{"id":"81c0ff59-24b6-4678-920d-b2a3edbc29d6","arxiv_id":"2411.13740","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A survey that organizes federated continual learning methods into class, domain, and task continual learning categories for edge-AI applications.","lead":"This paper surveys work that combines federated learning with continual learning for edge devices, sorting methods by whether new classes, new domains, or new tasks arrive over time. It is a reference map for researchers and engineers who want AI systems that keep learning on distributed devices without forgetting older knowledge.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The survey's organizing taxonomy is not a partition: Cross-FCL appears in both FCCL (Sec. 2.3) and FTCL (Sec. 4.2), and CFeD appears in both FCCL (Sec. 2.5) and FDCL (Sec. 3.2).","rationale":"The reader's weakest assumption is the meaningfulness and exhaustiveness of the three-category taxonomy, and the manuscript supports that concern with two concrete double-assignments: CFeD in Secs. 2.5 and 3.2, and Cross-FCL in Secs. 2.3 and 4.2. The Cross-FCL case is especially damaging because the paper defines FCCL and FTCL by opposite conditions on task identity at test time, so one method cannot validly belong to both. This is not a matter of disagreement with the field's consensus; it is an internal inconsistency in how the central organizational scheme is applied. Other concerns, such as the unsupported 'first comprehensive' claim or mechanical notation errors, are secondary: the first can be softened without changing the survey's utility, and the notation errors do not affect the structural argument. Given that the taxonomy is the main claimed contribution, the conditional verdict remains appropriate: the survey is useful and summaries are probably sound, but the central organizational scheme needs an explicit decision rule or an acknowledged non-partition before it can be recommended as a reliable navigation tool. No change to the reader's verdict is needed; my independent check reinforces it rather than overturning it.","tokens_in":32855,"tokens_out":3057,"duration_ms":32214,"concrete_test":"Build a section-assignment audit table: for every method in Tables 1–3, record the criterion from the method's own paper (is task identity known at test time? do new classes arrive? does the input distribution drift?) and compare it with the section where the survey places it. The decisive check is for Cross-FCL [41], which is reviewed in both Sec. 2.3 and Sec. 4.2: determine from the original paper whether task identity is provided at test time. If it is not, the FTCL placement is wrong; if it is, the FCCL placement is wrong; if the method is genuinely neither, the taxonomy lacks a rule for mixed settings. Repeating this check for CFeD [15] across Secs. 2.5 and 3.2 will settle whether the overlap is an editorial slip or a systematic ambiguity in the proposed taxonomy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of the survey is that FCL methods can be organized into three task-characteristic scenarios: federated class, domain, and task continual learning. For this organization to be a usable navigation map, the categories must be at least well-defined enough that a method is not assigned to two mutually exclusive scenarios. That condition fails within the manuscript itself. Cross-FCL [41] is reviewed under FCCL in Section 2.3 (parameter decomposition for class-incremental learning) and again under FTCL in Section 4.2 (architecture-based methods, where the task identity is explicitly provided). The defining line drawn in the paper is whether task identity is provided during testing: FCCL states task identity is not provided, while FTCL states it is. A single method cannot simultaneously satisfy both conditions, so either one of the two reviews is misassigned or the taxonomy's decision rule is not actually being applied. Similarly, CFeD [15] is covered under FCCL in Section 2.5 and under FDCL in Section 3.2; while domain drift and new classes can co-occur in practice, the survey presents these as distinct scenarios without giving a rule for mixed cases. The consequence is concrete: a newcomer who follows the taxonomy to locate all work on, say, flooding with new classes would miss or double-count the overlapping entries, and the 'first comprehensive survey' claim rests substantially on this taxonomy being a genuine organizational contribution. The summaries themselves may be accurate, but the load-bearing organizational structure needs an explicit exclusion condition or a stated allowance for overlap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript surveys federated continual learning (FCL) in the context of Edge-AI. It proposes a three-way taxonomy based on task characteristics: federated class continual learning (FCCL), federated domain continual learning (FDCL), and federated task continual learning (FTCL). For each category, it reviews representative methods (generative replay, parameter regularization, parameter decomposition, prompting, knowledge distillation for FCCL; data supplementation, knowledge learning, model enhancement, weight aggregation for FDCL; regularization, architecture, replay, meta-learning, and unsupervised methods for FTCL), summarizes them in tables, and then reviews applications (transportation, medical, IoT, UAVs, energy, digital twins, auditing, robotics) and future directions (benchmarks, explainability, algorithm-hardware co-design, foundation models). The paper claims to be the first comprehensive survey of FCL for Edge-AI.","tokens_in":33126,"tokens_out":4525,"duration_ms":41711,"significance":"If the central claims hold, the survey would be a useful entry point for researchers entering FCL, particularly because it brings together a large and recent body of work and organizes it by the type of continual-learning scenario, with summary tables that allow quick comparison of methods. The collection of application