{"id":"29ed638e-0da1-4ba4-873c-a0158b0bf710","arxiv_id":"2411.10487","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A systematic mapping study identifies ten architectural patterns, seven for the quantum-classical split and three for middleware, that describe how quantum components can be integrated into AI inference systems.","lead":"This paper reviews 113 papers on quantum AI systems and extracts a catalog of ten architectural patterns for mixing classical and quantum components. It is a reference for software architects deciding how to split work between quantum and classical machinery.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pattern catalog relies on an unreviewed manual classification; missing protocol and internal labeling errors (S93, S125) make the mapping of 113 studies to ten patterns unreproducible.","rationale":"Reading the paper in good faith, I take the central contribution to be an evidence-based catalog of ten architectural patterns. The catalog is plausible and potentially useful: each pattern has a clear template, problem/solution framing, benefits and drawbacks, and concrete known uses, and several patterns align with earlier pattern language (Leymann 2019; Weigold et al. 2021b). The abstract and Section 4, however, make an empirical mapping claim, and that claim is only as strong as the manual classification of the 113 papers. The reader's weakest-assumption analysis identifies exactly this. My review finds concrete evidence that the classification is not yet reliable: the promised supplementary extraction forms are missing; no inter-rater reliability or coding scheme is reported; and the text contains internal labeling contradictions (S93 in SP-4 prose vs SP-6 table; nonexistent S125). These are not cosmetic because they change the pattern counts and call into question whether the seven-vs-three taxonomy would survive an independent audit. That said, the concern is fixable. The pattern concepts themselves are grounded in the literature, and the paper's threats section is otherwise candid about the immaturity of the evidence base. Therefore I do not move the verdict to reject; the appropriate stance remains conditional on the authors supplying the protocol and demonstrating coding reliability. My read is consistent with the reader's conditional verdict and with its identification of reproducibility as the load-bearing assumption.","tokens_in":39808,"tokens_out":8320,"duration_ms":84348,"concrete_test":"Perform an inter-rater reliability audit: provide two independent coders with the pattern definitions (including an explicit decision rule separating SP-4 Quantum Head from SP-6 Intermediate Quantum Layer), the inclusion criteria, and the list of 113 primary studies; ask each coder to assign every study to the pattern(s) it instantiates, then compute Cohen's kappa and compare per-pattern counts to Figure 11. If kappa is below 0.6 or more than 15% of assignments differ, the catalog is not a reproducible mapping and the paper must be reframed as a proposal rather than a systematic mapping result. As a secondary check, read S93 (Schetakis et al. 2022) and S33 (Sünkel et al. 2023) to decide which pattern(s) they actually implement, resolving the paper's internal discrepancies.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that a systematic mapping study of 113 primary studies yields ten architectural patterns in two families—depends on the reliability of the manual classification of each study into patterns. This condition is not met, and there is visible evidence of its failure. Section 2 refers to a protocol and full extraction forms 'in the supplementary materials,' but the arXiv version includes no supplementary materials, so no audit trail exists for the assignments. No coding scheme, decision rules, or inter-rater reliability checks are reported. More concretely, the paper contradicts itself in the classification: SP-4's known-use prose cites Schetakis et al. (2022) [S93] as a Quantum Head example, yet Table 2 places S93 under SP-6 (and SP-3) rather than SP-4; Section 5 cites [S125], a study ID that does not exist in Table 1, whose entries end at S113. Such errors indicate the labeling is not stable enough to support quantitative claims such as the pattern counts in Figure 11. The threats-to-validity section discusses simulation and benchmarking limitations but never addresses the absence of coding reliability, which is the standard threat for a qualitative synthesis of this kind. If the assignments are arbitrary, the 'ten patterns' result is an expert-proposed taxonomy rather than a reproducible empirical finding.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper reports a systematic mapping study of 113 papers on quantum-enhanced AI systems, from which the authors derive a catalogue of ten architectural patterns: seven quantum-classical split patterns (SP-1 through SP-7) and three quantum middleware patterns (MP-1 through MP-3). Each pattern is described with a template including problem, solution, benefits, drawbacks, and known uses, and the paper further discusses reasons for using quantum computing in AI (RQ2), emerging trends, and threats to validity. The central claim is that the ten patterns constitute a reproducible, evidence-based categorization of architectural solutions for quantum AI systems.","tokens_in":40051,"tokens_out":4092,"duration_ms":40853,"significance":"If the classification is reliable, the paper provides a useful catalogue of design alternatives with explicit trade-offs across quality attributes such as efficiency, scalability, trainability, simplicity, portability, and deployability, which is a valuable contribution to the emerging field of quantum software engineering. The study follows a recognized systematic mapping methodology (Petersen et al., 2015), covers a