acceptodds
Under review as a conference paper at ICLR 2027

VEPA: Verified Evolving Formal Policy for Lifelong LLM Agent Guardrails

Abstract

LLM agents increasingly operate over sensitive resources, including private documents and mutable system state, creating substantial safety and security risks. Runtime guardrails should therefore provide inspectable, reproducible, and deterministic allow or refuse decisions before potentially irreversible actions are executed. Yet safety and security policies are rarely complete before deployment, requiring guardrails to continually adapt as new risks emerge. Existing adaptive guardrails learn from deployment experience, but typically apply safety policies through stochastic LLM inference, which can produce inconsistent and ambiguous decisions or policy interpretations. Conversely, formal guardrails provide deterministic enforcement, but typically rely on pre-deployment or task-specific policies and lack an adaptation mechanism for newly observed risks. We introduce VEPA, a framework for verified evolution of formal agent guardrails. VEPA represents the guardrail as a formal policy of executable symbolic rules over typed agent events. During deployment, we use an LLM guided by deployment feedback to propose policy patches that repair newly observed failures while preserving previously correct behavior. A deterministic verifier then admits only structurally valid, conflict-free policies. An evidence-aware authority layer subsequently activates verified rules with sufficient supporting evidence, preventing uncertain policy updates from directly affecting guardrail decisions. Empirically, across ASSEBench-Safety, ASSEBench-Security, and ATBench, VEPA consistently outperforms other baselines while progressively expanding the scope of deterministic formal-policy enforcement. These gains persist under other deployment conditions such as mixed streams, corrupted feedback, and longer horizons.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Agent security, agent safety, policy, verification

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_XpgGOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes VEPA, a technique to evolve formal policy at runtime, based on feedback about failures. VEPA also determines the appropriate time to activate or deactivate a newly proposed formal policy rule, based on runtime evidence.

Critical strengths

The paper aims to tackle an interesting problem involving LLM agents and automatically evolving formal policies for agents. It also distinguishes between accepting newly generated formal policy rules and allowing them to take effect at runtime.

Critical weaknesses

The name of the technique (and the paper as a whole) claims that the evolution of formal policy is "verified", but Sec. 3 describes an approach to ensure that the LLM-proposed patch for the policy is checked to be "structurally valid, conflict-free, and performs well on the observed evidence". The proposed technique does not offer a guarantee of correctness with respect to the intended safety policy. Using the term "verification" to describe such an approach gives an impression of a rigorous proof of correctness that VEPA does not offer, especially in the context of formal policy enforcement and guardrails.

The paper refers to a set of failures as "examples that are handled incorrectly", but does not provide a clear definition of what it means to handle something incorrectly. Does a failure refer to only incorrect ALLOW verdicts? Does it include incorrect REFUSE verdicts? Or does it also include policies that abstained when an intervention was necessary? The paper does not adequately discuss the differences between the above kinds of errors and whether each of them need different fixing strategies, which makes it difficult to understand the problem being targeted.

It is not clear what guarantee the LLM-proposed patches are associated with (e.g., an LLM-proposed patch could fit the limited collected evidence perfectly well, while rewriting the policy into something significantly different from the true policy that must be enforced.) The expressivity of the DSL seems restricted; The rules seem to support only conjunctions of predicates and the operators "in" and "not_in". What if the failure is not expressible in the DSL? Does the framework repeatedly try to incorporate failures or recognize that something is not expressible after a few tries? A justification for the choice of DSL constructs would shed more light on the appropriate usage scenarios for VEPA.

VEPA takes in a labeled sequence of events to be given as input to propose candidate patches, but the paper does not discuss how realistic this prerequisite is, let alone suggest a technique for obtaining them. The paper presents evidence for sparse labelling settings, but it does not discuss how one should go about obtaining authoritative labels in any case.

The evaluation uses recorded pre-determined agent trajectories, which does not accurately match a realistic deployment, where a guardrail's actions would affect the trajectory of an agent. This evaluation setting, while useful, does not provide insight into VEPA's performance as a runtime technique. The authors mention that to maintain uniformity across different experimental settings, they make this choice. An alternative way to handle this would have been to use the guardrails at runtime, and repeat the experiments for statistical confidence measures.

The evaluation does not sufficiently justify the claims made earlier in the paper, e.g., long horizon. To evaluate the effectiveness of VEPA in a long horizon setting, one would use it on a long horizon task as a runtime guardrail, and learn new rules as the agent performs hundreds if not thousands of tool calls, and repeat experiments for statistical significance. But in the evaluation, the paper presents a larger version of the 10 deployment round experiment by scaling it up to 30 deployment rounds.

Questions for authors
  • How were the DSL constructs chosen? What class of failures can it capture (or not capture)?

  • What kinds of errors might the LLM-driven parser make (the parser that populates the trace with action information)? How often might such errors be observed in realistic deployments?

  • What would be the performance of VEPA in a runtime enforcement setting, where it is actively modifying the trajectory of an agent?

5 comments
AuthorsOpenReview ↗

Response to Reviewer XpgG [Part 1/3]

We thank the reviewer for the constructive and detailed review of our work. We are glad that you found the problem of automatically evolving formal policies for LLM agents interesting and recognized the importance of separating the acceptance of generated rules from their deployment authority. We provide the responses to your comments below.


[W1] Verified policy

We would like to clarify the meaning and importance of verification in VEPA. First, our use of "verified" follows the standard definition in formal methods, where verification involves mechanically checking whether an artifact satisfies explicitly specified properties. Specifically, VEPA formally verifies two properties: well-formedness under the DSL and conflict-freedom. We also note that previous work on formal agent guardrails uses verification in a similar sense [1, 2].

