

[5] Andriushchenko et al. Jailbreaking Leading Safety-Aligned LLMs with Simple Adaptive Attacks (ICLR, 2025)


We are grateful for your insightful suggestions, which helped us clarify the contributions of uncertainty resolution and further evaluate robustness against attacks targeting the routing mechanism. We hope our additional experiments and analyses satisfactorily address your concerns.

If you have any further questions, please let us know.

Thank you very much,

Authors

Thank you for the careful clarification. Please make sure including these results in the revised version.

AuthorsOpenReview ↗

Response to Reviewer WPeY

Dear Reviewer WPeY,

We are pleased to hear that our rebuttal addressed your concerns. We will ensure that the additional results presented in Appendix E of our rebuttal are properly integrated into the camera ready, with clearer explanations and discussion. Please let us know if you have any further questions.

Thank you very much.

Sincerely,
Authors

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

This paper proposes RESTATE, a selective safety defense based on internal probes. An initial probe scores every request, allowing the system to directly handle confidently benign or harmful requests. For uncertain requests, a shared, learnable intent token is appended, and the frozen language model produces a new context-dependent representation that is scored by the same frozen probe. Only requests that remain uncertain are escalated to a downstream safety defense. Training updates only the token embedding using a classification objective with an adversarial perturbation term; both the backbone and probe remain frozen. The token is excluded from normal response generation. The authors evaluate the method across multiple models and classification and safety benchmarks, with component ablations, probe-transfer experiments, and adaptive attacks.

Critical strengths
  • The method is simple, modular, and parameter-efficient. Updating a single embedding while freezing the backbone and probe provides a concrete way to construct improved safety-detection representations for uncertain requests. The token can be discarded after detection, and the framework can accommodate different downstream defenses. This makes the design practically relevant for reducing safety-monitoring costs.

  • The gains on difficult examples and the component analyses are valuable. On Qwen3-8B, the authors report an increase in WJ balanced accuracy from 85.50% to 92.62%, and from 79.19% to 85.20% on the external mixture. The evaluation includes token and loss ablations, probe retraining, transfer to independently trained probes, and adaptive attacks against the complete policy. These analyses provide evidence that the reassessment component is useful beyond aggregate system-level comparisons.

Critical weaknesses
  • The independent contribution of the single-token mechanism remains insufficiently established. The authors do cite and compare against Anthropic’s suffix-probe approach, including Suffix and EMA → Suffix baselines; the concern is therefore not missing related work. Rather, appending content to improve classification representations already has a direct precedent, so the incremental advantage of a learnable single token needs to be established under comparable training conditions. In Table 11, Token + standard BCE achieves 86.73% WJ BAcc, only 1.23 percentage points above the 85.50% Anchor baseline. Robust BCE raises this to 92.62%, an additional 5.89 points. This indicates that robust training contributes substantially to the overall gain, but does not isolate its contribution from that of representation reassessment. Comparable robust training for trainable probe and representation-adaptation baselines, controlling training data, difficult-example selection, and routing calibration, would strengthen the attribution. For the fixed suffix baseline, the training conditions of its readout probe and its computational costs should be explicit. This issue materially affects the assessment of novelty and the claimed advantage over existing approaches.

  • The adaptive-robustness results cannot yet be fully reconciled, and aggregate gains obscure some regressions. Section 5.3 reports HarmBench any-query ASR of 71.3% for Always and 53.2% for RESTATE. However, applying the equal-weight averaging over the three attacks described in A.5 to Table 13 gives approximately 49.78% and 28.67%, respectively. The authors should clarify whether this reflects different aggregation rules or an unsynchronized revision. Moreover, some configurations deteriorate: Table 14 reports HarmBench / GPTFuzzer / IA any-query ASR increasing from 5.50% to 14.50%; Table 10 reports SORRY-Bench FNR increasing from 6.89% to 9.11% and XSTest FPR from 1.60% to 3.60%. These cases do not negate the average improvements, but limit conclusions about consistent improvements to the safety–utility trade-off. Since confidently misclassified harmful requests can bypass further checks, stage-specific failure analysis and consistent main-text and appendix reporting are important for assessing the safety claims.

