- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes ARGOS, a two-stage LLM-based framework for automatically generating Functional Safety Requirements (FSRs) from open-ended instructions for embodied agents. In the first stage, the framework extracts task-relevant entities and retrieves physical attributes from an expert-curated knowledge base containing 82 typed profiles. It then uses structured prompting to combine these attributes and generate potentially hazardous interaction scenarios. In the second stage, the generated hazards are combined with robot capability specifications and regulatory information to synthesize FSRs.
The authors evaluate ARGOS on 48 seed scenarios expanded into 365 hazard descriptions, using DeepSeek-V3.2 and GPT-4o. The evaluation includes comparisons with vanilla prompting, physics-aware chain-of-thought, full-rulebook long-context prompting, and an STPA/FMEA-structured baseline. Generated outputs are assessed using LLM-based scoring and human evaluations, with additional ablations examining the contributions of physical attribute grounding and regulatory retrieval.
Critical strengths
- Practically relevant problem formulation.
Automating hazard analysis and functional safety requirement generation for open-ended embodied tasks is a meaningful problem. The attempt to identify multi-factor hazards through reusable physical attributes, rather than relying exclusively on predefined task-level hazard descriptions, provides a potentially useful direction for safety engineering.
- Inclusion of stronger baselines and human evaluation.
The paper goes beyond comparisons with basic prompting methods by including full-rulebook long-context and STPA/FMEA-structured baselines. The additional blinded human evaluation and paired statistical analysis in Appendix H provide useful evidence for assessing the quality of generated outputs. These efforts improve the empirical presentation, although they do not resolve my main concerns about methodological novelty and the validity of the central scientific claims.
Critical weaknesses
- Limited methodological novelty and insufficient evidence for a distinct combinatorial reasoning mechanism.
The proposed framework appears to rely primarily on a combination of established techniques, including expert-curated knowledge retrieval, embedding-based matching, structured prompting, and LLM-based text generation. In particular, the central combinatorial reasoning operation in Section 3.1 is implemented through an LLM prompted with combinations of retrieved physical attributes. Appendix A.1 further identifies the Risk Injection Block, essentially a structured prompting template, as the core methodological contribution.
While such a pipeline may be practically useful, it is unclear what new generalizable reasoning or learning mechanism is introduced beyond existing retrieval-augmented generation and structured safety-analysis prompting.
More importantly, the experiments do not sufficiently isolate the contribution of the proposed combinatorial reasoning procedure. Although the authors include full-rulebook long-context and STPA/FMEA baselines, these comparisons do not directly establish whether explicit attribute combination provides benefits beyond supplying the same relevant physical knowledge through a simpler prompt. A controlled comparison using identical retrieved attributes, backbone models, and generation budgets, while varying only the proposed combination mechanism, is needed to support the claimed methodological contribution.
Consequently, I find the methodological novelty insufficiently established for ICLR.
- Insufficient validation of open-world hazard discovery, knowledge-base robustness, and safety requirement correctness.
A fundamental limitation is the framework's dependence on an expert-curated knowledge base containing only 82 typed physical profiles. Although the authors argue that combining reusable attributes avoids explicitly enumerating every hazard, successful discovery still depends on the relevant physical factors being represented, correctly retrieved, and appropriately combined.
The paper does not systematically evaluate what happens when essential physical attributes are absent from the knowledge base, when retrieval produces incorrect or incomplete results, or when hazards involve previously unseen physical mechanisms. The retrieval audit in Appendix H.4 evaluates matching against the existing expert-defined profiles, rather than whether those profiles provide sufficient coverage of hazards in genuinely novel scenarios. Therefore, the reported retrieval accuracy does not establish the claimed open-world scalability.
Furthermore, the evaluation primarily measures the perceived quality of generated hazard descriptions and FSRs rather than their correctness and completeness against independently established safety references. Although expert evaluations are included, these assessments do not establish hazard discovery recall, the physical validity of the proposed causal interactions, or the correctness and verifiability of the resulting safety requirements.