Second, we emphasize that resolving inconsistencies and conflicts in policy is especially an important problem for agentic security, as highlighted by recent work [3]. In our settings, this verification is also particularly important for evolving agent guardrails, where continually incorporating new rules based on deployment feedback can introduce policy conflicts and inconsistencies over time, potentially leading to policy regressions (we provide an analysis in Appendix C.2). Unlike natural-language policy approaches, which rely on LLM to make decisions and provide no effective ways for guaranteeing conflict-freedom, VEPA provides deterministic, mechanically checkable guarantees that these structural properties are preserved throughout policy evolution.


[W2] Set of failures

The failure set includes all three cases: (1) incorrect ALLOW (false negative), where the policy allows an unsafe action, (2) incorrect REFUSE (false positive), where the policy blocks a legitimate action, and (3) abstention, where the policy fails to provide a formal decision, including cases where an intervention is necessary. Specifically, in VEPA, an example is considered correctly handled only when the formal policy produces the ground-truth ALLOW or REFUSE decision. Otherwise, it is included in the failure set.

Our policy evolution framework aims to address all these failure types. The failure-repair objective encourages policy updates that correct wrong decisions and expand coverage, while the regression-avoidance objective preserves previously correct behavior. Different failures may require different policy revisions, such as adding rules to cover abstentions or refining existing rules to correct incorrect ALLOW/REFUSE decisions. We describe different operators in lines 229-233 of our manuscript.


[W3 + Q1] LLM patch and DSL

Yes, your observation that “an LLM-proposed patch could fit the limited collected evidence perfectly well, while rewriting the policy into something significantly different from the true policy that must be enforced” is correct. This is why we formulate policy evolution as a constrained bi-objective optimization problem, balancing the repair of newly observed failures against regressions on previously correct examples. VEPA searches for policy updates that correct failures while preserving established behavior. Moreover, our evidence-aware authority mechanism only accepts rules with sufficient evidence that must be enforced.

Regarding your question on DSL, we construct our DSL based on contextual attributes relevant to agent safety and security, motivated by the contextual integrity framework and common properties of agent environments, including properties such as actions, data, and tools. Our DSL supports conjunctions of predicates, while multiple rules can express disjunctions, which allows it to capture safety and security requirements such as unauthorized access, data leakage, and unsafe tool use. In our paper, we focus on evolving formal policies within specified environments with the fixed event schema. Therefore, failures requiring new environment properties or previously unsupported tools may not be expressible within the current DSL. We acknowledge this limitation in Section F of our manuscript. We view extending VEPA to dynamically evolve the event schema and policy language as a potential direction for future work, and complementary to the policy-evolution problem studied in this paper.


[W4] Labels for events

We follow the deployment-feedback setting studied in prior adaptive guardrails, particularly LiSA [4], where users provide corrective feedback on guardrail decisions. Learning from corrective feedback is also studied in related work, such as AGrail [5], which studies adapting agent policies or safety checks from experience. Moreover, we believe that collecting user feedback and corrections is a common practice for deployed AI systems [6, 7].

AuthorsOpenReview ↗

Response to Reviewer XpgG [Part 2/3]

[Q2] Evaluation of the LLM parser

We add two analyses to evaluate the reliability of the LLM parser and the impact on downstream guardrail performance. These experiments with more detailed settings are presented in the new Section 5.3 of the updated PDF.

Parsing Accuracy. First, we evaluate structured event extraction using three stronger LLMs as independent reference judges: Gemini-3.1-Pro, Gemini-3.8-Flash, and Claude-Opus-5. Each judge receives the original agent trajectories and the same DSL schema as VEPA's parser (Gemini-3.1-Flash-Lite), and independently extracts structured events.

We report three metrics: Field Agreement, which measures agreement across all schema fields; Critical-Field Agreement, which measures agreement on fields whose values can affect formal-policy decisions; and Decision Agreement, which measures whether the parser and reference judge produce identical formal-policy decisions. We additionally construct a three-judge consensus through field-level majority voting.

LLM JudgeField Agree. (%) ↑Critical Agree. (%) ↑Decision Agree. (%) ↑
Gemini-3.1-Pro94.294.697.8
Gemini-3.8-Flash92.893.996.5
Claude-Opus-593.792.597.6
3-Judge Consensus93.993.497.6

The table above shows that VEPA's lightweight parser achieves 92.8-94.2% field agreement and 92.5-94.6% critical-field agreement across the three independent judges. Moreover, decision agreement is 96.5-97.8%, showing that most field-level disagreements do not alter downstream formal-policy decisions. Agreement with the three-judge consensus is similarly high. The consistently high agreement across different reference models suggests that the lightweight parser provides a consistent and reliable structured representation for downstream formal-policy enforcement.

Robustness to Parsing Errors. Second, we study how parsing errors affect VEPA's end-to-end performance. We introduce controlled errors into structured events at rates of ρ∈{0,0.05,0.1,0.2}\rho \in \{0, 0.05, 0.1, 0.2\}. For each nonzero error rate, we randomly select the corresponding fraction of events and replace one randomly selected policy-relevant field with a different valid value from its DSL vocabulary. This simulates incorrect attribute extraction while preserving the structural validity of event representations. We rerun the online deployment protocol with corrupted events with the same experimental settings.