Questions for authors

The following are clarification and validation requests whose answers could materially affect my assessment.

  1. Attribution of gains: After matching training data, difficult-example selection, and applicable Robust BCE training conditions, how much advantage remains over probe retraining, suffix probes, or other lightweight representation-adaptation methods? Which existing comparisons already control these factors?

  2. Consistency of results: What aggregation procedures produce the HarmBench adaptive-attack results in Section 5.3 / Figure 5 and Tables 13–14? Please provide reconcilable denominators, attack budgets, and per-configuration results, and clarify whether any results were not synchronized across revisions.

  3. Failure paths: In configurations with worse safety or over-refusal, do errors primarily arise from initial high-confidence decisions, intent-token reassessment, or downstream defenses? What fraction of harmful requests is confidently allowed through by the initial probe?

  4. Actual computational costs: At matched safety and utility levels, how do end-to-end latency, throughput, and memory usage change? Do the savings from fewer downstream-defense calls outweigh the extra token computation, cache operations, and effects of routing on batching efficiency?

6 comments
AuthorsOpenReview ↗

Response to Reviewer Ymxx (1/4)

Dear Reviewer Ymxx,

We sincerely appreciate your careful and constructive review. We have carefully considered your concerns regarding (1) attribution of performance gains under comparable training conditions, (2) consistency of adaptive attack results, (3) analysis of safety failure paths, and (4) practical computational costs. Below, we address these points through clarifications and additional experimental evidence.


[W1, Q1] Attribution of gains under comparable training conditions

Thank you for emphasizing the importance of distinguishing the contribution of intent token reassessment from training data selection and robust optimization. While several relevant controls were included in the original manuscript (Section 4.3), we recognize that their respective roles in attributing the performance gains could have been more clearly distinguished. We clarify these distinctions below and made them more explicit in the revised manuscript.

(1) Matching training data and difficult example selection. Table 1 compares classifier baselines trained with access to the same combined training data, including the difficult examples selected for intent token training (except Anchor). These comparisons therefore control for data availability, although the baselines learn from these examples jointly rather than through a dedicated uncertainty resolution stage.

Table 2 further examines whether additional training on difficult examples alone explains the improvement. Retraining Anchor on the combined data achieves 86.39% BAcc, while sequential retraining on difficult examples yields 79.34%, compared with 92.62% for RESTATE. These results suggest that simply retraining the probe on additional difficult examples is less effective than our uncertainty focused reassessment approach.

(2) Contribution of RobustBCE to performance improvement. We acknowledge that Robust BCE contributes substantially to the improvement, as shown in Table 11 (BAcc: 86.73→92.62). To further isolate this contribution, we additionally evaluated probe retraining, Suffix, and joint training of probe and token under comparable BCE and Robust BCE conditions, using matched training data and perturbation settings where applicable.

MethodLossBAcc ↑FNR ↓AUC ↑
Linear probeBCE86.3915.2693.85
Robust BCE86.6515.1393.97
MLP probeBCE85.5011.3593.85
Robust BCE85.9311.8793.92
Suffix probeBCE84.3513.2692.11
Robust BCE84.9511.8793.27
Joint probe + tokenBCE85.719.1791.14
Robust BCE87.1314.1794.01
RESTATE (Ours)BCE86.7310.2694.53
Robust BCE92.627.7096.79

Applying Robust BCE also yields improvements for alternative methods, including the Suffix probe and lightweight representation adaptation approaches, although the gains remain relatively modest. For instance, Suffix improves from 84.35 to 84.95 BAcc, and joint probe and token training improves from 85.71 to 87.13, both remaining below RESTATE (92.62). These results suggest that combining robust optimization with targeted reassessment of uncertain requests is more effective than applying robust optimization to existing probe based or representation adaptation methods alone.