The evaluation rubric in Appendix E.2 also explicitly rewards multi-factor coupling, which is closely aligned with the structure enforced by the proposed generation prompts. This raises concerns about whether the reported improvements reflect genuinely better hazard discovery or greater conformity to the evaluation criteria.
The evidence against the strongest long-context baseline is also less conclusive than the main results suggest. In the additional 12-seed blinded audit (Appendix H, Tables 9–10), the overall improvement is only 0.153 points, with a paired sign-flip test p-value of 0.0898. The long-context baseline also performs better on the Long-tail Risk dimension in this audit.
Taken together, these limitations make it difficult to conclude that ARGOS provides a reliable, scalable, or substantially novel approach to discovering hazards in open-ended embodied environments. The current experiments primarily demonstrate improvements in the quality of structured safety-analysis text, rather than establishing the broader open-world hazard discovery capabilities claimed by the paper.
Questions for authors
Q1. What is the independent contribution of the proposed combinatorial reasoning mechanism?
Section 3.1 and Appendix A.1 describe the central combinatorial operation primarily as structured LLM prompting over retrieved physical attributes. Can the authors clarify what distinguishes this mechanism from conventional retrieval-augmented prompting? More importantly, can they provide a controlled comparison in which the same retrieved attributes, backbone model, and generation budget are used, with and without the proposed combination-specific prompting? This would help determine whether the improvement comes from the proposed reasoning mechanism or simply from supplying relevant physical knowledge.
Q2. How does ARGOS handle incomplete knowledge and retrieval failures?
The framework relies on 82 expert-curated physical profiles. What happens when a task involves a critical physical property or hazard mechanism that is not represented in this knowledge base? Similarly, how does the system respond when retrieval fails to return a relevant attribute or introduces an incorrect one? Do the authors have quantitative evidence from controlled knowledge-removal, retrieval-error, or unseen-hazard experiments? Such evidence is important for evaluating the claimed applicability to open-ended environments.
Q3. What evidence establishes the actual correctness and coverage of generated hazards and FSRs?
The reported evaluations focus primarily on rubric-based assessments of generated descriptions. Can the authors provide independent evidence that the identified multi-factor interactions are physically valid, that important hazards are not systematically omitted, and that the synthesized FSRs are consistent with the applicable regulatory and robot capability constraints? In particular, are there expert-validated reference hazards or safety requirements against which discovery recall, false hazard generation, and requirement correctness can be evaluated?
Answers to these questions would materially affect my assessment of both the methodological novelty and the strength of the experimental validation.
5 comments
Q2: Incomplete knowledge bases and retrieval failures
The 82-profile knowledge base should be understood as a collection of reusable physical primitives rather than a catalog of predefined hazards.
Its purpose is to provide physical premises that can be composed across different tasks and scenarios. The 48 seed instructions are expanded into 365 context-specific hazard scenarios through attribute retrieval and bounded interaction reasoning.
Importantly, we distinguish retrieval accuracy from knowledge-base completeness.
Retrieval reliability.
ARGOS uses BGE-based semantic retrieval with a configurable similarity threshold. We evaluate thresholds from 0.5 to 0.9.
At , a case-by-case audit found no hallucinated or missed profiles within the evaluated household-service subset. Lowering the threshold to 0.6 introduces approximately 25% hallucinated retrievals, whereas thresholds of 0.8 or above begin to miss valid semantic matches.
The framework also provides a configurable Hybrid Hard-Match Fallback, under which exact lexical matches override embedding disagreements for known entries. This combines semantic matching with deterministic handling of explicitly represented profiles.
Knowledge-base coverage.
The reviewer correctly distinguishes matching an existing profile from establishing that the knowledge base contains every relevant physical mechanism. These are different questions.
ARGOS does not claim that successful retrieval from the existing knowledge base proves exhaustive physical coverage. If a critical physical property or mechanism is absent from the knowledge base, the current pipeline does not provide a formal guarantee that it will independently recover that missing information.
Likewise, the current experiments do not include controlled knowledge-removal, retrieval-error injection, or systematically held-out unseen-mechanism evaluations. We do not interpret the retrieval audit as evidence of universal open-world coverage.