Parsing Error RateF1 ↑
0%0.86 ± 0.02
5%0.86 ± 0.04
10%0.85 ± 0.03
20%0.82 ± 0.04

Under moderate corruption (5%-10%), VEPA maintains comparable performance. Together with the high parser agreement in our experiment above, where field-level disagreement with independent LLM judges is below 8%, these results suggest that VEPA is relatively insensitive to moderate extraction errors. Even at a 20% corruption rate, performance decreases by only 0.04 F1.

Overall, these experiments show that (1) VEPA's lightweight parser achieves consistently high agreement with independent stronger LLM judges, particularly in downstream formal-policy decisions, and (2) VEPA's end-to-end performance remains stable under moderate upstream extraction errors.

AuthorsOpenReview ↗

Response to Reviewer XpgG [Part 3/3]

[W5 + Q3] VEPA in runtime enforcement setting

First, we note that our existing evaluation on recorded agent trajectories provides a controlled setting for studying guardrail adaptation. By evaluating different methods on the same agent events and ground-truth labels, we can directly compare their ability to learn from deployment feedback. Benchmarks such as ASSEBench and ATBench are also used in other guardrail works [8, 9].

To further strengthen the evaluation, following your suggestion, we conduct an additional runtime enforcement experiment on AgentDojo [10] under benign conditions and prompt-injection attacks. We use Gemini-3.8-Flash as the agent and follow the same guardrail configurations and adaptation protocol as our main experiments. Specifically, each guardrail checks every proposed tool call before execution. Allowed actions execute normally, while rejected actions are blocked and replaced with a tool-error message, after which the agent resumes with its existing history and can revise its actions or select alternative tools. We allow at most 10 execution-loop iterations per episode and repeat the evaluation across five runs, and reset the environment between executions. We compare our method with two strongest baselines (AGrail and LiSA).

We report three metrics: Benign Utility, which measures legitimate task completion without attacks; Utility under Attack, which measures legitimate task completion in attacked episodes; and Attack Success Rate (ASR), which measures the fraction of attacks that successfully achieve their objectives.

MethodBenign Utility ↑Utility under Attack ↑ASR ↓
AGrail90.8 ± 4.468.7 ± 5.312.5 ± 3.9
LiSA87.1 ± 2.872.9 ± 4.13.2 ± 1.4
VEPA (Ours)89.2 ± 3.278.4 ± 4.62.6 ± 1.2

The table above shows that VEPA achieves the highest utility under attack (78.4%), outperforming LiSA and AGrail by 5.5% and 9.7% respectively, while maintaining comparable benign utility. VEPA also achieves the lowest ASR of 2.6%, compared with 3.2% for LiSA and 12.5% for AGrail. These results show that VEPA remains effective when actively enforcing policies and modifying agent trajectories, preventing successful attacks while allowing agents to recover from blocked actions and continue legitimate tasks.

We present these experiments with more detailed settings in the new Section C.10 of the updated PDF.


[W6] Long-horizon setting

We would like to clarify that the long-horizon setting in our paper refers to longer deployment horizons, and not long-horizon agent tasks. We have made this explicit throughout the paper, including the introduction and experimental sections by using the term "longer deployment horizons". Studying longer deployment horizons is important because guardrails must maintain safety throughout deployment of agents while adapting to new risks that may emerge over time. This is a central motivation of VEPA and the primary focus of our long-horizon evaluation.

For long-horizon agent tasks involving hundreds or thousands of tool calls, VEPA can be applied at each action, since formal-policy enforcement operates on individual structured events. Therefore, the VEPA enforcement mechanism is not limited by the number of actions or tool calls in an agent task.

We hope our responses and additional experiments have addressed your comments and clarified the strengths of our approach.


References

[1] Miculicich, Lesly, et al. "Veriguard: Enhancing llm agent safety via verified code generation." arXiv preprint arXiv:2510.05156 (2025).

[2] Shi, Tianneng, et al. "Progent: Securing AI agents with privilege control." arXiv preprint arXiv:2504.11703 (2025).

[3] Bagdasarian, Eugene, et al. "Open and Emergent Problems in Agentic Privacy and Security: A Contextual Angle." (2026).

[4] Kim, Minbeom, et al. "LiSA: Lifelong Safety Adaptation via Conservative Policy Induction." NeurIPS 2026.

[5] Luo, Weidi, et al. "Agrail: A lifelong agent guardrail with effective and adaptive safety detection." ACL 2025.

[6] Pan, Melissa Z., et al. "Measuring agents in production." ICML 2026.

[7] Fanconi, Claudio, and Mihaela van der Schaar. "Cascaded language models for cost-effective human–ai decision-making." NeurIPS 2025.

[8] Zheng, Zhijie, et al. "StepGuard: Learning Step-Level Guardrails with Scalable Supervision and Safety–Utility Balancing." EMNLP 2026.

[9] Gu, Siyi, et al. "StepSafe: Learning Timely Intervention with Context-Aware Safety Reasoning." arXiv preprint arXiv:2610.20960 (2026).

[10] Debenedetti, Edoardo, et al. "Agentdojo: A dynamic environment to evaluate prompt injection attacks and defenses for llm agents." NeurIPS 2024.

