acceptodds
Under review as a conference paper at ICLR 2027

NormGuard: Evidence-to-Norm Reasoning for Dynamic Personalized Safety

Abstract

Personalized safety requires AI assistants to adapt their behavior to user-specific risks and constraints. Existing work, however, largely assumes that safety-relevant user context is available upfront or remains static. In real-world dialogues, such information is often revealed progressively across turns, requiring assistants to continuously accumulate evidence about the user. Beyond tracking this evolving context, a second challenge is determining how the observed evidence should affect the assistant’s behavior, as existing context- or profile-conditioned approaches typically leave this mapping implicit. To study these challenges under controlled information asymmetry, we introduce a dynamic personalized-safety benchmark in which safety-relevant user information is progressively revealed while the complete user profile remains hidden from the assistant. We further propose NormGuard, an evidence-to-norm framework that separates user-state estimation from behavioral constraint formation. NormGuard progressively constructs a perceived profile from observable dialogue evidence and induces structured, evidence-grounded safety norms that specify applicable constraints, priorities, and supporting evidence for response generation. Extensive experiments show that NormGuard improves personalized safety over competing baselines, while ablation studies demonstrate complementary benefits from perceived-profile tracking and explicit safety-norm guidance. These results show that personalized safety benefits from explicitly translating progressively revealed user evidence into behavioral constraints, rather than relying on profile information alone.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Personalized Safety, Safety Norms, Dynamic User Modeling, Multi-turn Dialogue, LLM Safety

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ba8tOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses “personalized safety”: AI assistants should adjust their behavior to each user’s risks and constraints, but existing methods mostly assume the user’s information is fully known upfront and never changes, whereas in real dialogues it is revealed gradually over multiple turns, and existing methods also leave unclear how observed evidence should actually change the assistant’s behavior. To address this, the authors build a dynamic benchmark in which safety-relevant user information surfaces progressively while the full profile stays hidden from the assistant, and propose the NORMGUARD framework, which continuously updates a “perceived profile” of the user from dialogue evidence while distilling that evidence into structured safety norms, with constraints, priorities, and supporting evidence, to guide response generation. Experiments show it outperforms the baselines, and ablations indicate that profile tracking and explicit norms each contribute; the conclusion is that the key to personalized safety lies in explicitly converting progressively revealed user evidence into behavioral constraints rather than relying on profile information alone.

Critical strengths

Well-chosen problem. Both observations, that user information is revealed progressively and that the mapping from evidence to behavioral constraints should be made explicit, are real issues that prior work had not addressed head-on.

Critical weaknesses

Multi-turn is a fixed script, not an interaction. User turns are generated once and frozen, never react to the evaluated method’s responses, and clarification questions go unanswered; “dynamic” reduces to “has context”. The core results are single-turn. The title and abstract emphasize cross-turn accumulation, yet the main tables are 8,039 single-turn samples, where ablating “profile tracking” is meaningless; the only real multi-turn result is one line chart, in which profile tracking contributes almost nothing. Generation, method, and judge are the same 9B model. Self-scores are clearly inflated (other judges give 3.0–3.5, it gives 4.6), and all ablations and the “near-oracle” claim rest on self-evaluation alone. The generator prompt even names the five evaluation dimensions directly.

Questions for authors

See Weekness

1 comment

We sincerely thank you for your feedback. Our responses to the key concerns follow.

Response to the Multi-Turn Setting, Profile Tracking, and Evaluation Reliability.

We clarify several aspects of our experimental design and evaluation.

(1) Dynamic Setting and Controlled User Trajectories. As defined in Section 3.1, “dynamic” refers to the progressive revelation and accumulation of safety-relevant user evidence across dialogue turns. User turns are generated from ground-truth profiles and scenario contexts, progressively disclosing safety-relevant information to reflect realistic self-disclosure. We keep user turns fixed across methods to ensure identical conversational evidence. Method-dependent user simulations would produce divergent trajectories with varying difficulty and information, reducing comparability. This design enables controlled evaluation of profile tracking and safety adaptation.

(2) Contribution of Perceived-Profile Tracking. Table 2 shows that removing perceived-profile tracking reduces the comprehensive score from 4.67 to 4.55 and personalization from 4.31 to 4.08, demonstrating its contribution even in single-turn settings. More importantly, Figure 2 shows NormGuard consistently outperforming w/o Profile across all five turns, directly supporting profile tracking as evidence accumulates. We also include a Single-Turn baseline, which processes turns independently without an evolving user representation. NormGuard outperforms Single-Turn across all five turns and other non-oracle baselines, supporting the combination of profile tracking and evidence-to-norm reasoning.

