acceptodds
Under review as a conference paper at ICLR 2027

CabinSafeBench: Evaluating Context-Adaptive Safety in Intelligent Cockpit Agents

Abstract

Large language model (LLM)-based cockpit agents are increasingly capable of completing complex in-vehicle tasks, yet existing evaluations largely ask whether an agent can achieve a user goal, not whether it can adapt how that goal is achieved when safety-relevant conditions change. We study context-adaptive safety: the capability to preserve goal achievement, when a safe realization exists, while adapting execution behavior to the current driving context. We introduce CabinSafeBench, a benchmark based on controlled context intervention that keeps user intent and task goals fixed while varying safety-relevant world states. This design isolates failures of safety adaptation from failures of task understanding. We further develop CockpitWorld, a closed-loop environment coupling Android Automotive OS with CARLA, enabling safety to be evaluated from realized state transitions and execution consequences rather than tool calls alone. Across six LLM-based cockpit agents, task success is nearly saturated at 99.85%, while safe task success is only 75.38%. Under controlled context interventions, task success remains essentially unchanged, whereas safe task success drops from 99.81% to 40.19%. These results show that strong task competence does not imply context-adaptive safety: agents may know what goal to achieve while failing to adapt how it should be achieved as the world changes.

Primary area:
datasets and benchmarks
Keywords:
Large Language Models, Intelligent Agents, Safety Evaluation, Context-Adaptive Safety, Benchmarking, Embodied AI, Intelligent Cockpit

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_ZMBUOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces CabinSafeBench to evaluate the safety adaptation of cockpit agents under changing driving contexts. It constructs paired intervention scenarios by keeping user requests and task goals fixed while changing safety-relevant contexts, and develops CockpitWorld, a closed-loop simulation environment that couples Android Automotive OS (AAOS) with CARLA. The authors evaluate six LLM-based cockpit agents using metrics including Task Success Rate (TSR), Safe Task Success Rate (STSR), and Context Adaptation Failure Rate (CAFR). The results show that task success remains high after context interventions, while safe task success drops substantially.

Critical strengths
  1. Practical research problem. The paper highlights the distinction between task completion and safe execution. By controlling context changes, it studies agents' safety adaptation and offers a reasonable perspective on safety evaluation beyond conventional task success rates.

  2. Executable closed-loop evaluation framework. CockpitWorld couples AAOS with CARLA and uses deterministic, machine-checkable rules to evaluate task execution. This allows action safety to be assessed under specific environmental states rather than relying solely on natural-language outputs.

Critical weaknesses
  1. Limited benchmark size, model coverage, and interpretability of aggregate results. The core intervention experiment contains only 90 paired scenarios, which may not adequately cover the diversity of cockpit safety challenges. The six models are also not compared under controlled model-scale settings, limiting the representativeness and generalizability of the findings. In addition, Table 1 shows dramatic differences in STSR, TSG, and CAFR across models released around the same time. The paper does not sufficiently discuss or explain these differences, nor does it provide error-analysis experiments to establish whether they stem from model capabilities rather than confounding factors such as inference configurations, tool-call integration, or the evaluation pipeline. The paper also does not explain why six models from different families form a representative model set. Averaging these heterogeneous models lacks a clearly defined statistical population, making the resulting scores difficult to interpret as indicators of broader cockpit-agent capabilities or specific capability gaps.

  2. Insufficient verification of goal preservation, safety ground truth, and simulation validity. Keeping the task goal unchanged is a core design principle, but the natural-language request and the underlying goal predicate may be ambiguous. The current evidence does not establish that alternative executions in all intervention pairs satisfy the same task goal, nor does it allow independent verification that safe feasibility is strictly guaranteed in both contexts. Moreover, deterministic, machine-checkable evaluations ensure consistent rule execution but do not establish that the rules themselves are valid. The paper provides insufficient justification for the sources, triggering conditions, and key thresholds of its safety constraints, leaving it unclear whether STSR and CAFR accurately reflect actual safety adaptation capabilities. The generation of dynamic driving contexts, state updates, and specific scenario configurations also remain insufficiently clear. The limit of four interactions per episode makes the coverage of long-horizon tasks difficult to assess, limiting independent verification of the simulation capabilities and experimental results.

Figure-related issues. The Dense Traffic example in Figure 1 does not clearly show dense traffic. Figure 4 contains overlapping elements within its subplots, reducing readability.

Questions for authors
  1. Benchmark size and coverage: Are 90 paired context-intervention scenarios sufficient? What supports the representativeness of these scenarios and the generalizability of the findings?