Reviewer_BN7FOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Summary

This paper introduces VEPA (Verified Evolving Formal Policy for Agent Guardrails), a framework for the lifelong online evolution of formal, executable symbolic guardrails for LLM agents. VEPA separates policy generation, verification, and runtime authority into three stages: (1) raw agent interactions are parsed into a typed domain-specific language (DSL) by a lightweight LLM; (2) an LLM proposer guided by archive-based Pareto search generates candidate policy patches that balance failure repair against non-regression while a deterministic verifier filters out malformed or conflicting rules; and (3) a Beta-Bernoulli posterior risk authority layer activates verified rules only after they accumulate sufficient empirical support. Evaluated across ASSEBench-Safety, ASSEBench-Security, and ATBench under online deployment settings, including mixed streams, noisy feedback, LLM fallback, and long horizons, VEPA consistently shows improvement over static LLMs, natural-language memory guardrails, and classical rule learners while progressively expanding the coverage of deterministic formal enforcement.

Critical strengths

Strengths

  • The paper addresses a critical gap in LLM agent safety by uniting the continuous adaptability of lifelong memory-based guardrails with the deterministic, reproducible, and low-latency runtime execution of formal symbolic policies.
  • The Bi-Objective Policy Evolution and verification procedure are well thought out. Formulating Policy Evolution as a constrained bi-objective optimization problem, solved via LLM-guided archive-based Pareto search with inverse behavioral-density sampling for diversity and hard deterministic verifications, effectively prevents the policy contradictions and behavioral regressions that plague natural-language memory systems.
  • Separating structural verification from runtime deployment authority via Beta-Bernoulli posterior risk quantiles over deterministically evaluated rule scopes, prevents rules being applied directly due to noisy observations and thus prevent premature governance of agent actions.
  • Extensive experiments across three benchmarks and challenging deployment conditions demonstrate strong F1 improvements, sub-millisecond post-parsing rule execution, and a substantially lower regression rate than competing baselines.
Critical weaknesses

Weaknesses

  • Because VEPA's determinism and conflict-freedom guarantees apply strictly after a temperature-0 LLM parses raw agent traces into a fixed categorical schema, the system remains vulnerable to upstream extraction errors or prompt injections. The manuscript could benefit from reporting isolated parser extraction accuracy.

  • Restricting the primary RQ1 and several RQ2 evaluations to events covered by VEPA's own high-confidence active policy creates a non-stationary evaluation distribution skewed toward clean, recurring event clusters; reporting full-stream evaluations (with and without LLM fallback) in Appendix, alongside active coverage across all individual benchmarks and sparse/noisy feedback settings would strengthen the empirical comparisons.

  • Minor Corrections: Several bibliography entries have uncapitalized acronyms in their titles (e.g., "ai", "llm"; "abac","nsga-ii" and "gui"). Protecting these acronyms with braces in BibTeX will ensure consistent capitalization.

Questions for authors

Please respond to the weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer BN7F

We thank the reviewer for the thoughtful review and constructive feedback. We are happy that you recognized VEPA's contribution in combining lifelong guardrail adaptation with deterministic, reproducible, and low-latency formal enforcement. We also appreciate your positive assessment of our bi-objective policy evolution, evidence-aware authority mechanism, and extensive evaluation across challenging deployment conditions. We address your comments below.


[W1] Evaluation of the LLM parser

We add two analyses to evaluate the reliability of the LLM parser and the impact on downstream guardrail performance. These experiments with more detailed settings are presented in the new Section 5.3 of the updated PDF.

Parsing Accuracy. First, we evaluate structured event extraction using three stronger LLMs as independent reference judges: Gemini-3.1-Pro, Gemini-3.8-Flash, and Claude-Opus-5. Each judge receives the original agent trajectories and the same DSL schema as VEPA's parser (Gemini-3.1-Flash-Lite), and independently extracts structured events.

We report three metrics: Field Agreement, which measures agreement across all schema fields; Critical-Field Agreement, which measures agreement on fields whose values can affect formal-policy decisions; and Decision Agreement, which measures whether the parser and reference judge produce identical formal-policy decisions. We additionally construct a three-judge consensus through field-level majority voting.

LLM JudgeField Agree. (%) ↑Critical Agree. (%) ↑Decision Agree. (%) ↑
Gemini-3.1-Pro94.294.697.8
Gemini-3.8-Flash92.893.996.5
Claude-Opus-593.792.597.6
3-Judge Consensus93.993.497.6

The table above shows that VEPA's lightweight parser achieves 92.8-94.2% field agreement and 92.5-94.6% critical-field agreement across the three independent judges. Moreover, decision agreement is 96.5-97.8%, showing that most field-level disagreements do not alter downstream formal-policy decisions. Agreement with the three-judge consensus is similarly high. The consistently high agreement across different reference models suggests that the lightweight parser provides a consistent and reliable structured representation for downstream formal-policy enforcement.

Robustness to Parsing Errors. Second, we study how parsing errors affect VEPA's end-to-end performance. We introduce controlled errors into structured events at rates of ρ∈{0,0.05,0.1,0.2}\rho \in \{0, 0.05, 0.1, 0.2\}. For each nonzero error rate, we randomly select the corresponding fraction of events and replace one randomly selected policy-relevant field with a different valid value from its DSL vocabulary. This simulates incorrect attribute extraction while preserving the structural validity of event representations. We rerun the online deployment protocol with corrupted events with the same experimental settings.

Parsing Error RateF1 ↑
0%0.86 ± 0.02
5%0.86 ± 0.04
10%0.85 ± 0.03
20%0.82 ± 0.04

Under moderate corruption (5%-10%), VEPA maintains comparable performance. Together with the high parser agreement in our experiment above, where field-level disagreement with independent LLM judges is below 8%, these results suggest that VEPA is relatively insensitive to moderate extraction errors. Even at a 20% corruption rate, performance decreases by only 0.04 F1.