(3) Performance gain decomposition of RESTATE. We clarify that RESTATE’s substantial gain is attributable to the combination of appropriate selective reassessment and Robust BCE, rather than Robust BCE alone. All data and loss ablations for the intent token in Table 3 use our uncertainty based escalation. To distinguish the contribution of selective escalation, Table 3 applies the same intent token trained with Robust BCE to a random subset at the same escalation rate (38.4%). This achieves 88.41% BAcc, compared with 92.62% under uncertainty based escalation. The additional +4.21%p gain, with the learned token and escalation budget held fixed, demonstrates the substantial contribution of targeting uncertain requests.

Overall, these analyses suggest that RESTATE's gains arise from the combination of uncertainty specialized representation learning, robust optimization, and targeted reassessment, rather than any single factor alone. We have made these distinctions and the corresponding experimental conditions more explicit in the revised manuscript (Appendix E.2).

AuthorsOpenReview ↗

Response to Reviewer Ymxx (2/4)

[W2, Q2] Consistency of adaptive-attack results and aggregation procedures

Thank you for carefully pointing out this inconsistency and providing us with the opportunity to clarify our aggregation procedures. We acknowledge that the distinction between the two aggregation methods was not sufficiently explained in the original manuscript, which may have caused confusion. We clarify the differences and provide reconciled results below.

First, the discrepancy between Figure 5 and Tables 13–15 arises from different aggregation procedures, which were not clearly distinguished in the original manuscript.

Our evaluation design reflects the different roles of the attacker and the defender. To jailbreak a single harmful behavior, an attacker can try multiple adaptive attack methods, each of which iteratively rewrites the query along its attack trajectory. The jailbreak succeeds once any of these queries elicits a harmful response. In contrast, under selective safety deliberation, a defender deploys a single inference-time defense chosen for its efficiency. We therefore measure attack success against each defense separately and report the average over the defenses. The following metrics formalize this design.

  • Any-query ASR: Within a single attack trajectory, a harmful behavior is counted as successfully attacked if at least one query, including the initial request and intermediate rewrites, produces a response judged to fulfill the original harmful objective. Any-query ASR is the fraction of behaviors for which this occurs.
  • Attack-wise mean (Appendix): We compute any-query ASR separately for PAIR, TAP, and GPTFuzzer, then average their results. Table 13 reports each attack’s results averaged over IA, RA, and ZS; Tables 14–15 provide the individual results for each attack, defense, and routing policy.
  • Attack union any-query ASR (Main text): For each harmful behavior and defense, we count success if any query from any of the three attack methods succeeds. This corresponds to an attacker trying all three methods against a system with single deliberation method and is similar to the At Least One Attack Success Rate (ALO-ASR) used in MixAT [1]. Thus, the union is taken over attacks, not over defenses: we compute the attack union ASR separately for each defense, and Figure 5 reports the average of these three per-defense values (IA, RA, and ZS).

Therefore, the scores reported in the main text and Figure 5 are attack union any-query ASR averaged over the three defenses, while the Appendix reports results separately for each attack, whose average gives the attack-wise mean.

The following results reconcile the two aggregation procedures using the same underlying evaluation records.

BenchmarkAggregationAlwaysRESTATE
HarmBench (N = 200)Attack-wise mean896 / 1,800 (49.78%)516 / 1,800 (28.67%)
Attack union428 / 600 (71.33%)319 / 600 (53.17%)
SORRY-Bench (N = 180)Attack-wise mean1,115 / 1,620 (68.83%)820 / 1,620 (50.62%)
Attack union454 / 540 (84.07%)402 / 540 (74.44%)

We have also included the per-defense attack-union ASR results for both benchmarks in the revised Appendix E.5 (Table 22), allowing the aggregate results to be verified for each defense configuration.

Regarding attack budgets, we used up to 30 target queries for PAIR (3 streams × 10 iterations), 29 for TAP (branching 2, width 3, depth 10), and 30 for GPTFuzzer (30 mutation iterations), per behavior and defense configuration, excluding the initial request (see Appendix E.5 for details).