  2. Cross-model differences and evaluation reliability: How can the extreme cross-model performance differences be explained or shown to reflect genuine model capabilities rather than evaluation artifacts? Why do the actual inference configurations differ across models?

  3. Interpretation of Macro Avg: What is the statistical meaning of averaging results across six different model families? How should this metric be interpreted without a clear justification for the representativeness of the selected models?

  4. Goal preservation: When the literal user request may differ from the underlying task goal, how do the authors verify that an alternative execution after intervention still satisfies the original goal? Can specific pairs illustrate how task goals are defined and safe feasibility is verified?

  5. Safety ground truth: What are the sources, operational definitions, triggering conditions, and key thresholds of the safety constraints? How is the validity of these rules independently verified to ensure that the evaluation results are not biased by the rules themselves?

4 comments
AuthorsOpenReview ↗

Benchmark Validity and Evaluation Reliability

We thank the reviewer for recognizing the importance of our research problem and executable evaluation framework. We clarify the experimental design and evaluation validity below.

W1 & Q1/Q3: Benchmark Coverage and Statistical Interpretation.

CabinSafeBench contains 220 episodes per agent, comprising 40 baseline episodes and 90 nominal–intervened pairs, spanning six task families and six safety-relevant context dimensions. The 90 pairs constitute the canonical controlled-intervention set, specifically designed to measure safety adaptation under fixed task objectives. Importantly, our conclusions are supported by paired statistical analysis rather than aggregate scores alone. As reported in Appendix D, aligned pair-cluster bootstrap resampling (10,000 replicates) yields a 59.63 percentage-point STSR decrease (95% CI: [56.67, 62.41]), while TSR remains essentially unchanged. The Macro Avg. represents the equally weighted descriptive mean across the six evaluated agents, not a population-level estimate of all cockpit agents. We additionally report individual model results and context-specific profiles to explicitly characterize heterogeneity rather than obscure it through aggregation.

W1 & Q2: Cross-Model Differences and Inference Reliability.

The six agents are evaluated on identical benchmark instances using the same system-prompt template, observation format, 20 semantic tools, and interaction budget. As detailed in Appendix B, all models use the same OpenAI-compatible endpoint, temperature (0.0), maximum output tokens (2,048), and automatic tool selection. The differences in timeout and retry settings concern provider-specific transport robustness, not task semantics or decoding parameters. Furthermore, all models achieve nearly identical TSR (99.55–100.00%), while their safety performance differs substantially. This distinction is supported by within-model intervention comparisons, not merely cross-model rankings. The common-eligibility analysis in Appendix D further shows that restricting CAFR to the same 89 eligible pairs preserves the complete model ranking, with a maximum change of only 0.76 percentage points. These results establish substantial differences in observable end-to-end safety adaptation under a standardized evaluation protocol.

W2 & Q4: Goal Preservation and Safe Feasibility.

Goal preservation is explicitly enforced by the benchmark construction rather than inferred from natural-language similarity. Each task is represented as τ=(u,g)\tau=(u,g), where uu is the user request and gg is an executable goal predicate. Both contexts share the identical (u,g)(u,g), while safety is independently evaluated through context-dependent violation predicates. As specified in Sections 3.2–3.6 and Appendix A.4, a pair is retained only when a safe goal-achieving realization exists in both contexts, verified through reference completion strategies and environment-state checks. For example, the request "let some fresh air in" preserves the functional goal of increasing cabin fresh air; opening a window and using an appropriate ventilation mechanism provide alternative realizations under different operating conditions. The reference strategies establish feasibility but are never exposed to evaluated agents. Thus, the evaluation distinguishes failure to adapt execution from changes in task semantics or inherent task infeasibility.

W2 & Q5: Safety Ground Truth and Simulation Validity.

Safety evaluation is defined through executable predicates over the current state, selected action, and resulting state transition, rather than subjective LLM judgments. As described in Appendix A.5, safety constraints have stable identifiers, applicable tasks, contextual triggering conditions, and rule provenance. Externally grounded requirements are distinguished from benchmark-specific operational assumptions. We will further clarify the individual rule conditions, thresholds, and supporting rationale to make these judgments directly auditable. CockpitWorld couples AAOS vehicle-function execution with CARLA driving-state evolution, propagating updated states across interaction steps. The four-step budget defines the evaluated interaction horizon; stateful execution and multi-step adaptation are supported within this budget. We will clarify this scope and improve the traffic depiction in Figure 1 and subplot readability in Figure 4.