(3) Cross-Evaluator Reliability and Cross-Model Generalization. Appendix A (Table 3) evaluates 300 fixed responses using six LLM evaluators, all favoring NormGuard over Vanilla across five dimensions. To examine model-specific biases in initialization, generation, and evaluation, we further test Qwen3.5-9B, Ministral-3-8B-Instruct, and Granite-4.1-8B across all six model assignments, using distinct models at each stage.

Init.Gen.Eval.MethodRiskEmpathyPersonal.UtilityComp.
MinistralGraniteQwenVanilla2.963.822.553.112.76
NormGuard3.084.002.803.222.91
MinistralQwenGraniteVanilla4.004.633.544.254.00
NormGuard4.224.793.734.444.24
GraniteMinistralQwenVanilla2.923.992.903.372.92
NormGuard3.144.293.303.563.23
GraniteQwenMinistralVanilla3.574.143.533.623.32
NormGuard3.904.403.853.903.75
QwenMinistralGraniteVanilla3.764.543.354.123.77
NormGuard3.994.753.644.284.02
QwenGraniteMinistralVanilla3.533.923.343.583.18
NormGuard3.664.053.483.683.36

Qwen, Ministral, and Granite denote Qwen3.5-9B, Ministral-3-8B-Instruct, and Granite-4.1-8B, respectively. Risk, Empathy, Personal., Utility, and Comp. denote risk awareness, empathy, personalization, utility, and comprehensive score. All scores use a five-point Likert scale.

NormGuard outperforms Vanilla across all six configurations and all five metrics, indicating robustness to model assignments. Together with Appendix A, these results support cross-model consistency.

We will include these results in the revision.

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

The paper studies personalized safety when the user reveals safety-relevant attributes gradually across turns. It contributes a synthetic benchmark built from PENGUIN domains × NormBank attributes (8,039 one-turn samples, 1,000 five-turn dialogues) and NormGuard, a prompting pipeline based safeguard: profile tracking → norm induction (with priority and evidence) → norm filtering → norm-conditioned generation. With Qwen3.5-9B as the backbone, NormGuard scores 4.67 Comp. against 4.15 for Vanilla (Table 1). This is close to the full ground-truth-profile reference (4.81) and above RAISE (4.26). Ablations credit most of the gain to the norm layer.

Critical strengths
  1. A well-controlled multi-turn evaluation setting. Safety-relevant attributes are revealed progressively over five-turn dialogues, while the full profile P* stays hidden from the assistant and is used only for data synthesis and evaluation. User turns are fixed and reused across methods, so comparisons are clean.
  2. The design is well controlled: the full profile P* is hidden from the assistant and used only for data synthesis and evaluation, and the user turns are fixed and reused across methods.
  3. The evidence → profile → norm → response decomposition is interpretable. Each norm comes with a priority and a cited source of evidence, which makes it auditable.
Critical weaknesses
  1. Self-judging. Qwen3.5-9B builds the data (the author indicates that all LLM components use Qwen3.5-9B “unless otherwise specified,”), runs the method, and is the default judge. It is far more lenient than other judges (Vanilla gets 4.15 vs. about 2.9–3.3; Table 3).
  2. The method gaps are just 0.1–0.4 --- it seems not significant based on the intervals shown in Table 1 and Table 2.
  3. Over-refusal is not measured. Several showcased responses are refusals or redirections (Fig. 4, C.1, C.2). There is no benign control set.
  4. Narrow evaluation scope. The eval data and the methods is test on only one week 9B backbone, with no frontier models. We don't know how the frontier models perform on the dataset and whether the proposed method is effective on the frontier models.
  5. NormGuard runs at least four LLM operations per turn (signal collection, norm induction, norm filtering, response generation; Sec. 4), and no cost or latency is reported. The baselines aren’t compute-matched.
  6. Short conversation horizon. The “dynamic” setting is at most five turns within a single session. Real personalized assistants accumulate evidence over long histories and across sessions, where the hard problems appear: evidence spread far apart, stale or contradicted attributes, and memory limits. Prior work studies exactly this long-horizon setting (PerMemSafe). Whether NormGuard’s turn-by-turn profile and norm updates scale to it is untested.
Questions for authors

none

4 comments
AuthorsOpenReview ↗

Response to Reviewer ETMv (Part 1)

We sincerely thank you for your feedback. Our responses to the key concerns follow.