We apologize for the insufficient clarification of these aggregation procedures. In the revised manuscript, Section 5.3 and the Figure 5 caption now state that the results are attack-union any-query ASR. Appendix A.5 defines each success criterion, and the new Appendix E.5 distinguishes the attack-wise mean from attack-union ASR and reconciles the two (Tables 21–22).


[1] Dékány et al. MixAT: Combining Continuous and Discrete Adversarial Training for LLMs (NeurIPS, 2025)

AuthorsOpenReview ↗

Response to Reviewer Ymxx (3/4)

[W2, Q3] Analysis of failure paths

Thank you for this important question. We agree that analyzing where failures originate helps clarify the effectiveness and limitations of each component. Below, we refer to the three stages in your question as initial high-confidence decisions (Initial), intent-token reassessment (Intent), and downstream defenses (Downstream).

(1) Classification errors and subsequent deliberation. In Table 10, RESTATE increases FNR on SORRY-Bench and FPR on XSTest. The table below attributes each classification error of RESTATE to the stage at which it first arises: Initial counts errors already made by the Anchor, including those that remain uncorrected after reassessment, and Intent counts new errors introduced by intent token reassessment. Intent token reassessment corrects 8 and 1 Anchor errors but introduces 18 and 6 new errors on SORRY-Bench and XSTest, respectively.

BenchmarkInitialIntentTotal
SORRY-Bench (FN)231841
XSTest (FP)369

However, a classification error does not necessarily lead to a safety failure in the full defense pipeline, since requests that remain within the intent uncertainty band are deferred to deliberation. The table below shows how many of the new errors at Intent are deferred to deliberation, and how many of these are resolved after deliberation, averaged over six defenses.

BenchmarkNew errors at IntentDeferred to deliberationResolved after deliberation
SORRY-Bench188 (44.44%)56.25%
XSTest66 (100.00%)94.44%

Thus, classification errors do not necessarily become response-level failures, because downstream deliberation rescues a substantial portion of the deferred errors.

(2) Failure paths under adaptive attacks. We traced the stage producing the first harmful response in the GPTFuzzer/IA configuration, where ASR increases. On HarmBench, the first harmful response comes from Initial for 23 behaviors, Intent for 0, and Downstream for 6. On SORRY-Bench, the corresponding counts are 32, 2, and 29. Thus, most observed failures arise from initial high-confidence allow decisions, before intent-token reassessment is invoked.

Nevertheless, these failures can be easily mitigated by sending more requests to the inexpensive intent-token reassessment. Because reassessment is cheap, higher-recall screening remains efficient while full deliberation stays selective. This follows the cascade principle of tolerating more initial flags for further assessment [1]. We tested two ways to implement this:

  1. Intent confirmation: apply the same learned intent token even when the Anchor is confident-benign, and allow a direct answer only when the intent readout also confirms confident-benign.
  2. Wider uncertainty band: expand the Anchor band by 50% on the allow side, sending more requests to the existing intent-token reassessment.

Reassessed requests are refused or sent to downstream deliberation when confident-benign is not confirmed. The table below groups behaviors by the stage producing their first harmful response. Both refinements reduce ASR below Always IA in these two configurations. The wider band increases deliberation rates by only 0.49%p on HarmBench and 0.91%p on SORRY-Bench.

BenchmarkSettingInitialIntentDownstreamASR (%)
HarmBenchAlways IA00115.50
HarmBenchRESTATE IA230614.50
HarmBenchIntent confirmation0063.00
HarmBenchWider band3064.50
SORRY-BenchAlways IA004525.00
SORRY-BenchRESTATE IA3222935.00
SORRY-BenchIntent confirmation023218.89
SORRY-BenchWider band423120.56

To further examine the effect on over-refusal, the table below reports ORR (%) on benign evaluation sets for the same IA configurations. Relative to the original RESTATE, the refinements reduce ASR by 10.00–16.11%p in the fixed-query evaluations above, while ORR remains unchanged or increases by at most 1.76%p.