Reviewer_ZMBUOpenReview ↗

Update Review

Thanks for detailed response. However, several key concerns remain insufficiently addressed.

1. Benchmark Size and Representativeness.

The paper still lacks sufficient analysis of coverage across task types, context combinations, and scenario complexity. Could the authors provide a sample-size convergence analysis or evaluation of additional tasks and intervention scenarios to demonstrate that the current benchmark size is sufficient to support the main conclusions?

2. Unexpected Cross-Model Performance and Evaluation Reliability.

The authors emphasize that all models use the same tool interface and basic inference settings, but this does not adequately explain the extreme performance differences, particularly the unexpected gaps between stronger and weaker models. The common-eligibility analysis only rules out differences in CAFR denominators; it does not establish that the results reflect genuine model capabilities rather than differences in actual inference configurations, tool integration, or scoring rules. As a benchmark, the paper should provide concrete failure analyses or controlled experiments to explain these results, which deviate substantially from typical capability expectations, rather than merely reporting score differences.

3. Interpretability of Macro Avg.

I understand that the authors define Macro Avg. as a descriptive mean across the six evaluated models rather than a population-level estimate. However, it remains unclear why these six models were selected and whether they are representative in terms of widespread use or state-of-the-art capabilities. Without a clear justification for model selection, the average is highly dependent on the chosen model set, limiting its value as a main aggregate result and potentially making it misleading.

We thank the reviewer for the follow-up comments. We provide additional quantitative evidence from our existing experimental data.

1. Benchmark Size and Representativeness.

CabinSafeBench contains 220 episodes per agent, including 40 baseline episodes and 90 aligned intervention pairs. We further conducted a sample-size stability analysis with 2,000 repeated subsamples, preserving pair alignment across models. At 15 pairs, the estimated ΔSTSR is 59.61 pp (95% interval: [53.33, 65.56]), compared with 59.63 pp over all 90 pairs; CAFR is 59.91% versus 59.97%. These results demonstrate consistent estimates of intervention-induced safety degradation across sample sizes within the evaluated suite. We additionally analyzed task distributions, context combinations, and interaction complexity to characterize the benchmark's operational coverage.

2. Cross-Model Performance and Evaluation Reliability.

Matched execution records directly reveal distinct action-selection behaviors across models. We aligned all 90 intervention pairs across four models and compared their actions under identical task and context specifications. In representative high-speed, heavy-rain, and closed-road scenarios, GPT and DeepSeek execute context-inappropriate target actions, whereas Gemini and Grok withhold the corresponding writes. Furthermore, our audit of five available direct-execution event streams confirms that all 1,083 executor-submitted tool calls had schema-valid recorded arguments and successful tool responses. These execution-level observations demonstrate genuine differences in action selection, independently observable before aggregate scoring. The analysis therefore provides concrete behavioral evidence for the pronounced cross-model differences.

3. Macro Average and Model Selection.

The six models were selected to provide cross-provider and cross-family coverage, enabling comparison of heterogeneous LLM agents under a unified evaluation protocol. Macro Avg. represents their equally weighted descriptive mean. We further conducted a leave-one-model-out sensitivity analysis, obtaining STSR values of 70.91–78.64% and CAFR values of 52.18–70.85%. We report these results alongside individual model scores and paired intervention analyses to characterize model-specific behavior and aggregate sensitivity. The primary contribution is a controlled evaluation framework that exposes context-adaptive safety failures through goal-preserving interventions.

Reviewer_ZMBUOpenReview ↗

Official Comment by Reviewer ZMBU

Thanks for response. Unfortunately, the authors still have not provided sufficient analysis of scenario coverage or detailed error analysis. Although they have added a subsampling stability experiment and execution-log audits for some models, this evidence is still insufficient to establish the representativeness of the current dataset or adequately explain the extreme performance differences across models. For example, DeepSeek V4.1 Flash has a CAFR of 98.89%, while Gemini 3.8 Flash has only 5.56%, a difference of 93.33 percentage points. Many similar discrepancies can be found in the main table. If such substantial score differences are primarily driven by a subset of specific scenarios or particular action-selection patterns, this further raises my concerns about the adequacy of scenario coverage and the validity of the evaluation design.

Furthermore, the authors still have not provided sufficient evidence or references to justify the representativeness of the six selected models. Although the authors define Macro Avg as a descriptive average, the paper continues to rely heavily on these averages throughout its analysis and conclusions. Without a clear justification for model selection, these results are highly susceptible to being overinterpreted as reflecting the overall capabilities of the field, which could be misleading.