broad corpus including NISQ-era applications and middleware, and is honest about the limited evidence for quantum advantage. However, the central result depends on a manual, unreviewed classification of the 113 primary studies into patterns, and the manuscript currently provides no audit trail (protocol, extraction forms, coding scheme, or inter-rater checks) and contains visible labeling inconsistencies. This reproducibility gap is load-bearing, so the contribution is not yet established to systematic-mapping standards.","major_comments":[{"comment":"The pattern classification is not reproducible. Section 2 states that the protocol and full extraction forms are 'provided in the supplementary materials,' but the arXiv version includes no supplementary materials, so there is no audit trail for the assignment of the 113 studies to the ten patterns. No coding scheme, decision rules, or inter-rater reliability checks are reported. This is not a cosmetic gap: the entire catalogue is built on these manual assignments. Concrete evidence of instability appears in the text itself. In SP-4 (Quantum Head), the 'Known uses' paragraph cites Schetakis et al. (2022) [S93] as an example of the Quantum Head pattern, yet Table 2 lists S93 under SP-3 (Quantum Feature Engineering) and SP-6 (Intermediate Quantum Layer), not under SP-4. Similarly, Section 5 cites [S125] as evidence for inference speed-up and space complexity, but Table 1 ends at S113, so [S125] does not designate any primary study in the corpus. These inconsistencies indicate that the assignments are not stable enough to support the quantitative claims built on them, such as the pattern counts in Figure 11. The authors should provide the complete protocol and extraction forms, define the coding scheme, and report inter-rater reliability (or otherwise justify the objectivity of the assignments).","section":"Section 2 and Table 2 / Section 4.1"},{"comment":"The threats-to-validity section does not address the most important threat for a qualitative synthesis of this kind: the reliability of the manual pattern classification. Section 7 discusses the predominance of simulation-based results, the absence of comprehensive benchmarking, and the difficulty of establishing quantum speedup, but it never discusses the absence of a coding scheme, inter-rater reliability, or the possibility that the pattern assignments are subjective. Since the paper's primary contribution is the pattern catalogue, the classification procedure is a core methodological element, and the failure to treat it as a validity threat is a load-bearing omission.","section":"Section 7"},{"comment":"The pattern counts in Figure 11 are not well defined. Several studies appear in multiple pattern rows of Table 2 (for example, S93 appears in SP-3 and SP-6, S111 appears in SP-2, SP-3, and SP-4, and S113 appears in SP-6 and SP-7). The method does not state whether a primary study can be assigned to more than one pattern or how the numbers in Figure 11 were computed from the overlapping rows. Without this definition, the statement 'the most widely used pattern is the quantum monolith' (Section 6) and the other frequency comparisons are ambiguous and cannot be verified from the data presented.","section":"Figure 11 and Table 2"}],"minor_comments":[{"comment":"In the 'Known uses' paragraph of SP-4, the prose cites S93 as a Quantum Head example, but Table 2 places S93 under SP-3 and SP-6; this internal contradiction should be resolved either by correcting the text or by reclassifying the study.","section":"Section 4.1, SP-4"},{"comment":"The reference to [S125] is invalid because Table 1 contains only S1 through S113; the citation should be replaced with the actually intended reference, or the list of primary studies should be extended if S125 exists.","section":"Section 5"},{"comment":"The claim that 'all 113 selected papers on quantum AI systems discuss hybrid quantum-classical software architectures' is stated without qualification, but some entries in Table 1 (e.g., S44 'Layered Architecture for Quantum Computing' and S92 'Relevance of Near-Term Quantum Computing in the Cloud') do not appear to describe hybrid quantum-classical AI architectures; this generalization should be either substantiated or softened.","section":"Section 4.1, introductory paragraph"},{"comment":"There is a typo in the phrase 'In the course of the systemtic mapping study' — 'systemtic' should be 'systematic'.","section":"Section 3"},{"comment":"There are several spelling errors, including 'diminsionality' in SP-4, 'Deployablity' in MP-2, and 'efficiency' in SP-1 and SP-7; a careful proofreading pass is needed.","section":"Various"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely topic and the pattern catalogue could be a useful contribution to JSS if the methodological reproducibility issues are resolved. The main concern is that the systematic mapping is presented as an empirical, reproducible result, but the pattern assignments are not auditable and contain visible errors. I would encourage the editor to require the authors to provide the full protocol, extraction forms, and coding scheme (either as supplementary material or in an online repository), to report inter-rater reliability or otherwise justify the objectivity of the classification, and to correct the internal inconsistencies (S93, S125, overlapping pattern counts). The reference architecture resembles the authors' prior work (Lu et al. 2024b); the relationship should be made explicit to avoid any perception of self-derivation. These fixes are within the scope of a major revision and do not require a new study."