ORR (%)Always IARESTATE IAIntent confirmationWider band
WildJailbreak29.8010.2011.9611.57
XSTest0.000.801.200.80

These findings identify concrete refinement targets and an adjustable deliberation trade-off. We have included these analyses in the revised Appendix E.6.

Reviewer_AQBoOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RESTATE, a selective safety defense for LLMs. A linear probe (the "anchor") reads the last-token hidden state of the request and routes it into one of three paths. Requests the probe scores as confidently harmful are refused directly; confidently benign requests are answered normally. Uncertain requests get a single learned "intent token" appended, with the LLM and probe both frozen, and the same probe re-scores the new hidden state. Only requests that remain uncertain after this second look are sent to an inference-time deliberation defense. Both uncertainty bands are calibrated to target false-negative and false-positive rates. On Qwen3-8B, RESTATE improves balanced accuracy over the anchor by 7.1 points on WildJailbreak and 6.0 points on seven external benchmarks. Averaged over six deliberation methods, it reduces ASR from 37.7% to 10.2% and over-refusal from 13.7% to 7.2% compared with always deliberating, while deliberating only 14.2% of requests and keeping task utility within about one point of vanilla generation.

Critical strengths
  1. Clear and well-motivated idea. Section 3.1 links each design choice to an observed failure. The model sometimes complies with requests the probe confidently flags as harmful, which justifies direct refusal. Adversarial rewrites and most probe errors sit near the decision boundary, which justifies reassessing only that region. Improving the representation the probe reads, rather than the probe's capacity, is a useful angle.

  2. Low cost and easy to deploy. The method adds only d parameters and one extra forward step, only for uncertain requests, and reuses the KV cache. The generation path is unchanged, and utility on IFBench, MATH-500 and GPQA-Diamond stays within about one point of vanilla, while always deliberating loses up to 6.4 points.

  3. Strong classification results in the main setting. On WildJailbreak, RESTATE outperforms nine probe baselines covering added capacity, multi-layer/multi-token aggregation, a fixed safety suffix, and cascades.

  4. General framework with consistent defense-level gains. RESTATE lowers ASR and over-refusal together across six quite different deliberation methods, and any defense can serve as the deliberation step. The evaluation covers seven external datasets (including over-refusal sets and real user traffic), three adaptive attacks and additional backbones, and the appendix documents calibration, the inference algorithm and per-benchmark results in detail.

Critical weaknesses
  1. The gains shrink outside the main setting.

    • On Qwen3-32B and GPT-OSS-20B, BAcc improves by only about 1.1 points, and GPT-OSS-20B AUC is essentially unchanged (91.59 → 91.63). This is no larger than the "at most 1.9 points" the paper uses to dismiss higher-capacity probes.
    • The external pool is dominated by ToxicChat (4,947 of 8,240 requests). Per dataset (Tab. 10), RESTATE is worse than the anchor on SORRY-Bench (FNR 6.9 → 9.1) and XSTest (FPR 1.6 → 3.6).
    • On external AUC, the Suffix baseline (91.86) beats RESTATE (90.76).
    • The WildJailbreak evaluation set includes unused training-split data. Adversarial variants derived from the same vanilla prompt could appear on both sides of the split.
  2. The defense-level comparison omits stronger probes. RESTATE is compared only with Anchor and Random routing. It is not shown whether the intent token still helps when a stronger probe from Tab. 1 (SIREN, Token×Layer or Suffix) is used in the same three-way routing.

  3. Calibration and deployment realism are not established.

    • A 1% error target on 400 calibration examples per class corresponds to about four errors, so the thresholds are likely noisy. The realized error rates of the confident exits on the external datasets are not reported.
    • The WildJailbreak evaluation set is 82% harmful, unlike the benign-dominated traffic the paper uses as motivation.
    • Efficiency is the main motivation, but no latency or token cost is reported, even though the six deliberation methods differ greatly in cost.
  4. The claim that the token "refines the representation itself" is weakly supported. Transfer is shown only to probes trained on the same backbone, layer and WildJailbreak data, which are likely close in direction to the anchor. This does not rule out the token simply exploiting a shared probe direction.