Q1: Self-judging concern.

R1: Appendix A (Table 3) examines NormGuard using six LLM evaluators from different model families: Qwen3.5-9B, Qwen3.8-Max, Kimi K3, GLM-5.2, DeepSeek-V4-Pro, and Gemini3.8-Flash. Despite differences in absolute scoring scales, all six evaluators consistently favor NormGuard over Vanilla across all five evaluation dimensions. These results demonstrate that the observed improvements are not solely attributable to the evaluation preferences of Qwen3.5-9B.

Additional cross-model pipeline evaluation. We test all six assignments of Qwen3.5-9B, Ministral-3-8B-Instruct, and Granite-4.1-8B to initialization, response generation, and evaluation, with a different model at each stage.

Init.Gen.Eval.MethodRiskEmpathyPersonal.UtilityComp.
MinistralGraniteQwenVanilla2.963.822.553.112.76
NormGuard3.084.002.803.222.91
MinistralQwenGraniteVanilla4.004.633.544.254.00
NormGuard4.224.793.734.444.24
GraniteMinistralQwenVanilla2.923.992.903.372.92
NormGuard3.144.293.303.563.23
GraniteQwenMinistralVanilla3.574.143.533.623.32
NormGuard3.904.403.853.903.75
QwenMinistralGraniteVanilla3.764.543.354.123.77
NormGuard3.994.753.644.284.02
QwenGraniteMinistralVanilla3.533.923.343.583.18
NormGuard3.664.053.483.683.36

Qwen, Ministral, and Granite denote Qwen3.5-9B, Ministral-3-8B-Instruct, and Granite-4.1-8B, respectively. Risk, Empathy, Personal., Utility, and Comp. denote risk awareness, empathy, personalization, utility, and comprehensive score. All scores use a five-point Likert scale.

NormGuard outperforms Vanilla in all six configurations and four dimensions, supporting robustness to model assignments. We will add these results to the revision.


Q2: Over-refusal and benign control evaluation.

R2: NormGuard aims to improve context-sensitive safety rather than indiscriminately increase refusals. Its evidence-to-norm reasoning translates user-specific risks into actionable behavioral guidance, including avoiding inappropriate recommendations, providing safer alternatives, and preserving helpful responses when restrictive intervention is unnecessary.

Importantly, our evaluation includes Utility, which assesses whether responses remain clear, relevant, and actionable under applicable safety constraints. As shown in Table 2, NormGuard improves Utility from 4.22 (Vanilla) to 4.58, alongside Risk Sensitivity (4.29 → 4.68) and Comprehensive Score (4.15 → 4.67). These concurrent improvements provide evidence that NormGuard enhances safety while maintaining practical helpfulness, rather than merely adopting more restrictive responses.

We will further clarify this distinction in the revised manuscript and consider dedicated over-refusal and benign-helpfulness evaluations in future work.


Q3: Narrow evaluation scope and frontier models.

R3: As reported in Appendix A (Table 3), our evaluation already includes six LLM evaluators from different model families, including larger and more capable models such as Qwen3.8-Max, Kimi K3, GLM-5.2, DeepSeek-V4-Pro, and Gemini-3.8-Flash. Despite differences in their absolute scoring scales, all six evaluators consistently favor NormGuard over Vanilla across all five evaluation dimensions, suggesting that the observed improvements are not specific to the default Qwen3.5-9B evaluator.

Beyond evaluator diversity, the six cross-model pipeline assignments in R1 vary initialization, generation, and evaluation models independently. NormGuard exceeds Vanilla across all four reported dimensions in every configuration. This tests sensitivity to the generation backbone and judge. We will include these experiments and clarify the scope of our claims.

AuthorsOpenReview ↗

Response to Reviewer ETMv (Part 2)

Q4: Performance margins and statistical significance.