In conclusion, I believe these core concerns remain unresolved. I remain seriously concerned about the evaluation validity of this benchmark and the reliability of its conclusions, and therefore maintain my initial assessment and rating.

Reviewer_D1N7OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces CabinSafeBench, a benchmark for evaluating agents can preserve safe task completion when safety-relevant environmental conditions change. Its central design holds user requests and task goals fixed while constructing paired contexts that both admit safe solutions. The accompanying CockpitWorld environment couples Android Automotive OS with CARLA to support stateful execution and deterministic safety evaluation. Across six agents, overall task success reaches 99.85%, whereas safe task success is only 75.38%. Under controlled context interventions, safe task success drops from 99.81% to 40.19%, despite essentially unchanged functional performance. A rule-based SafetySupervisor substantially recovers these failures, although it shares safety knowledge with the evaluator. The results highlight a substantial gap between functional competence and context-appropriate safe execution.

Critical strengths
  1. The paper clearly distinguishes achieving a user’s goal from choosing a safe way to achieve it under the current conditions. Holding task semantics fixed and requiring safe solutions in both contexts helps separate adaptation failures from changes in the objective or task infeasibility. This provides a meaningful framework for evaluating safety while preserving task utility.

  2. The benchmark reveals substantial differences in safety behavior among agents whose task success rates are nearly indistinguishable. The per-model and per-context analyses expose meaningful heterogeneity, while aligned pair-cluster bootstrap estimates and the common-eligibility analysis strengthen confidence that the principal findings are not artifacts of aggregation or differing CAFR denominators.

Critical weaknesses

My main concern is the gap between the breadth of the research question and the narrow coverage of the benchmark. The paper studies context-adaptive safety, but its evaluation is confined to a limited range of cockpit interactions. The taxonomy predominantly describes vehicle-specific environmental and operational conditions, without systematically covering broader sources of context-dependent risk, such as changes in authorization, information trustworthiness, or privacy constraints. Consequently, it remains unclear whether the benchmark captures sufficiently diverse forms of safety adaptation to provide insights beyond cockpit-specific operational constraints.

Questions for authors

See Weaknesses.

4 comments
AuthorsOpenReview ↗

Scope and Coverage of Context-Adaptive Safety

We thank the reviewer for recognizing the value of our problem formulation and controlled evaluation methodology. It is important to distinguish the generality of the research question from the scope of its empirical evaluation. Context-adaptive safety is the capability under investigation, while intelligent cockpit interaction provides a concrete, safety-critical domain for systematically evaluating this capability. CabinSafeBench covers diverse contextual changes, including driving conditions, occupant states, and road/navigation conditions, each requiring agents to adapt their execution strategies while preserving user goals. These scenarios represent distinct safety adaptation requirements rather than variations of a single operational rule. The benchmark is therefore designed to systematically evaluate context-dependent safe execution within intelligent cockpits, rather than exhaustively enumerate all possible categories of agent safety risks.

The methodological contribution lies in isolating and measuring safety adaptation, rather than maximizing the number of risk categories. By holding user intent and task feasibility constant while intervening on safety-relevant context, CabinSafeBench distinguishes adaptation failures from task comprehension failures or infeasibility. CockpitWorld further grounds this evaluation in executable actions and observable state transitions. Authorization, information trustworthiness, and privacy are important dimensions of agent safety, but introduce distinct security and trust considerations beyond the physical and operational safety addressed in this work. Our findings demonstrate that near-saturated task success does not guarantee context-appropriate safe execution, while the controlled intervention methodology provides a systematic approach to identifying this gap. The generality of an evaluation principle should not be conflated with exhaustive coverage of safety risk categories. We will clarify this distinction in the revised paper.

Reviewer_D1N7OpenReview ↗

Response to the Rebuttal

Thank you for the response. To clarify, my concern is not that a single benchmark must exhaustively cover every safety risk category, so framing it as a request for exhaustive enumeration misses the main point.

  1. Evaluating only cabinsafe is not a foundamental or major problem in agent safety evaluation. I still think the authors should add more scenes on context-adaptive safety to make the benchmark more meaningful and acceptable for ARR series.

  2. And the benchmark focuses on a narrow set of physical vehicle-control operations, remaining unclear whether the observed model behaviors reflect generalizable context-adaptive safety or domain-specific sensitivity to driving rules. Clarifying the scope in the revision will be helpful, though my concern regarding the gap between the broad framing and narrow empirical coverage remains.