Overall, these experiments show that (1) VEPA's lightweight parser achieves consistently high agreement with independent stronger LLM judges, particularly in downstream formal-policy decisions, and (2) VEPA's end-to-end performance remains stable under moderate upstream extraction errors.


[W2] Empirical comparisons

We clarify that our experiments in Section 4 separate the decision quality of formal authority from the performance of the complete system. In RQ1, all baselines are evaluated on exactly the same events covered by VEPA's active formal policy. This ensures that the comparison measures decision quality on the same events, and its purpose is to show the effectiveness of the formal policy compared to the natural language policy approaches. We additionally evaluate the complete deployment stream in Figure 4C, Section 4.3, where events covered by the active policy get deterministic symbolic decisions, while uncovered events are routed to the deployment-time LLM fallback. VEPA maintains strong performance in this setting. We also report formal-policy coverage for each dataset in Appendix C.3, with final-round coverage reaching approximately 90% across different benchmarks.


Minor corrections

Thank you for identifying these issues. We have corrected them in our revision.


We hope our responses and additional experiments have addressed your comments and clarified the strengths of our approach.

Dear Reviewer BN7F,

Thank you again for reviewing our paper and providing constructive feedback. We wanted to check in to see whether our rebuttal, including new experiments (LLM parser analysis), has addressed your comments. We would be happy to discuss any remaining questions you might have.

Reviewer_E75kOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes VEPA (verified evolving formal policy for lifelong LLM agent guardrails), which is a framework that enables continual adaptation of formal safety policies for LLM agents based on deployment feedback. VEPA addresses the limitations of existing neural guardrails, which can adapt to new risks but may produce inconsistent decisions, and formal guardrails, which provide deterministic enforcement but typically rely on static policies. VEPA represents agent interactions as structured events and maintains executable symbolic rules for ALLOW/REFUSE decisions. Its policy evolution mechanism uses LLM-generated rule edits and archive-based Pareto search to balance failure repair against regression prevention, while a deterministic verifier checks structural validity and rule conflicts. Additionally, an evidence-aware authority mechanism employs Beta-Bernoulli posterior risk estimation to activate only rules with sufficient supporting evidence. The authors evaluate VEPA on ASSEBench-Safety, ASSEBench-Security, and ATBench under simulated online deployment settings. Results demonstrate improvements over neural, memory-based, and symbolic rule-learning baselines, with additional experiments examining heterogeneous deployment streams, noisy feedback, longer adaptation horizons, and the contributions of individual components.

Critical strengths
  1. The research question is well-motivated. The paper identifies a meaningful gap between adaptive neural guardrails and deterministic formal enforcement.
  2. The proposed framework is well-structured with complementary components.
  3. The experimental results are consistent and the ablation studies are useful.
Critical weaknesses
  1. Insufficient evaluation of the LLM-based event parser. VEPA relies on an LLM to convert raw agent interactions into structured events, but the reliability of this parser is not evaluated. Errors in critical attributes, such as authorization status or data sensitivity, could directly affect downstream safety decisions. I suggest reporting field-level parsing accuracy and testing how controlled parsing errors (e.g., 5%, 10%, and 20%) affect end-to-end F1 and false-ALLOW rates.

  2. Limited evaluation of lifelong adaptation under distribution shifts. Although VEPA is evaluated on mixed-risk streams and longer deployment horizons, these experiments mainly use randomly shuffled samples from fixed datasets. This does not fully capture scenarios where new risks emerge over time. I suggest evaluating sequential risk shifts (e.g., unauthorized access → data exfiltration → prompt injection), measuring how quickly VEPA adapts to new risks and whether it retains performance on previously learned ones.

  3. Limited comparisons and insufficient analysis of methodological contributions. VEPA combines several existing techniques, including LLM-guided rule generation, Pareto search, and symbolic verification. While the ablations show improvements over greedy search, they do not clearly isolate the benefits of the Pareto formulation and diversity-aware sampling. Comparisons with simpler alternatives, such as single-objective search using a weighted combination of failure repair and regression avoidance, Pareto search with uniform sampling instead of diversity-aware sampling, or counterexample-guided rule refinement, under comparable computational budgets would help better establish the contribution of the proposed approach.

Questions for authors

See above.

I am happy to discuss during rebuttal process.

5 comments
AuthorsOpenReview ↗

Response to Reviewer E75k [Part 1/2]

We thank the reviewer for the thoughtful review and constructive suggestions. We are happy that you found our research question well-motivated, recognized the meaningful gap between adaptive neural guardrails and deterministic formal enforcement, and the complementary framework components and consistent experimental results. We address your comments below.


[W1] Evaluation of the LLM parser

We add two analyses to evaluate the reliability of the LLM parser and the impact on downstream guardrail performance. These experiments with more detailed settings are presented in the new Section 5.3 of the updated PDF.

Parsing Accuracy. First, we evaluate structured event extraction using three stronger LLMs as independent reference judges: Gemini-3.1-Pro, Gemini-3.8-Flash, and Claude-Opus-5. Each judge receives the original agent trajectories and the same DSL schema as VEPA's parser (Gemini-3.1-Flash-Lite), and independently extracts structured events.

We report three metrics: Field Agreement, which measures agreement across all schema fields; Critical-Field Agreement, which measures agreement on fields whose values can affect formal-policy decisions; and Decision Agreement, which measures whether the parser and reference judge produce identical formal-policy decisions. We additionally construct a three-judge consensus through field-level majority voting.