R4: We conduct two-sided paired t-tests on 8,039 matched samples, reporting score differences and 95% confidence intervals:

ComparisonΔ Comp.95% CIp-value
vs. Vanilla0.522[+0.501, +0.543]< 10⁻³⁰⁰
vs. w/o Profile0.128[+0.111, +0.144]4.65 × 10⁻⁵²
vs. w/o Safety Norms0.438[+0.416, +0.459]< 10⁻³⁰⁰
vs. Explicit Profile−0.050[−0.066, −0.035]2.83 × 10⁻¹⁰

Δ Comp. = NormGuard minus baseline; paired sample-level tests.

All differences are significant, with safety norms (+0.438) contributing more than profile tracking (+0.128). RAISE accesses about five ground-truth attributes via MCTS; RAISE + Norm adds safety norms; Explicit Profile accesses nine attributes. NormGuard uses dialogue evidence alone, scoring 4.67 versus RAISE (4.26), RAISE + Norm (4.60), and Explicit Profile (4.72).

Human evaluation (300 samples, 10 annotators) confirms the gains: NormGuard receives 77.8% of individual votes and wins 84.7% by sample majority (95% CI: [80.2%, 88.3%], p < 0.001). Gains are statistically reliable, though not all practically large. We will include these analyses.


Q5: Inference efficiency, computational costs, and performance gains.

R5: Single-call Supervised Fine-Tuning (SFT). We train an SFT variant for personalized safety responses in one LLM call, splitting 8,039 samples into 6,039 training, 1,000 validation, and 1,000 test samples. Both methods share the test set.

MetricNormGuardSFTΔ
Risk Sensitivity4.694.73+0.04
Emotional Empathy4.634.66+0.03
Personalization4.324.08−0.24
Utility4.564.67+0.11
Comprehensive4.674.67+0.00

Δ = SFT minus NormGuard; values rounded to two decimals.

SFT maintains overall performance but reduces Personalization by 0.24.

Inference cost analysis. Estimated per-turn costs:

MethodCalls/TurnTokens/Turn (In/Out)Latency (s)
NormGuard3–41,500–8,000 / 300–2,0009–50
NormGuard (SFT)1500–2,000 / 100–5003–12.5
Vanilla1–2500–4,000 / 100–1,0003–25
Self-Refine3–41,500–8,000 / 300–2,0009–50
SentinelMem3+k–4+k¹(3+k)×500–(4+k)×2,000 / (3+k)×100–(4+k)×5003(3+k)–12.5(4+k)

Latency assumes 1,000 input tokens/s, 50 output tokens/s, and 0.5 s overhead/call; offline evaluation excluded.

¹ kk denotes memory-state transitions per turn. SentinelMem uses three core calls plus UPDATE and optional summarization. For k=1k=1: 4–5 calls, 2,000–10,000 input tokens, 400–2,500 output tokens, and 12–62.5 s. kk is unbounded.

SFT reduces calls by 67–75% (3–4 to 1) and estimated latency from 9–50 s to 3–12.5 s with comparable quality. We will add these findings in the revision.


Q6: Conversation horizon and long-term scalability.

R6: Five turns were chosen because longer single-concern dialogues became repetitive. We extend evaluation to 10, 15, and 20 turns:

TurnsRisk SensitivityEmotional EmpathyPersonalizationUtilityComprehensive
54.64 ± 0.584.40 ± 0.494.16 ± 0.744.56 ± 0.644.55 ± 0.66
104.64 ± 0.584.38 ± 0.484.15 ± 0.754.57 ± 0.644.55 ± 0.65
154.63 ± 0.594.38 ± 0.474.14 ± 0.754.57 ± 0.634.54 ± 0.65
204.62 ± 0.594.37 ± 0.474.14 ± 0.744.56 ± 0.634.53 ± 0.65

Mean ± standard deviation; five-point Likert scale.

Performance remains stable across 5–20 turns: comprehensive scores decline only 0.02 (4.55 → 4.53), with stable Risk Sensitivity and Utility. This supports within-session robustness, not cross-session scalability. Unlike PerMemSafe/SentinelMem's memory-management focus, NormGuard converts accumulated evidence into safety norms. Future work will address conflicting attributes, memory limits, and cross-session behavior.

