{"id":"8ea6175c-7cd1-4f0e-b8b4-1d1dda29750f","arxiv_id":"2508.07163","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A survey mapping how neurosymbolic AI could address safety, regulatory, and operational challenges in advanced air mobility.","lead":"This paper reviews how neurosymbolic AI, which combines neural networks with symbolic rules, could be applied to the safety, regulation, and operations of advanced air mobility systems. The survey maps the research landscape and claims to be the first comprehensive roadmap for this intersection, making it a starting point for researchers and funders.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The survey's 'fragmented research landscape' premise is weakly evidenced: most Section 3 citations are adjacent-domain or speculative, so the 'comprehensive survey of integration' may overstate existing work.","rationale":"The reader's verdict is CONDITIONAL, focusing on the unsupported 'first review' claim, self-citations, and a battery energy-density misstatement. My stress-test concedes those points but identifies a deeper, more load-bearing issue: the survey's substantive content may not actually document a body of neurosymbolic-AAM research. If the cited works are mostly adjacent-domain or general, then the paper's claim to 'classify current advancements' and present a 'fragmented research landscape' is not supported by the evidence it marshals. This affects the central claim directly: a survey that catalogs potential applications rather than actual integrations should be framed as a research agenda, not a comprehensive survey. However, this is not grounds for rejection: a forward-looking roadmap can still be useful, and the paper is coherent as a proposal. The condition is that the authors must either demonstrate that a sufficient corpus of AAM-specific neurosymbolic work exists (via the concrete test) or revise the paper's title, abstract, and conclusion to explicitly state that it is a survey of opportunities and adjacent work, not existing integrated systems. The battery-energy-density error (300–400 Wh/kg and 'over 800 Wh/kg' for narrow-body aircraft in Section 3.1) is a factual mistake but secondary to the argument; it should be corrected but does not undermine the neurosymbolic thesis. I disagree with any stronger verdict because surveys are not expected to contain formal proofs, and the organizational structure, case studies (FAA/EASA), and risk taxonomy have value as a framework. The agreement is 'partial' because the reader's weakest_assumption centers on novelty and self-citation, while my concern is about the evidentiary basis for the alleged existing research landscape—related but distinct.","tokens_in":11882,"tokens_out":2759,"duration_ms":26361,"concrete_test":"For each of the eight domains in Section 3, classify every cited reference as: (a) implemented neurosymbolic system evaluated in an AAM context, (b) adjacent-domain neurosymbolic, (c) general neurosymbolic/methodology, or (d) non-neurosymbolic AAM background. Then count how many domains have at least one (a) reference. If fewer than 4 of 8, the 'comprehensive survey of integration' claim is unsupported and the paper should be re-scoped as a research agenda. Also check whether Acharya et al. 2025 is peer-reviewed.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central value claim is that it surveys a fragmented but real body of research on neurosymbolic AI applied to AAM, and classifies 'current advancements' (Abstract, Section 1, Section 3). Yet the citations in Section 3 reveal that most referenced works are not neurosymbolic-AAM integrations. For example: Section 3.4 Predictive Maintenance cites Yang et al. 2018 (fault diagnosis for embedded software) and Siyaev et al. 2023 (industrial digital twin), neither AAM-specific. Section 3.5 Safety cites Jansen et al. 2020 (probabilistic shields, general safe RL), Sharifi et al. 2023 (autonomous driving), and Kimura et al. 2021 (logical neural networks, not aviation). Section 3.6 Autonomy cites Wang et al. 2024, a general neurosymbolic-computing survey. Section 3.7 Cybersecurity cites Grov et al. 2024 (cyber attacks, not AAM) and Petit and Shladover 2014 (automated vehicles). Only Section 3.8 Demand Modeling cites directly AAM-relevant neurosymbolic work, and that work includes the authors' own arXiv preprint (Acharya et al. 2025) plus Kohaut et al. 2024 (ProMis). Sections 3.1 and 3.2 have no neurosymbolic citations at all, simply asserting neurosymbolic potential. Thus, the claimed 'fragmented research landscape' may actually be nearly empty; the paper is better characterized as a proposal for how neurosymbolic techniques could be applied, not a survey of existing integrations. If this is true, the 'comprehensive survey' framing and its implicit roadmap-to-date are not supported by the cited evidence. This is a correctness risk for the survey's central claim, independent of the 'first review' novelty assertion.