We thank the reviewer for the clarification.

1. Significance of Context-Adaptive Safety. CabinSafeBench investigates a fundamental challenge in agent execution: preserving task objectives while adapting actions to changing safety constraints. The same user goal may require different execution strategies under different environmental conditions. Our controlled intervention methodology isolates this capability by keeping user intent and task goals fixed while modifying safety-relevant contexts. By evaluating executable actions and resulting state transitions, CockpitWorld provides a concrete basis for measuring context-dependent safety adaptation.

2. Generalizability and Empirical Scope. The key methodological contribution is goal-preserving controlled context intervention, which separates task competence from safety-conditioned execution. CabinSafeBench instantiates this methodology in intelligent cockpits, where changes in vehicle motion, weather, traffic, occupant state, and navigation conditions affect the safety of agent actions. We will make the task distribution and intervention coverage more explicit to characterize the empirical scope. The resulting framework provides a reusable evaluation principle for investigating how agents adapt their execution strategies under changing safety constraints.

Reviewer_D1N7OpenReview ↗

Further Response

I sincerely thank the authors for their response. However, my concerns about this work remain, and I believe Reviewer ZMBU raises similar concerns. While I agree that context-adaptive safety is an important research topic, the benchmark still has limitations in its scope and evaluation format. I will therefore maintain my original score.

Reviewer_TgGBOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates the performance of Large Language Model (LLM)-driven cockpit agents in safety-critical scenarios. The authors introduce the concept of "context-adaptive safety"—the ability of an agent to adjust its execution strategy based on a changing driving environment while still achieving its task objectives. To this end, they developed the "CabinSafeBench" benchmark, which employs "controlled context intervention" to alter safety-relevant environmental variables while keeping user intent and task goals constant; this allows for the isolation of "task comprehension failures" from "safety adaptation failures." Furthermore, the authors created "CockpitWorld," a closed-loop simulation environment that integrates Android Automotive OS with the CARLA driving simulator, enabling actionable, state-transition-based safety evaluations. Experiments reveal that while the Task Success Rate (TSR) for six mainstream LLM-based cockpit agents is near-saturated (99.85%), the Safety Task Success Rate (STSR) is only 75.38%. Under controlled context intervention, the TSR remains virtually unchanged, whereas the STSR plummets from 99.81% to 40.19%—a significant finding that highlights a critical flaw: strong task-execution capability does not equate to robust safety adaptability.

Critical strengths
  1. Clear and Significant Problem Definition: The paper precisely identifies and formalizes the challenge of "context-adaptive safety"—a critical yet under-researched area. In safety-critical domains such as autonomous driving and intelligent cockpits, distinguishing between "knowing what to do" and "knowing how to do it safely" is paramount. This work fills a notable gap in existing evaluation frameworks, offering significant theoretical and practical value.
  2. Innovative and Rigorous Methodology: The proposed "controlled contextual intervention" method is a key highlight. By keeping task objectives constant while varying environmental states to construct paired samples, the authors cleverly decouple failures in task comprehension from failures in safety adaptation. Furthermore, the integration of "CockpitWorld"—a closed-loop, executable environment—transforms safety assessment from subjective, language-based judgment into an objective, reproducible metric grounded in actual state evolution. This experimental paradigm is highly versatile and thought-provoking.
Critical weaknesses
  1. Need for Clarification on the Origin and Universality of Safety Rules: The paper mentions that safety rules are derived from sources such as "regulatory guidance" and "vehicle/OEM operating guidance," yet it lacks a detailed discussion regarding the specific origins and authority of these rules. For instance, does a rule like "opening a window while driving at high speed is unsafe" apply universally to all vehicle models? Do these rules vary across different countries or regulatory jurisdictions? Readers may question whether these benchmark rules possess sufficient universality and representativeness, which could impact the overall credibility of the benchmark itself. The authors should provide a more transparent discussion—either in the main text or an appendix—regarding the rationale behind the selection of the safety rules and their potential limitations. [1]
  2. The interpretation of the "SafetySupervisor" experiment requires greater caution. The paper demonstrates that a SafetySupervisor utilizing a shared set of rules can significantly restore safety performance. While the authors clarify that this serves as "shared-rule recoverability evidence" rather than an "independently deployable defense," the experimental design itself carries a risk of circular reasoning. Since the rules employed by the SafetySupervisor are derived from the same constraints used in the evaluation benchmarks, the process essentially amounts to correcting model outputs using the "answer key." This experiment primarily validates the effectiveness of the rules themselves, rather than proposing a viable, general-purpose safety enhancement solution. It is recommended to downplay this component as a primary conclusion or, alternatively, to reframe it as an "upper-bound analysis."[2][3]

[1] Luque-Cerpa A, Wang M, Carlsson E, et al. Learning contextual runtime monitors for safe AI-based autonomy[C]//2026 ACM/IEEE 17th International Conference on Cyber-Physical Systems (ICCPS). IEEE, 2026: 130-141.