Thank you for the detailed response. The cross-judge and cross-model experiments show that the gains over Vanilla and the ablations do not depend on the Qwen3.5-9B self-judge, and the paired t-tests resolve the significance question. However, I will keep my rating:

  1. Over-refusal is still not evaluated. The LLM-judged Utility score cannot replace a benign (over-refuse) control set.
  2. Frontier models are still not tested. The new experiments vary the judges, but all generation models are 8–9B.
  3. The contribution is limited. The response acknowledges that the progressive-disclosure setting is not new, and a single-call SFT variant matches NormGuard’s comprehensive score (4.669 vs. 4.667).
  4. The comparisons with RAISE, RAISE + Norm, Self-refine, Explicit Profile, the ground-truth-profile reference, and SentinelMem were never re-judged by an independent model. The claims of outperforming RAISE and SentinelMem therefore still rest on the self-judge alone.
AuthorsOpenReview ↗

Response to Reviewer ETMv

Thanks for the reviewer' s feedback. First, regarding over-refusal, the improvements in Utility provide preliminary evidence that NormGuard enhances response usefulness rather than inducing excessive refusal. The Risk Awareness scores further support its ability to appropriately address safety-related concerns while maintaining utility.

Second, frontier models are included in the evaluation stage, as reported in Table 3. Given the large number of possible model configurations across the initialization, response generation, and evaluation stages, we primarily use 8B-scale models to ensure computational feasibility and systematic comparisons.

Third, RAISE, RAISE + Norm, Self-Refine, Explicit Profile, and Ground-Truth Profile have already been included in our main experiments. The additional cross-LLM experiments are specifically designed to examine the effectiveness and robustness of our proposed components across different model configurations. Therefore, we focus on Vanilla, w/o Profile, w/o Safety Norms, and NormGuard, which allow us to directly assess the contributions of perceived-profile tracking and norm awareness. While including all other baselines is feasible, doing so would substantially increase the experimental complexity and presentation length without directly serving the primary objective of these experiments. We therefore retain the most relevant comparisons in the cross-LLM evaluation. We appreciate the reviewer's understanding.

Reviewer_66P2OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1(−1)
Read review
Paper summary

This work addresses the problem of safety of AI-driven user interactions in the context of personalized dynamic safety. The distinction that this work draws from other works in this space is that normally user profiles are assumed to be static and revealed to the AI assistant from the start of the interaction, whereas this paper takes a more dynamic approach to this where. safety-relevant user information is incrementally revealed throughout the multiple turns of the interaction. The main contributions of this work are, 1) a personalized and dynamic safety benchmark built primarily from incorporating PENGUIN's 7 high-risk domains with the attribute branch of NormBank's SCENE taxonomy, where user turns are generated leveraging the ground-truth profile and kept fixed across different generation methods such that only assistant turns vary, 2) they propose a framework, NormGuard, the aims at decoupling user-state estimation from behavioral constraint formation into 3 stages: perceived-profile tracking, evidence-to-norm reasoning, and nom-guided response generation.

NormGuard uses an LLM-Judge based framework (using Likert scores) across various dimension (Comprehensive, Risk, Empathy, Personalization, Utility) (but using the same LLM backbone for NormGuard and the evaluator). Experiments are conducted on single-turn for comparing profile-discovery strategies, multi-turn, and human evaluation, with accompanying ablations. The main result of the paper shows NormGard with an inferred profile achieving a Comprehensive score of 4.67 compared to the strongest profile-discovery baseline of 4.26 and the ground-truth profile reference (4.81).

Critical strengths
  1. The problem is well-defined and the following discussions, synthetic data generation, evaluations, and ablations are rather well-motivated. The results show that the dynamic revelation framing gains over static personalized safety profile approach. Ablations also do a decent job of isolating different effects, for example, Self-Refine control differentiates norm structure from extra inference-time compute, and the held-out non-norm-first construction shows whether the improvements arise from the norm-first benchmark pipeline.

  2. The different steps in NormGuard are explicit, auditable, and well-motivated.

Critical weaknesses
  1. The main limitation of this work is that one model is used to construction, simulation, and evaluation, namely Qwen3.5-9B. Therefore, all numbers and conclusions derived from the experiments rests on how this LLM backbone performs given the different scenarios and instructions. Tables 1-2 show a ceiling performance of 4.81 and all following deltas are, therefore, based on this single model grading its own generations. In the Appendix Table 3, we are shown how NormGuard's values differ compared to Vanilla framework by small margins across different backbones, but none of the other approaches for comparison are included (e.g. RAISE + Norm,,, Explicit profile, and the other ablations).