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper surveys the intersection of Neurosymbolic AI and Advanced Air Mobility (AAM). It reviews background on neurosymbolic paradigms and AAM technologies, then organizes applications into eight domains (electrification, aircraft design, training/simulation, predictive maintenance, safety, autonomy, cybersecurity, demand modeling). It presents FAA and EASA AI roadmaps as case studies, discusses technological, ethical, and regulatory challenges, and concludes that neurosymbolic AI promises transparent, adaptable, high-performance solutions for AAM. The authors explicitly claim this is the first comprehensive review of neurosymbolic AI in AAM and that it provides a roadmap for researchers and practitioners.","tokens_in":12305,"tokens_out":2445,"duration_ms":23720,"significance":"If the paper's framing holds, it would be a valuable organizing resource: it compiles disparate threads, links them to AAM domain requirements, and highlights certification and explainability challenges that are central to aviation deployment. The paper also gives concrete attention to FAA/EASA regulatory trajectories, which is useful. However, the central value claim—that there is a measurable, fragmented body of existing neurosymbolic-AAM integration work that this survey comprehensively classifies—is weakly supported by the cited evidence. Many cited works are adjacent-domain or general neurosymbolic studies, not AAM integrations. The paper also contains a clear factual error in the battery energy density claim, and the 'first comprehensive review' assertion is made without a documented search protocol. These issues affect the paper's core contribution as currently framed, though they are fixable through reframing and correction.","major_comments":[{"comment":"The text states lithium-ion batteries offer 'over 800 Wh/kg for narrow-body aircraft' and cites [Barrera et al., 2022]. This is a factual error: current lithium-ion cells are around 250–300 Wh/kg, and even next-generation concepts discussed in the cited paper are well below 800 Wh/kg at the battery level. This error casts doubt on the accuracy of the domain-specific statements throughout the survey. It should be corrected and re-verified against the cited source.","section":"§3.1 (Electrification)"},{"comment":"The paper's core claim of being 'the first comprehensive review' and of surveying 'ongoing research efforts' at the neurosymbolic-AAM intersection is not substantiated. No search protocol, inclusion criteria, or database coverage is provided. More importantly, the Section 3 citations do not support the existence of a substantial body of neurosymbolic-AAM integrations: §3.4 cites a fault-diagnosis paper for embedded software (Yang et al. 2018) and an industrial digital-twin paper (Siyaev et al. 2023) that are not AAM-specific; §3.5 cites probabilistic shields (Jansen et al. 2020), an autonomous-driving neurosymbolic RL paper (Sharifi et al. 2023), and logical neural networks (Kimura et al. 2021) with no AAM application; §3.6 cites a general neurosymbolic computing survey (Wang et al. 2024); §3.7 cites cyber-attack work on automated vehicles (Petit and Shladover 2014) and general neurosymb","section":"§1 (Introduction) and §3 (Application Areas)"},{"comment":"Several load-bearing claims about the promise of Neurosymbolic Reinforcement Learning and its benefits are supported exclusively or primarily by the authors' own prior work: [Acharya et al., 2023] in the Introduction and §5.2, [Acharya et al., 2025] in §3.8, and [Sharifi et al., 2023] in §3.5. At least two of these are arXiv preprints rather than peer-reviewed publications. For a survey whose contribution is synthesis, the authors should either justify the representativeness of these sources with independent corroboration or clearly flag the evidence level (e.g., 'the authors' own proposal') when using them to assert that a method 'has shown potential.' Otherwise the survey risks circularity in its central claims.","section":"§1, §2.1, §5.2 (reliance on self-citations)"}],"minor_comments":[{"comment":"There is a typo: 'UA Vs' should be 'UAVs.' Please also standardize spelling of 'unnanned/autonomous' terminology throughout.","section":"§3.6 (Autonomy)"},{"comment":"The paper uses 'illustrates' and 'depicts' for Figures 1–3, but the figures are not explicitly referenced