  5. Novelty is moderate. The method is close to a one-token learned version of the Suffix probe placed in a cascade. Natural baselines are missing, such as training the probe and token jointly or using a multi-token soft suffix.

Questions for authors
  1. How much of the WildJailbreak gain survives when vanilla prompts and their adversarial variants are split by seed, so that no seed appears in both training and evaluation?

  2. Why is the gain much smaller on Qwen3-32B and GPT-OSS-20B than on Qwen3-8B? Were the token training and calibration settings tuned on Qwen3-8B and reused unchanged?

  3. If a stronger probe (SIREN, Token×Layer or Suffix) replaces the anchor in the same three-way routing, does adding the intent token still lower ASR and over-refusal at the same deliberation rate?

8 comments
AuthorsOpenReview ↗

Response to Reviewer AQBo (1/6)

Dear Reviewer AQBo,

We sincerely appreciate your thoughtful and constructive review. We carefully considered the concerns you raised regarding (1) data disjointness and generalization across benchmarks and backbones, (2) defense-level comparisons with stronger probes, (3) calibration reliability and deployment efficiency, (4) the interpretation of representation refinement, and (5) the novelty of the learned intent token and alternative designs. Below, we address these concerns through clarifications and additional experimental evidence.


[W1, Q1] Vanilla prompt (source)-disjoint evaluation on WildJailbreak

Thank you for highlighting the importance of source-level disjointness between vanilla requests and their adversarial variants.

We clarify that the 600 WildJailbreak examples drawn from the training split for our original evaluation were already source-disjoint from all Anchor fitting, intent token training, and calibration subsets. Thus, adversarial variants derived from the same source request did not appear across these partitions.

The remaining 2,210 examples come from WildJailbreak's official evaluation split. As described in Appendix C.2 of the original WildJailbreak paper [1], its 2,000 adversarial harmful examples were generated from vanilla behaviors that do not overlap with the training vanilla harmful requests. For the remaining 210 adversarial benign examples, the official evaluation release does not provide source identifiers, preventing independent verification of source-level disjointness.

Accordingly, our original WildJailbreak improvement of +7.12%p BAcc over Anchor (Table 1) was obtained under this split construction. We have clarified these details in Appendix E.1.

[1] Jiang et al. WildTeaming at Scale: From In-the-Wild Jailbreaks to (Adversarially) Safer Language Models (NeurIPS, 2024)

AuthorsOpenReview ↗

Response to Reviewer AQBo (2/6)

[W1, Q2] Generalization across larger backbones and benchmarks

Thank you for highlighting the importance of generalization across different backbones and benchmarks. We agree that further examining performance beyond the main Qwen3-8B setting would strengthen our evaluation.

(1) Generalization across larger backbones. We first clarify that each backbone uses its own trained Anchor and intent token, with separately calibrated thresholds. Regarding the pointed smaller gains with these models, it occurs as the original cross-backbone experiments employed a restricted hyperparameter search space, primarily to establish the applicability of RESTATE across different models and scales rather than to fully optimize each backbone.

To fairly assess its effectiveness, we expanded the hyperparameter search for each backbone with the original search space, and selected configurations based on a held-out validation set. RESTATE achieves BAcc improvements of +1.86%p on Qwen3-32B and +3.69%p on GPT-OSS-20B over their respective Anchors, along with improvements in FNR and AUC (see Appendix E.4 for details).

MetricBAcc ↑FNR ↓AUC ↑
BackboneAnchorOursAnchorOursAnchorOurs
Qwen3-32B83.6785.5311.096.7892.5093.58
GPT-OSS-20B82.4786.1615.6512.3991.5992.75

These results provide further evidence that RESTATE's benefits extend across different model scales and architectures.

(2) Generalization across external benchmarks. We appreciate your observation regarding the dominance of ToxicChat (4,947 out of 8,240 requests) in the external evaluation pool. To assess whether this imbalance drives the overall improvements, we additionally evaluated all methods after excluding ToxicChat, as shown below. Here, we additionally report Avg. Rank, the mean rank across BAcc, FNR, F1, and AUC, computed over all 11 methods (lower is better).