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuine synthesis, not a token survey. The authors read 113 primary studies and organize the integration of quantum components into AI systems into seven split patterns and three middleware patterns, each with problem, solution, benefits, drawbacks, and known uses. Most ingredients exist in earlier work—quanvolution, quantum head, and service wrapper are not new names—but the quantum-AI-specific catalog with quality-attribute trade-offs is new and useful. The RQ2 discussion is also honest: no evidence that NISQ QML systematically beats classical accuracy, and most results come from simulations. That restraint raises my confidence in the survey portion.\n\nSoft spots are real but mostly fixable. The systematic-mapping claim outruns the evidence: the protocol, extraction forms, and quality-assessment details are promised in supplementary materials that are not present in the arXiv version, no coding scheme or inter-rater reliability check is reported, and the assignments in Table 2 drive the counts in Figure 11. That is a standard validity threat for this kind of synthesis, and the threats section does not address it. I can point to visible instability: the SP-4 known uses name Schetakis et al. (2022) as a Quantum Head example, but Table 2 places S93 under SP-6 and SP-3, not SP-4; Section 5 cites [S125], but the primary-study list ends at S113. These are exactly the labeling errors that make the catalog hard to rely on as a systematic mapping result.\n\nThe reference architecture in Section 3 is said to be similar to a prior self-cited paper (Lu et al. 2024b); that is not a flaw here, just a continuity note. Self-citation is not a problem because the cited papers are external or from the same group's prior line.\n\nBottom line: treat this as an expert-proposed taxonomy with a strong literature basis, not yet as a reproducible empirical mapping. For a reader working on quantum software architecture or QML systems, the catalog is worth engaging with; the pattern descriptions alone are a good basis for discussion. I would accept it for peer review and ask for the missing protocol, a coding/reliability report, and correction of the citation and labeling errors. That is a heavy-revision route, not a reject.","headline":"A useful pattern catalog for quantum AI software, built on a plausible but under-documented manual classification that needs cleaning before the mapping claims can be trusted.","tokens_in":40557,"tokens_out":2026,"would_cite":true,"duration_ms":19856,"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 paper claims to derive ten architectural patterns for quantum AI software from a systematic mapping study of 113 primary studies, seven for splitting work between quantum and classical inference and three for the middleware that…","keywords":["quantum AI","software architecture","architectural patterns","quantum machine learning","quantum-classical split","quantum middleware","NISQ","systematic mapping study"],"falsifier":"Independently re-classify the 113 primary studies using the paper’s pattern definitions and extraction form, with two or more coders working separately; if coders frequently assign the same study to different patterns, or if fewer than, say, 80 percent of assignments agree, then the claimed ten-pattern structure is not a stable, reproducible result.","tokens_in":39628,"feed_emoji":"🧩","tokens_out":8482,"duration_ms":82548,"temperature":0.7,"pith_summary":"The paper tries to turn the currently ad-hoc practice of building quantum-enhanced AI systems into a reusable design vocabulary. From a systematic mapping study of 113 primary studies, it claims to identify ten architectural patterns: seven “quantum-classical split” patterns that decide which parts of an inference pipeline run on a quantum computer, and three “quantum middleware” patterns that manage communication and orchestration between quantum and classical components. Each pattern is described with its problem, solution, benefits, drawbacks, and known uses, and each realizes a trade-off among software quality attributes such as efficiency, scalability, trainability, simplicity, portability, and deployability. The paper also surveys the reported reasons for using quantum components in AI and concludes that current NISQ-era evidence mostly shows accuracy comparable to classical methods rather than systematic practical advantage.","feed_headline":"Ten patterns show how to split AI between quantum and classical chips","feed_subtitle":"Catalogued trade-offs let architects pick patterns by efficiency, scalability, or deployability.","key_machinery":"The central object is the quantum-classical split, treated as a family of architectural patterns rather than a single design: the decision of which tasks are delegated to a quantum component and how classical and quantum components are arranged in an inference pipeline. It is complemented by the quantum middleware layer, the set of software services that manage routing, wrapping, and orchestration between quantum and classical systems. The study organizes these into a multi-layered reference architecture (operation, supply chain, system, and hardware layers) and classifies each of the 113 primary studies against a pattern template with name, summary, problem, solution, benefits, drawbacks, and known uses. That template is the mechanism that turns the literature into a comparable catalog.","core_discovery":"The central claim is that the architectural design space for quantum AI systems can be organized into ten named patterns, grounded in a systematic mapping study. The seven quantum-classical split patterns are: quantum monolith (inference fully delegated to a quantum component), multi-layer (several learnable quantum components connected classically), quantum feature engineering (quantum component performs initial feature extraction), quantum head (quantum component performs final inference after classical dimensionality reduction), quanvolution (small quantum