LLM JudgeField Agree. (%) ↑Critical Agree. (%) ↑Decision Agree. (%) ↑
Gemini-3.1-Pro94.294.697.8
Gemini-3.8-Flash92.893.996.5
Claude-Opus-593.792.597.6
3-Judge Consensus93.993.497.6

The table above shows that VEPA's lightweight parser achieves 92.8-94.2% field agreement and 92.5-94.6% critical-field agreement across the three independent judges. Moreover, decision agreement is 96.5-97.8%, showing that most field-level disagreements do not alter downstream formal-policy decisions. Agreement with the three-judge consensus is similarly high. The consistently high agreement across different reference models suggests that the lightweight parser provides a consistent and reliable structured representation for downstream formal-policy enforcement.

Robustness to Parsing Errors. Second, we study how parsing errors affect VEPA's end-to-end performance. We introduce controlled errors into structured events at rates of ρ∈{0,0.05,0.1,0.2}\rho \in \{0, 0.05, 0.1, 0.2\}. For each nonzero error rate, we randomly select the corresponding fraction of events and replace one randomly selected policy-relevant field with a different valid value from its DSL vocabulary. This simulates incorrect attribute extraction while preserving the structural validity of event representations. We rerun the online deployment protocol with corrupted events with the same experimental settings.

Parsing Error RateF1 ↑
0%0.86 ± 0.02
5%0.86 ± 0.04
10%0.85 ± 0.03
20%0.82 ± 0.04

Under moderate corruption (5%-10%), VEPA maintains comparable performance. Together with the high parser agreement in our experiment above, where field-level disagreement with independent LLM judges is below 8%, these results suggest that VEPA is relatively insensitive to moderate extraction errors. Even at a 20% corruption rate, performance decreases by only 0.04 F1.

Overall, these experiments show that (1) VEPA's lightweight parser achieves consistently high agreement with independent stronger LLM judges, particularly in downstream formal-policy decisions, and (2) VEPA's end-to-end performance remains stable under moderate upstream extraction errors.

AuthorsOpenReview ↗

Response to Reviewer E75k [Part 2/2]

[W2] Evaluation under distribution shifts

We would like to note that the benchmarks already cover several distinct risk types (prompt injection, unauthorized access, data exfiltration, unsafe tool use) across different agent applications. Therefore, the mixed stream experiment (Figure 4A) requires a single policy to learn and adapt for multiple different types of risks.

To strengthen the evaluation, following your suggestion, we further conduct a more controlled experiment where different security risks emerge sequentially. Specifically, using ASSEBench-Security's risk annotations, we construct three consecutive phases: Unauthorized Access & Control → Data Exfiltration & Leakage → Prompt Injection. We report two metrics: Final-average F1, which measures the mean F1 across all three risk categories after the final phase, and Average Regression, which measures the percentage of previously correct examples that become incorrect after learning subsequent risks, averaged over the first two risk categories. Following our regression analysis in Appendix C.2, we evaluate regression on examples correctly handled by all evaluated methods at the checkpoint after learning each risk. We compare our method with two strongest baselines (AGrail and LiSA).

MethodFinal-Avg. F1 ↑Avg. Regression (%) ↓
AGrail0.75 ± 0.038.7 ± 2.4
LiSA0.77 ± 0.066.5 ± 2.0
VEPA (Ours)0.82 ± 0.042.5 ± 0.8

VEPA achieves the highest final-average F1 of 0.82, outperforming the strongest baseline, LiSA, by 0.05 F1. Moreover, VEPA substantially reduces regression to 2.5%, compared with 6.5% for LiSA and 8.7% for AGrail. These results show that VEPA can adapt to newly emerging risks while better preserving previously learned behavior. This further supports the effectiveness of our policy evolution framework in balancing failure repair and regression avoidance.

We have added this experiment to the new section C.8 of the updated PDF.


[W3] Comparisons with other search strategies

Thank you for this suggestion. We conduct an additional experiment to study the contributions of our Pareto search and diversity-aware sampling method. We compare our method against two additional variants: (1) Weighted Sum, which replaces the bi-objective Pareto formulation with a single scalarized objective, and (2) Uniform Sampling, which retains the Pareto frontier but samples parent policies uniformly. Specifically, Weighted Sum combines failure repair and regression avoidance with equal weights 0.50.5, and selects the final policy with rankings based on the scalarized objective. For Uniform Sampling, we preserve the original Pareto search and final policy selection but sample each parent policy from the frontier with equal probability. All methods use the same budget for LLM search. Similar to the experiment above, we report final-round F1 score and regression rate.

Search StrategyF1 ↑Regression (%) ↓
Weighted Sum0.82 ± 0.037.2 ± 3.5
Uniform Sampling0.79 ± 0.043.6 ± 1.3
VEPA (Ours)0.86 ± 0.023.8 ± 1.2

VEPA achieves the highest F1 of 0.86 while maintaining a low regression rate of 3.8%. Compared with Weighted Sum, our Pareto formulation improves F1 by 0.04 while reducing regression by 3.4%. Moreover, VEPA outperforms Uniform Sampling by 0.07 F1, with a comparable regression rate. We observe that weighted-sum optimization can favor one objective at the expense of the other, leading to larger regressions, while uniform sampling can repeatedly explore similar policies and struggle to discover effective updates in many evolution rounds. Together with the existing Greedy ablation, these results highlight the complementary benefits of Pareto optimization and diversity-aware exploration for policy evolution. We have added this experiment to the new section C.9 of the updated PDF.