in the surrounding text with the expected regularity. Ensure each figure is cited at the point where it is first discussed and that captions are self-contained.","section":"§1 and §2.2"},{"comment":"The sentence 'The FAA has made limited progress...' cites a DOT OIG audit report via a URL in a footnote. For a survey, please provide the full formal reference (report number, date, title) in the reference list so readers can verify the claim without following an unlabeled URL.","section":"§5.3 (Certification)"},{"comment":"Some references lack full page numbers or are arXiv preprints without a 'submitted/under review' status. For a survey, it would help to add DOIs where available and to mark preprint status consistently. Also, the 'Gilpin and Ilievski, 2021' reference has page numbers '15(3):123–145' that appear to be an invented journal-style volume; please verify this reference.","section":"References"},{"comment":"The FAA and EASA case studies are informative but are summarized from official documents without critical assessment. The paper does not explain how neurosymbolic AI specifically addresses the certification gaps beyond generic statements. A short table or bullet list mapping each roadmap gap to a concrete neurosymbolic mechanism would improve the utility of this section.","section":"§4 (Case Studies)"}],"recommendation":"major_revision","confidential_remarks":"The paper has promise as a research agenda, but its current 'comprehensive survey' framing is not supported by the evidence. The battery energy density error is concrete and should be corrected. The lack of a search protocol and the heavy reliance on adjacent-domain or self-cited works would likely attract criticism from readers familiar with either the neurosymbolic or the AAM literature. I recommend major revision, not rejection, because the paper can be reframed as a 'roadmap/research opportunities' review with an explicit methodology and careful separation of demonstrated results from conjectured applicability. The authors should also address the 'first' claim by either providing reproducible search details or softening it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—worth a skim, but not for the reasons the authors claim. This is a plausible mapping of where neurosymbolic methods could plug into AAM—demand modeling, safety, autonomy, certification—and it usefully summarizes the FAA and EASA AI roadmaps. If you need an entry point to that intersection, the taxonomy (electrification, design, predictive maintenance, etc.) is a reasonable organizing scaffold.\n\nThe problems are in the framing. The abstract and intro call it a comprehensive survey of current advancements, but the actual inventory in Section 3 is thin and mostly adjacent. Most citations are general neurosymbolic work or other domains—autonomous driving, industrial digital twins, cybersecurity—not AAM-specific integrations. Section 3.2 and 3.6 have no direct neurosymbolic-AAM citations at all; 3.5 leans on safe-RL shields and autonomous driving. The paper reads more like a proposal/roadmap than a survey of an existing fragmented landscape. That's not fatal, but the 'first comprehensive review' claim is unsupported and should be reworded to 'first roadmap' or 'research opportunities' framing.\n\nThere's also a concrete factual error in 3.1: it says lithium-ion batteries offer over 800 Wh/kg for narrow-body aircraft. That's off by a factor of two or more at cell level, and it muddles the short-range vs. narrow-body distinction. Needs a correct citation and number.\n\nOn self-citation: the load-bearing positive claims lean on the authors' own 2023 survey and a 2025 arXiv preprint, plus a co-author's autonomous-driving paper. That's not disqualifying, but the survey would be stronger with more independent evidence. The lack of a documented search protocol makes the 'no comprehensive review' claim unverifiable.\n\nBottom line: the organizational value is real; the empirical density is not. I'd send it to peer review with a request for major framing changes and fact-checking—this kind of survey can be a useful community resource if it says what it actually covers.","headline":"Useful roadmap with an overreaching 'comprehensive survey' claim and a battery error; worth refereeing after rework.","tokens_in":12777,"tokens_out":2649,"would_cite":false,"duration_ms":23438,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This survey argues that neurosymbolic AI — neural pattern learning fused with explicit symbolic rules — is the key to making advanced air mobility safe, transparent, and certifiable, and maps where the field stands.","keywords":["neurosymbolic