circuits act as convolutional filters), intermediate quantum layer (quantum component between classical stages), and quantum accelerator (quantum component evaluates a well-defined function, usually without trainable parameters). The three middleware patterns are: service or microservice wrapper (wrapping quantum resources behind classical service interfaces), quantum API gateway (runtime routing of requests to the most suitable quantum computer), and quantum workflow orchestrator (automated deployment and scheduling of hybrid quantum-classical workflows). The paper further claims that all 113 surveyed systems are hybrid classical-quantum, making the split between classical and quantum work the central architectural decision.","pith_inferences":["The pattern catalog could be turned into a decision procedure: given a dataset’s dimensionality, NISQ constraints, and target quality attributes, a system architect could shortlist split patterns and middleware patterns mechanically.","The paper’s pattern assignments are testable: a replication with independent coders and an inter-rater reliability measure would either confirm the ten-pattern boundary or show that some patterns, such as intermediate quantum layer versus quantum head, are not crisply separable.","Because the survey finds no production deployments, the relative frequencies of patterns will likely shift once real systems are built; for instance, quantum accelerator and middleware patterns may become more prominent as fault-tolerant hardware matures.","The quality-attribute trade-offs stated per pattern (e.g., efficiency versus portability) could be tested on real hardware by measuring wall-clock time, network latency, and deployment effort for two implementations of the same task using different patterns."],"forward_implications":["An architect can choose among the seven split patterns by quality attribute: for example, quantum monolith maximizes simplicity and deployability but sacrifices scalability on NISQ hardware, while quantum head and intermediate quantum layer trade communication overhead for compatibility with high-dimensional data.","The three middleware patterns make quantum services hardware-agnostic: wrapping quantum resources as services, routing requests through an API gateway, and orchestrating workflows let applications shift among vendors and quantum computers at runtime.","The catalog gives quantum AI research a shared vocabulary, so new system designs can be described as instances of named patterns rather than one-off hybrids.","If the survey’s reading of the evidence is right, practitioners should not expect systematic accuracy advantages from NISQ-era quantum AI today; reported gains are mostly comparable accuracy, algorithmic speed-ups in simulation, parameter reduction, or robustness properties.","The identity of the most-used pattern (quantum monolith) and the least-used (multi-layer) can guide where near-term research effort is focused."],"supporting_citations":[{"why":"supplies the systematic mapping study guidelines that the entire methodology follows.","marker":"(Petersen et al., 2015)"},{"why":"introduces the idea of a pattern language for quantum software, which this catalog extends toward AI systems.","marker":"(Leymann, 2019)"},{"why":"defines patterns for hybrid quantum algorithms, the basis for the quantum-classical split family.","marker":"(Weigold et al., 2021b)"},{"why":"defines the NISQ device constraints that motivate the split and middleware patterns.","marker":"(Preskill, 2018)"},{"why":"provides the reference architecture that the paper adapts for quantum-enhanced AI systems.","marker":"(Lu et al., 2024b)"},{"why":"supplies the quantum convolutional neural network design that underlies several split patterns such as quanvolution.","marker":"(Cong et al., 2019)"}],"fun_headline_variants":["10 architectural patterns map quantum AI design space","Quantum AI: 10 patterns for splitting classical and quantum workloads","New catalog lays out 10 patterns for hybrid quantum-classical AI","Study finds 10 ways to architect quantum-enhanced AI systems","Ten patterns revealed for hybrid quantum-classical AI architecture"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire ten-pattern catalog rests on the assumption that the authors’ manual assignment of the 113 selected papers to the named patterns is objective and reproducible; if that classification is arbitrary, the catalog’s boundaries lose their validity.","fun_headline_variants_meta":{"raw":{"variants":["10 architectural patterns map quantum AI design space","Quantum AI: 10 patterns for splitting classical and quantum workloads","New catalog lays out 10 patterns for hybrid quantum-classical AI","Study finds 10 ways to architect quantum-enhanced AI systems","Ten patterns revealed for hybrid quantum-classical AI architecture"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000543,"raw_usage":{"total_tokens":2581,"prompt_tokens":909,"completion_tokens":1672,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":1592}},"tokens_in":525,"tokens_out":1672,"duration_ms":11003,"temperature":1.0,"reasoning_tokens":1592,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:55:04.879612+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Independently re-classify the 113 primary studies using the paper’s pattern definitions and extraction form, with two or more coders working separately; if coders frequently assign the same study to different patterns, or if fewer than, say, 80 percent of assignments agree, then the claimed ten-pattern structure is not a stable, reproducible result.","supporting_citations":[{"cited_title":", year 2019","cited_arxiv_id":null,"evidence_quote":"introduces the idea of a pattern language for quantum software, which this catalog extends toward AI systems."}],"review_version":1}