Regarding your suggestion on counterexample-guided rule refinement, we note that VEPA already incorporates counterexample-driven feedback into policy evolution. It is provided to the LLM proposer to guide subsequent policy revisions (details in lines 278-285 of our manuscript).


We hope our responses and additional experiments have addressed your comments and clarified the strengths of our approach.

Dear Reviewer E75k,

Thank you again for reviewing our paper and providing constructive feedback. We wanted to check in to see whether our rebuttal, including new experiments (LLM parser analysis, risk distribution shifts, search strategy ablations), has addressed your comments. We would be happy to discuss any remaining questions you might have.

Reviewer_Lm7sOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

VEPA learns an evolving formal guardrail policy from deployment feedback. An LLM maps traces to typed events and proposes policy edits; a deterministic verifier rejects malformed or conflicting policies. Pareto search balances correction of observed failures against regression on previously correct cases, while a Bayesian evidence criterion determines which rules become active. The paper evaluates sequential adaptation on three benchmarks, including sparse and noisy feedback, ablations, and a hybrid system with an LLM fallback.

Critical strengths

The architecture provides a useful separation between proposing rules, checking their structural validity, and granting them deployment authority. The formal claims are explicitly bounded by the event schema and do not incorrectly imply that the neural parser or overall system is verified.

The empirical evaluation includes relevant adaptive guardrails and symbolic rule-learning baselines, multiple seeds, sparse and noisy feedback, and component ablations. The full-stream hybrid evaluation is particularly useful because policy coverage alone would not establish practical guardrail performance.

Critical weaknesses

The authority mechanism lacks sufficient prospective calibration evidence. The default δ = 0.5 uses the posterior median, and rules are evaluated on an archive that also influenced their construction and selection. Appendix D.2 appropriately acknowledges this limitation, but it remains unclear whether the resulting activation criterion controls future rule errors at a useful coverage level. This matters because evidence-dependent activation is a central contribution.

The “lifelong” evaluation primarily uses sequential replay of fixed benchmark distributions and a fixed event schema. It does not establish adaptation to changing policies, obsolete evidence, or new kinds of safety-relevant events. In addition, the formal guarantees are conditional on the neural parser, whose errors and susceptibility to adversarial traces are not sufficiently characterized.

Questions for authors

How do future error rates and coverage change when authority is estimated using held-out or post-creation feedback? How sensitive are these results to a more conservative δ?

Under an explicit policy change, how quickly are previously well-supported rules demoted or revised, given that historical evidence remains in the archive?

Can the authors quantify parser errors against audited event annotations and show their effect on full-stream false-allow and false-refuse rates?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Lm7s [Part 1/2]

We thank the reviewer for the thoughtful review and insightful comments. We are glad that you found our separation of policy proposal, structural verification, and deployment authority useful, and recognized the clearly scoped formal guarantees and comprehensive empirical evaluation. We provide the responses to your comments below.


[W1 + Q1] Future error rates and coverage under different authority settings and more conservative δ\delta

Following your suggestions, we extend our evaluation to three authority-estimation settings: (1) Retrospective (our default setting), which uses all accumulated feedback; (2) Held-out, which separates evidence used for policy evolution from evidence used for authority estimation; and (3) Post-creation, which uses only feedback collected after a rule is generated. We evaluate future error rates and formal-policy coverage on unseen deployment events using the same posterior-risk criterion and other settings as in our main experiments.

MethodFuture Error ↓Coverage ↑
VEPA (Default)7.2 ± 1.892.2 ± 4.7
Held-out6.8 ± 2.283.6 ± 5.2
Post-creation7.0 ± 1.780.8 ± 5.7

The table above shows that using all accumulated feedback achieves a 7.2% future error rate with the highest coverage at 92.2%. Using held-out evidence and post-creation evidence slightly reduces the future error, but decreases coverage significantly. Overall, accumulating all evidence achieves a similar error rate while covering substantially more events, showing the benefit of leveraging accumulated feedback in our design.

Regarding your question on sensitivity to more conservative thresholds, we conduct an additional experiment that varies δ∈{0.5,0.2,0.1,0.05}\delta \in \{0.5, 0.2, 0.1, 0.05\}. We keep the other settings the same. Smaller corresponds to a more conservative posterior-risk quantile, which requires more evidence before rule activation.

δ\deltaFuture Error ↓Coverage ↑
0.50 (Default)7.2 ± 1.892.2 ± 4.7
0.205.9 ± 2.185.7 ± 4.4
0.105.2 ± 1.778.0 ± 3.4
0.054.5 ± 1.872.2 ± 2.8

The table above shows that reducing δ\delta from 0.5 to 0.05 consistently decreases future error from 7.2% to 4.5%, while coverage decreases significantly from 92.2% to 72.2%. These results demonstrate a trade-off between future error rate and formal-policy coverage. We choose δ=0.5\delta=0.5 because it provides a favorable balance between error rate and formal-policy coverage, with comparable error rates compared to other values while having faster rule activation and substantially higher coverage under sparse feedback.

We present these experiments with more detailed settings in the new Section C.7 of the updated PDF.


[W2] Lifelong evaluation under diverse risks

We would like to clarify that our existing benchmarks already cover diverse safety and security risks, such as prompt injection, unauthorized access, data exfiltration, and unsafe tool use across different agent applications. Therefore, our mixed-stream experiment (Figure 4A) requires a single evolving policy to continually adapt to multiple types of risks while preserving previously learned behavior. Our longer-horizon experiments further show that these advantages hold as policy updates accumulate over time.