AI","advanced air mobility","urban air mobility","eVTOL","neurosymbolic reinforcement learning","air traffic management","aviation safety","survey"],"falsifier":"Locate a peer-reviewed survey published before 2025 that already reviews neurosymbolic AI applications across multiple AAM domains; that single finding would reduce the paper's contribution from 'first comprehensive roadmap' to one more taxonomy.","tokens_in":11808,"feed_emoji":"🚁","tokens_out":5839,"duration_ms":51153,"temperature":0.7,"pith_summary":"The paper sets out to show that advanced air mobility (AAM) — electric vertical take-off and landing aircraft, drone delivery, and urban air taxis — faces problems pure deep learning cannot solve alone. It argues that neurosymbolic AI, which couples neural pattern recognition with explicit symbolic rules, is a natural fit for a safety-critical, regulation-heavy industry. The survey organizes a fragmented literature into eight application areas and claims to be the first review of the intersection as a whole. If the framing holds, it gives researchers, operators, and regulators a shared map of where hybrid methods already help, where they fall short, and what must be built before certification becomes feasible.","feed_headline":"Hybrid AI maps a path to certifiable air taxis","feed_subtitle":"Neural learning plus symbolic rules could meet the safety and regulatory demands that block flying taxis and drone delivery.","key_machinery":"The organizing mechanism is the hybrid neurosymbolic architecture itself: a neural component learns patterns from large, heterogeneous data streams while a symbolic component encodes airspace rules, safety protocols, and regulatory constraints. The survey uses this pairing as a lens to classify applications into learning-for-reasoning, reasoning-for-learning, and fully integrated learning-reasoning systems, and to evaluate which AAM tasks are ready for hybrid methods and which are not.","core_discovery":"The central claim is that neurosymbolic AI can deliver the transparency, adaptability, and performance that AAM demands, where neither pure neural nor pure symbolic systems suffice. Across electrification, aircraft design, training and simulation, predictive maintenance, safety, autonomy, cybersecurity, and demand modeling, the paper identifies concrete hybrid mechanisms: symbolic logic constraints that shield neural policies from unsafe actions, case-based reasoning combined with Bayesian networks for fault diagnosis, decision trees injected into neural networks for interpretable demand forecasts, and knowledge graphs plus logic tensor networks for cyber defense. The paper's own assessment","pith_inferences":["A natural next step the paper leaves implicit is a standardized benchmark: run neural-only and neurosymbolic policies on the same AAM conflict-resolution scenarios and measure safety violations, data efficiency, and runtime.","The evidence the survey assembles suggests progress will come as much from regulatory sandboxes and certification experiments as from new architectures, since the aviation-authority roadmaps it examines are where the hard requirements are set.","The same rule-plus-learning pattern may transfer to other safety-critical mobility domains — autonomous rail, maritime shipping, or ground logistics — where black-box decisions are the main adoption barrier.","A testable extension of the demand-modeling results is whether injecting symbolic constraints reduces the amount of training data needed to reach a target forecast accuracy in a new city or airspace."],"forward_implications":["Neurosymbolic reinforcement learning can optimize dynamic flight paths and traffic management in real time while keeping decisions inside regulatory and safety bounds.","Symbolic 'shields' can block unsafe actions during neural policy training, a concrete route toward the safety assurance aviation regulators require.","Demand forecasting gains interpretability without losing accuracy when rule-based models such as decision trees are embedded in neural networks.","The same hybrid pattern extends across the AAM stack, from battery and propulsion design to predictive maintenance and cyber defense, creating a unified research agenda.","The roadmap identifies certification of evolving, adaptive AI systems as a critical gap that must be closed before deployment can proceed."],"supporting_citations":[{"why":"Defines neurosymbolic AI as the 'third wave' of AI and supplies the foundational definition the survey builds on.","marker":"[Garcez and Lamb, 2023]"},{"why":"Provides the six-category taxonomy of neurosymbolic systems the survey uses to structure the field.","marker":"[Kautz, 