  2. The margins between the different approaches on the same backbone LLM as shown in Tables 1 and 2 are very small: 4.67 NormGuard, 4.6 RAISE + Norm, 4.72 Explicit profile. These are all within less than 1.0 standard deviation, and all attributed to the top of the Likert scale (1.5 scale) with no significance tests or independent judge provided. Given these values, the quantitative utility of NormGuard seems to be oversold and not load-bearing enough for the paper's claims.

  3. Crucially, NormGaurd involves four to five sequential LLM calls per assistant turn. This is a substantial latency cost for the setting of user-interfaced, high-stakes AI assistants. No latency, or token-count, or throughput measurements are provided. Also, numbers of turns and how much Likert score is gained across the different approaches are not clear, either.

  4. The progressively-revealed dynamic implicit-evidence setting is very close to some of the prior work, notably CURATe and PerMemSafe/SentinelMem. And SentinelMem is only included as a baseline in passing in Figure 2.

Questions for authors

In addition to the central gaps outlined in the critical weaknesses, here are a few other questions/comments that would be great to be followed up with:

  1. The full comparison between NormGuard and at least two other independent evaluators should be added to Tables 1 and 2 (not just Vanilla as in Table 3).This will be insightful to see whether the close margins across the board are directional or statistically significant when a different judge LLM (pother than the one used as the generation backbone) is used.

  2. Please report latency costs for both per-turn and per-dialogue scenarios, as well as token counts for NormGuard vs Vanilla, Self-Refine, and SentinelMem. This is absolutely crucial to put the Likert score numbers in the context of latency trade-off.

  3. Multi-turn results provided in Figure 2 that include SentinelMem and Single-Turn can be better provided as a comprehensive table with per-turn means, variance, and significance tests. This will hugely help clarify how NormGuard compares with SentinelMem which is the closes prior method NormGuard builds on.

  4. From the results, it is not clear whether the gains in NormGuard (and if so, to what extent) are attributed to increased over-refusal. In the absence of any overrefusal, benign-helpfulness control the Risk/Comprehensive gains cannot be put into perspective.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 66P2 (Part 1)

We sincerely thank you for your feedback. Our responses are as follows.

Q1: Single-LLM dependency and cross-backbone evaluation.

R1: Appendix A (Table 3) evaluates NormGuard using six LLM judges (Qwen3.5-9B, Qwen3.8-Max, Kimi K3, GLM-5.2, DeepSeek-V4-Pro, and Gemini3.8-Flash). All six favor NormGuard over Vanilla across five dimensions, despite different scoring scales, suggesting that the improvements are not specific to the Qwen evaluator.

Additional cross-model evaluation. We further evaluate all six assignments of three LLMs (Qwen3.5-9B, Ministral-3-8B-Instruct, and Granite-4.1-8B) to data initialization, response generation, and evaluation, assigning distinct models to each stage.

InitializationGenerationEvaluationMethodRisk awarenessEmpathyPersonalizationUtilityComprehensive score
MinistralGraniteQwenVanilla2.963.822.553.112.76
NormGuard3.084.002.803.222.91
MinistralQwenGraniteVanilla4.004.633.544.254.00
NormGuard4.224.793.734.444.24
GraniteMinistralQwenVanilla2.923.992.903.372.92
NormGuard3.144.293.303.563.23
GraniteQwenMinistralVanilla3.574.143.533.623.32
NormGuard3.904.403.853.903.75
QwenMinistralGraniteVanilla3.764.543.354.123.77
NormGuard3.994.753.644.284.02
QwenGraniteMinistralVanilla3.533.923.343.583.18
NormGuard3.664.053.483.683.36

Qwen, Ministral, and Granite denote Qwen3.5-9B, Ministral-3-8B-Instruct, and Granite-4.1-8B, respectively. All scores are measured on a five-point Likert scale.

NormGuard consistently outperforms Vanilla across all six model assignments, improving all five metrics.

Baseline clarification. RAISE + Norm and Explicit Profile are auxiliary references accessing approximately five and nine ground-truth attributes, respectively. NormGuard instead infers profiles from dialogue evidence alone. We therefore focus cross-model comparisons on Vanilla and NormGuard to assess effectiveness without privileged profile access. We will clarify these settings and include the results in the revision.


Q2: Small performance margins and statistical significance.