AuthorsOpenReview ↗

Response to Reviewer Lm7s [Part 2/2]

[Q2] Policy changes and well-supported rules

Well-supported rules are harder to demote because they have accumulated substantial supporting evidence. Therefore, they require more contradictory evidence to exceed the posterior-risk threshold. For example, under our default setting, a rule with 20 supporting examples requires only 1 contradiction to be demoted, whereas a rule with 100 supporting examples requires 5 contradictions. On the other hand, newly introduced or weakly supported rules are more sensitive to new evidence and can be demoted more quickly when contradictions arise. These mechanisms make the policy stable under noisy feedback, as shown in our experiment in Figure 4B.


[W3 + Q3] Evaluation of the LLM parser

We add two analyses to evaluate the reliability of the LLM parser and the impact on downstream guardrail performance. These experiments with more detailed settings are presented in the new Section 5.3 of the updated PDF.

Parsing Accuracy. First, we evaluate structured event extraction using three stronger LLMs as independent reference judges: Gemini-3.1-Pro, Gemini-3.8-Flash, and Claude-Opus-5. Each judge receives the original agent trajectories and the same DSL schema as VEPA's parser (Gemini-3.1-Flash-Lite), and independently extracts structured events.

We report three metrics: Field Agreement, which measures agreement across all schema fields; Critical-Field Agreement, which measures agreement on fields whose values can affect formal-policy decisions; and Decision Agreement, which measures whether the parser and reference judge produce identical formal-policy decisions. We additionally construct a three-judge consensus through field-level majority voting.

LLM JudgeField Agree. (%) ↑Critical Agree. (%) ↑Decision Agree. (%) ↑
Gemini-3.1-Pro94.294.697.8
Gemini-3.8-Flash92.893.996.5
Claude-Opus-593.792.597.6
3-Judge Consensus93.993.497.6

The table above shows that VEPA's lightweight parser achieves 92.8-94.2% field agreement and 92.5-94.6% critical-field agreement across the three independent judges. Moreover, decision agreement is 96.5-97.8%, showing that most field-level disagreements do not alter downstream formal-policy decisions. Agreement with the three-judge consensus is similarly high. The consistently high agreement across different reference models suggests that the lightweight parser provides a consistent and reliable structured representation for downstream formal-policy enforcement.

Robustness to Parsing Errors. Second, we study how parsing errors affect VEPA's end-to-end performance. We introduce controlled errors into structured events at rates of ρ∈{0,0.05,0.1,0.2}\rho \in \{0, 0.05, 0.1, 0.2\}. For each nonzero error rate, we randomly select the corresponding fraction of events and replace one randomly selected policy-relevant field with a different valid value from its DSL vocabulary. This simulates incorrect attribute extraction while preserving the structural validity of event representations. We rerun the online deployment protocol with corrupted events with the same experimental settings.

Parsing Error RateF1 ↑
0%0.86 ± 0.02
5%0.86 ± 0.04
10%0.85 ± 0.03
20%0.82 ± 0.04

Under moderate corruption (5%-10%), VEPA maintains comparable performance. Together with the high parser agreement in our experiment above, where field-level disagreement with independent LLM judges is below 8%, these results suggest that VEPA is relatively insensitive to moderate extraction errors. Even at a 20% corruption rate, performance decreases by only 0.04 F1.

Overall, these experiments show that (1) VEPA's lightweight parser achieves consistently high agreement with independent stronger LLM judges, particularly in downstream formal-policy decisions, and (2) VEPA's end-to-end performance remains stable under moderate upstream extraction errors.


We hope our responses and additional experiments have addressed your comments and clarified the strengths of our approach.

Dear Reviewer Lm7s,

Thank you again for reviewing our paper and providing constructive feedback. We wanted to check in to see whether our rebuttal, including new experiments (calibration evidence, LLM parser analysis), has addressed your comments. We would be happy to discuss any remaining questions you might have.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response and Paper Revision

Dear Reviewers and Area Chairs,

Thank you for taking the time to review our paper and provide thoughtful and constructive feedback.

In our rebuttal, we have provided responses to each reviewer's comments and revised our manuscript. All changes are highlighted in blue text. The updates include:

  1. LLM parser analysis (Section 5.3): We provide a comprehensive analysis of the LLM parser, including its consistent and reliable performance and the robustness of VEPA under controlled parsing errors.

  2. Additional experiments: We conduct several new experiments to further show the effectiveness of VEPA under different settings:

  • Rule-authority calibration (Appendix C.7): VEPA achieves consistent prospective performance across different evidence sources.
  • Adaptation under sequential shifts (Appendix C.8): VEPA consistently outperforms adaptive guardrail baselines while effectively reducing regressions when the risk categories are structured sequentially.
  • Pareto-based search (Appendix C.9): VEPA's search strategy is more effective than alternative search strategies, including weighted-sum optimization and uniform sampling from the Pareto frontier.
  • Runtime enforcement (Appendix C.10): VEPA effectively prevents successful attacks while preserving legitimate task completion during interactive agent execution.
  1. Clarifications of our experimental setting and method: We clarify the user deployment feedback (Section 3.1) meaning of verification (Section 3.2.1), make the long-horizon deployment setting more explicit (Section 5.2 and Appendix B.1), and the failure set (Section B.2).

We hope that our responses and the additional experiments address the reviewers' comments and clarify our paper's strengths and contributions. We would be happy to discuss if there are any remaining questions during the discussion period.

Thank you once again for your valuable feedback and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.