areas and the discussion of open challenges, especially algorithm-hardware co-design and foundation-model integration, are valuable and generally accurate in their individual descriptions of cited works. However, the value of the survey depends critically on the taxonomy being a reliable navigation map. The manuscript itself contains duplicate assignments of methods to different taxonomy categories, and the formal definitions contain notation errors. These issues undermine the claimed organizational contribution and the 'first comprehensive survey' claim, so the manuscript needs substantive revision before the central claims can be accepted.","major_comments":[{"comment":"Cross-FCL [41] is reviewed as an FCCL parameter-decomposition method in Section 2.3 and again as an FTCL architecture-based method in Section 4.2. This is not merely a presentation redundancy: the survey defines FCCL by the absence of task identity during testing (Section 2) and FTCL by the explicit provision of task identity during learning and testing (Section 4). One method cannot satisfy both conditions, so either one review is misassigned or the taxonomy's decision rule is not actually being applied. The authors should state a clear assignment rule and remove or justify each duplicate placement.","section":"§2.3 and §4.2"},{"comment":"CFeD [15] is listed under FCCL knowledge distillation (Table 1, Section 2.5) and under FDCL domain knowledge learning (Table 2, Section 3.2). The survey presents FCCL and FDCL as distinct scenarios but never gives a rule for methods that address both new classes and domain drift, nor does it discuss mixed scenarios. This makes the taxonomy non-exhaustive for real methods and potentially misleading for a newcomer who uses the taxonomy to locate all relevant work. The authors should either justify a mixed-category treatment or choose a single primary assignment for each method.","section":"§2.5, Table 1, §3.2, Table 2"},{"comment":"The problem formalization for FDCL contains notation errors that make the definition unusable as written. The displayed definition D^t_k = {(x^t_i, y^t_i)}^{|K|}_{i=1} uses |K| as the upper bound of the sample index i, although i should index samples in client k and the number of samples is client-specific; the text later uses D^t_g = {D^t_1, ..., D^t_K}, which conflates the number of clients (K) with the sample-count bound. Please correct the indexing and clearly distinguish the number of clients from the number of samples per client.","section":"§3 (FDCL formalization)"},{"comment":"The paper claims in Section 1.2 and Section 7 to be 'the first comprehensive survey of federated continual learning for Edge-AI,' but Section 1.1 states that 'Yang et al. [24] conducted a survey of FCL.' As written, the claims are internally inconsistent. The authors should either temper the novelty claim or explicitly delineate the Edge-AI-specific scope, inclusion criteria, and how the present survey differs from [24] in a way that justifies the word 'first.'","section":"§1.2, §1.1, §7"},{"comment":"A survey whose central contribution is comprehensiveness should state its literature search protocol: databases searched, time window, keywords, inclusion/exclusion criteria, and how representative methods were selected. The manuscript does not describe any such protocol, so the 'comprehensive' claim cannot be independently verified or updated. Adding a short methodology subsection would substantially strengthen the paper.","section":"§1 (no methodology section)"}],"minor_comments":[{"comment":"The sentence 'Paul et al. [87] extend FedWeIT [41]' cites the wrong reference: FedWeIT is Yoon et al. [40], while [41] is Cross-FCL. Please correct the citation.","section":"§4.5"},{"comment":"The introduction to Section 6 says 'we highlight and discuss three future directions,' but the section contains four subsections (6.1 FCL Benchmark, 6.2 Explainable FCL, 6.3 Algorithm-Hardware Co-design, 6.4 FCL with Foundation Models). Please adjust the count or the section structure.","section":"§6 (intro)"},{"comment":"The heading 'UA Vs' should read 'UAVs.'","section":"§5.4"},{"comment":"The phrase 'regulation-based approaches' should be 'regularization-based approaches' to match the terminology used elsewhere in the paper.","section":"§4.6"},{"comment":"The phrase 'Spare matrix multiplication' should be 'Sparse matrix multiplication.'","section":"§6.3"},{"comment":"The dataset name 'SHVN' appears to be a typo for 'SVHN.'","section":"§6.1"},{"comment":"In the conclusion, 'applications empowered by federated continual learning In addition' is missing a period before 'In addition.'","section":"§7"}],"recommendation":"major_revision","confidential_remarks":"The survey has useful content and the individual method summaries appear broadly faithful to the cited works, but the central organizational claim needs repair: the taxonomy must be made internally consistent (no duplicate assignments of the same method to mutually exclusive categories), the formalizations should be corrected, and the 'first comprehensive survey' claim must be reconciled with the acknowledged prior survey [24]. These are fixable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The survey gives the FCL community something real: a readable organization of the literature by three task scenarios—class, domain, and task continual learning—plus an applications chapter that actually covers nine domains. I checked a few of the method summaries against the originals I know (GLFC, TARGET, FedWeIT, CFeD) and they are faithful enough. For a newcomer, this is the fastest way to see what exists and where the open problems are. That is the paper's value, and it is genuine.\n\nThe soft spots are real but not fatal. First, the abstract and conclusion claim 'first comprehensive survey' while Section 1.1 cites Yang et al. [24], a prior FCL survey. The authors differentiate themselves by the task-characteristic taxonomy and the Edge-AI angle, which is legitimate, but the priority claim is too strong as written. 