[2] Meng X, Zhang Y, Huang Z, et al. DSBench: A Comprehensive Benchmark for Evaluating External and In-Cabin Risks[J]. arXiv preprint arXiv:2511.14592, 2025.

[3] Kirmayr J, Stappen L, André E. Car-bench: Evaluating the consistency and limit-awareness of llm agents under real-world uncertainty[C]//Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026: 40599-40618.

Questions for authors
  1. Regarding the generalizability of safety rules: Could you provide further details on the origin of these safety rules? For instance, in the case of a rule such as "it is unsafe to open vehicle windows at high speeds," have factors like variations in vehicle airtightness, sound insulation, or air conditioning systems been considered? Are there edge cases where such a rule might not be fully applicable?

  2. Regarding potential bias in model selection: The paper evaluates six state-of-the-art closed-source models, which differ significantly in terms of training data, architecture, and core capabilities. Beyond differences in "context-adaptive safety" proficiency, could the substantial variance in experimental results (e.g., the 90% gap in CAFR between Gemini and DeepSeek) also stem from variations in how each model interprets specific system prompt formats or their differing levels of aptitude in tool use? Have any ablation studies been conducted to control for these variables?

2 comments
AuthorsOpenReview ↗

Safety Rule Validity and Evaluation Robustness

We thank the reviewer for recognizing the significance of our problem formulation and evaluation methodology. We address the concerns below.

W1 & Q1: Safety Rule Provenance and Generalizability.

CabinSafeBench evaluates context-dependent safety constraints rather than establishing universal safety prohibitions. The applicability of a safety constraint depends on the operating conditions and vehicle assumptions under which it is defined. This distinction is central to our formulation of context-adaptive safety.

For example, opening a window at high speed is not universally unsafe. The relevant question is whether the action satisfies the safety requirements under the specified operating context. Our benchmark formalizes such requirements as executable constraints, drawing on regulatory guidance, vehicle/OEM operating guidance, and operational safety considerations.

We will further clarify rule provenance, applicability conditions, and operational assumptions in the appendix, distinguishing externally established requirements from benchmark-specific constraints. This ensures that safety judgments are interpreted within their defined operational scope.

W2: Interpretation of the SafetySupervisor Experiment.

The shared-rule configuration is an intentional experimental design. SafetySupervisor serves as an oracle-informed diagnostic reference to investigate a specific question: To what extent can observed safety failures be recovered when the relevant safety constraints are explicitly available during execution?

By providing the supervisor with the same constraints used for evaluation, the experiment isolates the effect of explicit safety-rule availability on agent execution. Its contribution lies in quantifying safety recoverability under known constraints, rather than establishing an independent safety defense mechanism.

We will clarify this diagnostic interpretation and distinguish the recoverability analysis from the benchmark's primary findings.

Q2: Prompt Sensitivity and Tool-Use Confounding.

The central evidence for context-adaptive safety failures comes from within-model paired comparisons, rather than absolute performance differences between models.

Under controlled context intervention, each model is evaluated with consistent task objectives, prompt structure, and tool interfaces, while safety-relevant environmental conditions are varied. Consequently, model architecture and general tool-use proficiency remain fixed within each paired comparison. This design isolates the performance changes associated with context intervention from differences in baseline model capabilities.

The results demonstrate a clear separation between functional task completion and safety adaptation: TSR remains near saturation (99.85% overall), whereas STSR decreases from 99.81% to 40.19% under context intervention. This substantial divergence demonstrates that successful task execution does not necessarily entail safe execution under changing conditions.

Furthermore, CabinSafeBench measures end-to-end context-adaptive safety, encompassing instruction interpretation, decision-making, and tool execution. Cross-model CAFR differences characterize the variation in this integrated capability under a standardized evaluation protocol, while the within-model intervention results provide the primary evidence for the task–safety gap.

Thank you for the response. It addresses most of my concerns. My remaining concern is the provenance of the safety rules. It is still unclear which rules are based on regulatory or OEM guidance, which were defined for the benchmark, and under what conditions each rule applies. I will therefore keep my original score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.