The contribution of the current system is compositional hazard discovery using reusable physical attributes, within a defined household-service domain and bounded interaction space.
We will revise the manuscript to explicitly distinguish retrieval reliability from knowledge-base coverage, clarify the scope of the 82 profiles, and add a limitation statement concerning missing physical attributes and previously unseen mechanisms. We will also narrow the open-world generalization claims to avoid implying guaranteed discovery beyond the represented physical knowledge.
Q3: Correctness and coverage of generated hazards and FSRs
The evaluation of ARGOS involves distinct criteria at different stages, and these should not be conflated.
1. Physical validity and seed consistency
The Stage-I evaluation measures Physical Reliability, Long-tail Risk, and FSR Derivation Capability.
The evaluation is not based merely on whether a generated scenario contains multiple factors. Physical Reliability explicitly assesses real-world physical plausibility and adherence to the seed's closed-world constraints. The generation procedure also requires the hazard to be causally attributable to the interaction of the specified factors.
In addition, three robot-safety researchers independently evaluated 108 anonymized and shuffled scenarios. ARGOS obtained the highest human mean scores across the three dimensions:
| Metric | ARGOS | CoT | Vanilla |
|---|---|---|---|
| Physical Reliability | 7.94 | 7.34 | 6.97 |
| Long-tail Risk | 7.26 | 5.29 | 6.43 |
| FSR Derivation | 7.34 | 5.46 | 6.49 |
These results provide independent human assessments of the physical quality and requirement relevance of the generated hazard descriptions.
2. Interpretation of the multi-factor evaluation criterion
The reviewer notes that the Long-tail Risk rubric explicitly rewards multi-factor coupling.
This criterion is intentional: discovering coupled physical risks is the central objective of ARGOS. However, the metric does not simply count the number of mentioned factors. It evaluates whether the generated scenario identifies a credible physical mechanism, remains within the seed context, and reveals a meaningful safety boundary or secondary consequence.
Physical Reliability and FSR Derivation are also evaluated separately, preventing the Long-tail Risk score from being interpreted as a complete measure of overall scenario correctness.
Furthermore, the three human evaluators assessed anonymized outputs using the same predefined rubric, providing an assessment beyond the automatic generation model's own output.
We will revise the manuscript to explain more explicitly why multi-factor coupling is part of the target construct, distinguish it from mere factor enumeration, and clarify the complementary roles of the three evaluation metrics.
3. FSR correctness and hardware compatibility
Stage II uses a dedicated audit protocol that evaluates hardware compatibility, physical risk coverage, and logical robustness.
The audit explicitly checks whether each FSR assumes unsupported sensor capabilities, violates hardware limitations, misses secondary physical consequences, or lacks coherent triggering and recovery conditions.
The ablation results show that ARGOS achieves the highest Overall FSR quality on the primary backbone (7.97) and maintains its advantage in the controlled GPT-4o study (7.59).
We emphasize that these results support requirement-level quality and hardware consistency under the declared evaluation protocol. They are not equivalent to independent certification of the applicable standards clauses or physical validation of an implemented controller.
We will revise the manuscript to make the validation scope explicit and distinguish the current rubric-based requirement audit from downstream expert certification, TSR verification, and physical robot testing. We will also clarify the role of regulatory retrieval in connecting the hazard descriptions to the available standards knowledge base.
4. Interpretation of the strongest baseline comparison
We agree that the 12-seed blinded audit should be interpreted according to its paired statistics rather than relying solely on the Overall mean.
ARGOS obtains an Overall score of 8.45, compared with 8.30 for full-rulebook long-context prompting. The Overall difference is 0.153 (), which is not statistically significant.
However, the dimension-level results are informative:
- Physical Reliability: +0.283 ().
- Long-tail Risk: −0.192 ().
- FSR Derivation: +0.367 ().
Long-context performs better on Long-tail Risk in this subset, while ARGOS demonstrates significant improvements in Physical Reliability and FSR Derivation.