2022]"},{"why":"Supplies the learning-for-reasoning / reasoning-for-learning / learning-reasoning classification used to organize applications.","marker":"[Yu et al., 2023]"},{"why":"Surveys neurosymbolic reinforcement learning and planning, the method the paper proposes for dynamic AAM optimization.","marker":"[Acharya et al., 2023]"},{"why":"Demonstrates a neuro-symbolic deep reinforcement learning approach for safe driving policies, the template for AAM safety.","marker":"[Sharifi et al., 2023]"},{"why":"Introduces probabilistic shields that filter neural-network outputs, the mechanism the paper cites for enforcing aviation rules.","marker":"[Jansen et al., 2020]"},{"why":"Reviews UAM demand analysis, the existing demand-modeling baseline neurosymbolic methods would improve.","marker":"[Long et al., 2023]"},{"why":"Presents Probabilistic Mission Design, an example of embedding legal frameworks into neuro-symbolic systems.","marker":"[Kohaut et al., 2024]"},{"why":"Reviews aircraft design concepts for AAM, grounding the aircraft-design application area.","marker":"[Kiesewetter et al., 2023]"}],"fun_headline_variants":["Neurosymbolic AI maps route to certified air mobility","Hybrid neural-symbolic methods tackle AAM's safety hurdles","Survey: neurosymbolic models key to trustworthy air taxis","Symbolic constraints make neural AI safer for skies","Blending logic and learning for certifiable air transport"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The paper's standing as a 'first roadmap' rests on the claim, made without a documented search protocol, that no comparable review of neurosymbolic AI in advanced air mobility already exists.","fun_headline_variants_meta":{"raw":{"variants":["Neurosymbolic AI maps route to certified air mobility","Hybrid neural-symbolic methods tackle AAM's safety hurdles","Survey: neurosymbolic models key to trustworthy air taxis","Symbolic constraints make neural AI safer for skies","Blending logic and learning for certifiable air transport"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000565,"raw_usage":{"total_tokens":2458,"prompt_tokens":628,"completion_tokens":1830,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":372,"completion_tokens_details":{"reasoning_tokens":1763}},"tokens_in":372,"tokens_out":1830,"duration_ms":13030,"temperature":1.0,"reasoning_tokens":1763,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:16:40.675457+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Locate a peer-reviewed survey published before 2025 that already reviews neurosymbolic AI applications across multiple AAM domains; that single finding would reduce the paper's contribution from 'first comprehensive roadmap' to one more taxonomy.","supporting_citations":[{"cited_title":"Neurosymbolic ai: The 3 rd wave","cited_arxiv_id":null,"evidence_quote":"Defines neurosymbolic AI as the 'third wave' of AI and supplies the foundational definition the survey builds on."},{"cited_title":"The third ai summer: Aaai robert s","cited_arxiv_id":null,"evidence_quote":"Provides the six-category taxonomy of neurosymbolic systems the survey uses to structure the field."},{"cited_title":"Neurosymbolic reinforcement learning and planning: A survey","cited_arxiv_id":null,"evidence_quote":"Surveys neurosymbolic reinforcement learning and planning, the method the paper proposes for dynamic AAM optimization."},{"cited_title":"Towards safe autonomous driving policies using a neuro-symbolic deep reinforcement learning ap- proach","cited_arxiv_id":null,"evidence_quote":"Demonstrates a neuro-symbolic deep reinforcement learning approach for safe driving policies, the template for AAM safety."},{"cited_title":"Safe re- inforcement learning using probabilistic shields","cited_arxiv_id":null,"evidence_quote":"Introduces probabilistic shields that filter neural-network outputs, the mechanism the paper cites for enforcing aviation rules."},{"cited_title":"Demand analysis in urban air mobility: A literature review","cited_arxiv_id":null,"evidence_quote":"Reviews UAM demand analysis, the existing demand-modeling baseline neurosymbolic methods would improve."},{"cited_title":"Probabilistic mission design in neuro-symbolic systems","cited_arxiv_id":null,"evidence_quote":"Presents Probabilistic Mission Design, an example of embedding legal frameworks into neuro-symbolic systems."},{"cited_title":"A holistic review of the current state of research on air- craft design concepts and consideration for advanced air mobility applications","cited_arxiv_id":null,"evidence_quote":"Reviews aircraft design concepts for AAM, grounding the aircraft-design application area."}],"review_version":1}