R2: Performance margins should account for differences in profile access. RAISE retrieves approximately five ground-truth attributes using MCTS; RAISE + Norm additionally incorporates safety norms; and Explicit Profile accesses nine ground-truth attributes. NormGuard relies solely on dialogue evidence without accessing ground-truth profile attributes, yet achieves a comprehensive score of 4.67, exceeding RAISE (4.26) and RAISE + Norm (4.60), and approaching Explicit Profile (4.72).

Paired statistical significance. We conduct two-sided paired t-tests on 8,039 matched samples:

ComparisonΔ Comp.95% CIp-value
vs. Vanilla0.522[+0.501, +0.543]< 10⁻³⁰⁰
vs. w/o Profile0.128[+0.111, +0.144]4.65 × 10⁻⁵²
vs. w/o Safety Norms0.438[+0.416, +0.459]< 10⁻³⁰⁰
vs. Explicit Profile−0.050[−0.066, −0.035]2.83 × 10⁻¹⁰

Δ Comp. denotes NormGuard minus the comparison method.

All differences are significant, confirming the contributions of profile tracking (+0.128) and safety norms (+0.438). The standard deviations in Tables 1–2 do not directly determine paired significance, which accounts for matched samples.

Independent human evaluation. Across 300 samples and 10 annotators, NormGuard receives 77.8% of individual preference votes over Vanilla and wins 84.7% of samples by majority vote (95% CI: [80.2%, 88.3%], p < 0.001). These results independently support the improvements beyond automatic evaluation.

We will include these statistical analyses in the revision.

AuthorsOpenReview ↗

Response to Reviewer 66P2 (Part 2)

Q3: Inference efficiency, computational costs, and performance gains.

R3: Single-call Supervised Fine-Tuning (SFT). We explore an SFT variant that generates personalized safety responses directly from dialogue histories and user queries in a single LLM call. We split 8,039 samples into 6,039 training, 1,000 validation, and 1,000 test samples, evaluating both methods on the same test set.

MetricNormGuardSFTΔ
Risk Sensitivity4.6874.730+0.043
Emotional Empathy4.6264.660+0.034
Personalization4.3204.077−0.243
Utility4.5574.665+0.108
Comprehensive4.6674.669+0.002

Δ denotes SFT minus NormGuard.

SFT maintains comparable overall performance (4.669 vs. 4.667), improving Risk Sensitivity, Empathy, and Utility, but reducing Personalization (−0.243).

Inference cost analysis. We further estimate per-turn computational costs:

MethodCalls/TurnTokens/Turn (In/Out)Latency (s)
NormGuard3–41,500–8,000 / 300–2,0009–50
NormGuard (SFT)1500–2,000 / 100–5003–12.5
Vanilla1–2500–4,000 / 100–1,0003–25
Self-Refine3–41,500–8,000 / 300–2,0009–50
SentinelMem3+k–4+k¹(3+k)×500–(4+k)×2,000 / (3+k)×100–(4+k)×5003(3+k)–12.5(4+k)

Latency estimates assume 1,000 input tokens/s, 50 output tokens/s, and 0.5 s overhead per call, excluding offline evaluation.

¹ k denotes SentinelMem's memory-state transitions per turn. Three core calls handle extraction, profile updating, and response generation; each UPDATE adds one call, while optional summarization may add another. Estimated latency is 9–50 s for k = 0 and 12–62.5 s for k = 1; k is not capped at one.

SFT reduces sequential calls by 67–75% (3–4 to 1) and estimated latency from 9–50 s to 3–12.5 s, while maintaining comparable comprehensive performance. This suggests a promising efficiency–personalization trade-off. We will include these results in the revision.


Q4: Comparisons with CURATe and PerMemSafe/SentinelMem.

R4: We acknowledge that progressively revealed safety-relevant information has been studied previously. In NormGuard, this primarily serves as an evaluation setting. Our methodological contribution is evidence-to-norm reasoning, connecting evolving perceived profiles to actionable, context-dependent safety constraints.

Methodological distinctions. CURATe evaluates retention and use of previously disclosed safety-critical information, while PerMemSafe examines implicit and evolving personalized risks. Both primarily provide evaluation benchmarks. SentinelMem focuses on risk-aware memory management and personalized risk inference. In contrast, NormGuard induces, filters, and prioritizes safety norms from dialogue evidence, translating perceived user characteristics into explicit behavioral constraints for response generation.