We therefore do not claim statistically significant Overall superiority over long-context prompting in this 12-seed audit. The evidence supports a more specific conclusion: explicit physical grounding and bounded interaction composition improve physical reliability and the derivability of safety requirements.
We will revise the manuscript's Abstract, Results, and Discussion to make this distinction explicit and avoid overstating the strength of the Overall comparison.
Overall methodological positioning
ARGOS addresses a specific problem in embodied-AI safety engineering: deriving physically grounded hazard hypotheses and functional safety requirements from open-ended natural-language task instructions.
Its contribution is not unrestricted physical-world discovery or a new general-purpose LLM learning mechanism. It is a structured two-stage procedure that connects:
- Reusable typed physical attributes.
- Bounded, causally constrained interaction reasoning.
- Scenario-specific hazard descriptions.
- Regulatory- and hardware-grounded FSR synthesis.
The experiments demonstrate the benefits of this complete procedure across multiple backbones, stronger baselines, ablation studies, and human evaluation.
We will revise the Abstract, Introduction, Methodology, and Discussion to make the scope, methodological contribution, evaluation boundaries, and limitations more precise. In particular, the revised manuscript will consistently describe ARGOS as an offline design-time framework for open-ended task instructions with bounded physical-attribute composition, rather than suggesting unrestricted open-world coverage or runtime safety control.
Rebuttal to Reviewer 9uP9
We appreciate the reviewer's recognition of the practical relevance of automated hazard analysis for embodied AI, as well as the acknowledgment of the stronger baselines and blinded human evaluation.
The central issue is how the methodological contribution and experimental evidence of ARGOS should be interpreted. We clarify these points below.
Q1: Independent contribution of the combinatorial reasoning mechanism
ARGOS does not claim to introduce a new neural architecture or a new general-purpose learning mechanism. Its contribution is a structured physical reasoning methodology for hazard discovery and FSR synthesis.
The distinction from conventional retrieval-augmented prompting lies in how retrieved knowledge is represented, constrained, and used.
In conventional retrieval-based prompting, retrieved information is generally supplied as additional context for text generation. ARGOS instead treats typed physical profiles as explicit reasoning premises and constructs bounded interaction hypotheses from them.
The proposed procedure has four connected components:
This is more specific than simply retrieving related text and asking an LLM to generate hazards.
For example, in the Rose-and-Balloon case, the relevant physical properties are not merely supplied as descriptions. The generation procedure explicitly requires reasoning about their interaction: a rigid thorn and a low-puncture-resistance membrane jointly produce a puncture mechanism. This leads to a force-level safety constraint rather than an unrelated generic speed limit.
The contribution is therefore the operational formulation of attribute-grounded, bounded combinatorial hazard reasoning and its integration into downstream FSR synthesis. We do not rely on a claim that the Risk Injection Block alone constitutes a new learning architecture.
Regarding the experimental isolation of combinatorial reasoning:
The full-rulebook baseline receives all 82 physical profiles in one prompt without truncation. Both ARGOS and this baseline use the same backbone, seed inputs, robot capability profiles, target counts, and evaluation rubric.
This comparison tests whether presenting the complete physical knowledge base jointly is equivalent to explicitly selecting and composing relevant physical premises. The results show that it is not: ARGOS obtains higher Physical Reliability and FSR Derivation, while the full-rulebook baseline remains competitive on Long-tail Risk.
The comparison therefore provides evidence for the value of structured physical knowledge organization and interaction composition.
We also distinguish this evidence from a stricter causal ablation. The current experiments do not include a fully matched comparison that holds the exact retrieved attributes, generation budget, and backbone constant while removing only the combination-specific interaction mechanism. We do not claim that the existing full-rulebook comparison independently isolates the contribution of that single component.
We will revise the manuscript to distinguish the methodological contribution of the complete ARGOS procedure from the causal attribution of individual components. We will also clarify that the current baseline comparisons establish the advantage of the complete structured workflow, rather than claiming that they independently isolate the effect of combinatorial reasoning under identical token and generation budgets.