MethodBAcc ↑FNR ↓F1 ↑AUC ↑Avg. Rank ↓
Static classifiers
Anchor80.8118.4382.1789.604.75
Bilinear79.0921.4080.2288.168.25
MLP79.6222.6480.2787.858.00
Suffix82.8822.8782.5991.623.75
Token × Layer81.8518.2682.9589.993.00
SIREN80.4916.6982.3788.674.00
Adaptive classifiers
Anchor → MLP77.7119.3379.8282.159.50
EE-MLP79.8824.3879.9887.779.25
TPC79.4721.2480.5382.618.00
EMA → Suffix82.6823.5482.2688.256.00
RESTATE (Ours)84.7117.3085.2691.541.50

Even without ToxicChat, RESTATE achieves the highest BAcc (84.71) and F1 (85.26), while ranking second in both FNR (17.30) and AUC (91.54). This yields the best average rank of 1.50, indicating consistently strong performance across all four metrics rather than relying on a single metric.

Regarding SORRY-Bench and XSTest, we acknowledge that RESTATE exhibits modest degradations on them and that Suffix achieves higher AUC on the original external pool (91.86 vs. 90.76). Nevertheless, RESTATE achieves better overall BAcc (85.20 vs. 83.02), F1 (75.53 vs. 73.87), and FNR (16.41 vs. 22.93) than Suffix. These metrics complement AUC by reflecting classification performance at the operating decision threshold, where both harmful-request detection and benign-request acceptance matter.

Overall, these additional results support RESTATE's effectiveness across different backbones and its balanced classification performance across diverse external benchmarks.

AuthorsOpenReview ↗

Response to Reviewer AQBo (3/6)

[W2, Q3] Defense-level comparison with stronger probes

Thank you for this constructive suggestion. We agree that extending the defense-level comparison to stronger probes would help assess whether intent token reassessment remains effective beyond the original Anchor.

The classifier baselines in Table 1 primarily focus on harmfulness detection rather than downstream defense decisions. To enable a defense-level comparison, we adapted Suffix, SIREN, and Token × Layer to the same three-way routing framework. Each probe was trained on the original probe-fitting data, excluding samples reserved for intent token training, and then frozen. The intent token was trained separately for each probe, with uncertainty bands independently calibrated.

To evaluate the benefit of intent token reassessment at matched deliberation rates, we compared it against a three-way routing baseline that randomly selects the same number of requests for deliberation within the probe's uncertain region, while the remaining requests follow the probe's predictions. We also report the conventional binary deliberation gate, which invokes deliberation for requests classified as harmful, using its original, unmatched deliberation rate. Results on WildJailbreak are reported as ASR / ORR (%), averaged over six downstream defenses.

First-stage probe2-way (unmatched)3-way (random)3-way + intent tokenDelib. (%)
Anchor (RESTATE)43.20 / 3.43 (70.1%)17.46 / 11.8110.20 / 7.1614.2
Suffix39.62 / 4.22 (77.5%)9.36 / 10.967.64 / 8.438.3
SIREN41.31 / 3.27 (75.4%)11.77 / 9.979.28 / 7.8113.9
Token × Layer41.48 / 3.17 (74.9%)12.92 / 9.8411.72 / 7.1715.0

At matched deliberation rates, intent token reassessment consistently reduces both ASR and ORR across all three stronger probes. Compared with random deliberation within the uncertain band, ASR decreases by 1.20–2.49%p and ORR by 2.16–2.67%p. The binary gate, in contrast, deliberates 70–78% of requests while retaining ASR around 40–43%, whereas the three-way routing with intent token requires deliberation for only 8–15% of requests.

Overall, these additional results support that RESTATE's three-way routing and intent token reassessment remain effective when integrated with stronger probe architectures, improving the safety and over-refusal trade-off at matched deliberation rates.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.