Baseline comparison. We therefore compare SentinelMem in the multi-turn evaluation (Figure 2), rather than the single-turn profile-discovery comparison (Table 1). NormGuard outperforms SentinelMem across all five turns by +0.275 to +0.458 comprehensive points, with Holm-adjusted p < 0.05 throughout.

We will clarify these methodological distinctions and baseline choices in the revision.


Q5: Over-refusal and benign helpfulness.

R5: NormGuard targets context-sensitive safety rather than indiscriminate refusal. Evidence-to-norm reasoning guides safer recommendations while preserving helpfulness when restrictive intervention is unnecessary.

Our evaluation includes Utility, measuring clarity, relevance, and actionability under safety constraints. NormGuard improves Utility from 4.22 to 4.58, alongside Risk Sensitivity (4.29 → 4.68) and Comprehensive Score (4.15 → 4.67). These concurrent gains suggest improved safety without sacrificing practical helpfulness.

We will clarify this distinction and consider dedicated benign-helpfulness and over-refusal evaluations in future work.

AuthorsOpenReview ↗

Response to Reviewer 66P2 (Part 3)

Q6: Cross-evaluator comparisons and statistical reliability.

R6: We extend cross-evaluator analysis to the main ablations using four independent LLM judges (Qwen3.8-Flash, GLM-5.3-Flash, DeepSeek-V4.1-Flash, and Gemini-3.8-Flash) under the same protocol.

EvaluatorMethodComp.RiskEmpathyPersonal.Utility
Qwen3.8-FlashVanilla3.123.043.513.123.09
w/o Profile3.293.333.743.273.15
w/o Norms3.253.123.863.233.17
NormGuard3.463.433.923.373.23
GLM-5.3-FlashVanilla3.263.433.413.033.18
w/o Profile3.503.683.583.203.31
w/o Norms3.373.493.763.183.23
NormGuard3.613.733.783.303.35
DeepSeek-V4.1-FlashVanilla3.233.293.452.933.20
w/o Profile3.473.603.773.093.28
w/o Norms3.363.363.863.063.22
NormGuard3.653.683.953.263.41
Gemini-3.8-FlashVanilla2.983.233.182.723.02
w/o Profile3.323.623.512.973.29
w/o Norms3.203.333.682.953.17
NormGuard3.493.773.763.153.41

All scores use a five-point Likert scale; bold indicates the highest score per evaluator and dimension.

NormGuard ranks highest across all five dimensions under all four evaluators, improving comprehensive scores over Vanilla by +0.34–0.51, w/o Profile by +0.11–0.17, and w/o Norms by +0.21–0.29. These consistent gains support both components' contributions beyond the original evaluator's preferences.

We will add these results to the revision.


Q7: Multi-turn performance and statistical significance.

R7: We provide turn-level comprehensive scores, variances, and Holm-adjusted paired significance for NormGuard, SentinelMem, and Single-Turn over five interaction turns.

TurnMethodMeanVarianceΔ Comp.p (Holm)
1NormGuard4.6630.480——
SentinelMem4.2051.0940.4581.24 × 10⁻³⁵
Single-Turn4.6170.4970.0460.0535
2NormGuard4.5710.625——
SentinelMem4.2601.0740.3111.14 × 10⁻¹⁷
Single-Turn4.3460.8980.2262.17 × 10⁻¹⁰
3NormGuard4.5850.672——
SentinelMem4.2851.1250.3007.74 × 10⁻¹⁵
Single-Turn4.3040.9870.2818.11 × 10⁻¹⁴
4NormGuard4.6480.500——
SentinelMem4.3640.9460.2845.90 × 10⁻¹⁵
Single-Turn4.3920.8830.2562.55 × 10⁻¹³
5NormGuard4.6310.606——
SentinelMem4.3560.9710.2757.74 × 10⁻¹⁵
Single-Turn4.3671.0240.2642.44 × 10⁻¹¹

Δ Comp. = NormGuard minus baseline. Two-sided paired t-tests with Holm correction across ten comparisons.

NormGuard significantly outperforms SentinelMem at every turn (+0.275 to +0.458). Against Single-Turn, improvements are significant at Turns 2–5 (+0.226 to +0.281), but not Turn 1 (+0.046, p = 0.0535), supporting the benefit of cross-turn evidence accumulation.

Across five turns, NormGuard averages 4.620, versus 4.294 for SentinelMem and 4.405 for Single-Turn. We will include these results in the revision.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.