'First organized by these task characteristics' would be accurate and still worth saying.\n\nSecond, the three-category taxonomy is not a partition, and the paper never says whether overlap is allowed. Cross-FCL appears in both FCCL (Sec. 2.3) and FTCL (Sec. 4.2); CFeD appears in both FCCL (Sec. 2.5) and FDCL (Sec. 3.2). The stress-test note is right: a newcomer using the taxonomy to find all work on a scenario will double-count or miss entries. This is fixable by stating the categories are lenses, not exclusive buckets, and by giving a rule for mixed cases.\n\nThird, minor mechanical issues: notation errors in the formalizations (e.g., the FDCL dataset indexed by |K|) and no stated search protocol for the survey coverage. For a survey of roughly 170 papers, a Methods or scope paragraph would help reproducibility, though for a venue like ACM CSUR this is more of a reviewer request than a deal-breaker.\n\nMy overall read: the taxonomy is useful even though it is not a partition, the summaries are reliable, and the flaws are addressable in revision. The paper deserves a serious referee. I would send it out with a request to fix the 'first' claim, state the overlap rule, and clean the notation.","headline":"A useful newcomer's map of federated continual learning, organized by three task scenarios, though the taxonomy overlaps and the 'first' claim needs rewording.","tokens_in":33667,"tokens_out":2553,"would_cite":true,"duration_ms":24650,"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 claims that federated continual learning for Edge-AI can be organized by three task characteristics—new classes, drifting domains, and known task identities—and that each demands its own family of anti-forgetting methods.","keywords":["Federated Continual Learning","Edge-AI","Edge Computing","Artificial Intelligence","Lifelong Learning","Incremental Learning","Federated Learning","Catastrophic Forgetting"],"falsifier":"Run a coverage test over the surveyed corpus: assign each method to exactly one of federated class, domain, or task continual learning; if methods such as CFeD must be placed in more than one category, or if newly published FCL methods fall outside all three, then the taxonomy's partition claim fails.","tokens_in":32642,"feed_emoji":"🧩","tokens_out":8846,"duration_ms":78654,"temperature":0.7,"pith_summary":"The paper sets out to be the first comprehensive survey of federated continual learning for Edge-AI, and its organising claim is that the field splits into three task scenarios: federated class continual learning, where new classes arrive; federated domain continual learning, where data distributions drift while classes stay fixed; and federated task continual learning, where task identity is known. For each scenario the survey reviews representative methods, their formal problem settings, and their limitations, grouping them into strategy families such as generative replay, parameter regularization, parameter decomposition, prompting, and knowledge distillation for the class scenario. The takeaway for a reader is that FCL is not a single problem but a small set of related problems, and that the choice of anti-forgetting strategy should follow from which of the three task characteristics is present. The survey also catalogs applications in transportation, medicine, IoT, UAVs, energy, digital twins, financial audit, and robotics, and identifies benchmarks, explainability, algorithm-hardware co-design, and foundation models as open directions.","feed_headline":"Three task types organize federated continual learning","feed_subtitle":"New classes, drifting domains, and known task IDs each call for different forgetting-prevention strategies.","key_machinery":"The central object is the three-way task-characteristic taxonomy: federated class continual learning (task identity withheld, new classes appear), federated domain continual learning (class set fixed, distributions drift), and federated task continual learning (task identity supplied). This taxonomy does the organizing work of the survey: it determines which forgetting problems are central in a given scenario and which solution families are even available, since methods that rely on task-specific components are natural only when task identity is known. Secondary organizing devices include the tri-level division into data-centric, model-centric, and algorithmic approaches for the class scenario, and the four domain-focused strategy groups of data supplementation, knowledge learning, model enhancement, and weight aggregation.","core_discovery":"The survey's central claim is a map: federated continual learning for Edge-AI is best understood through three task characteristics, distinguished by what changes over time and whether task identity is available at test time. In federated class continual learning, clients encounter new classes and task identity is withheld, so the failure modes are intra-task forgetting, where the global model loses knowledge contributed by a client that did not participate in a round, and inter-task forgetting, where new tasks degrade performance on old ones. In federated domain continual learning, the class set stays fixed while local and global data distributions drift, and the model must generalize across client-specific and unknown domains while adapting to known domain drift. In federated task continual learning, task identity is provided, which makes task-specific components and task-aware methods available. The paper then subdivides each scenario into concrete strategy families, reviews representative methods within the families, and argues that these families cover the current state of the art in FCL for Edge-AI.","pith_inferences":["If the taxonomy is read as a design guide rather than a strict partition, it suggests a testable rule: a method built for one scenario should be re-evaluated when task-identity availability or the type of drift changes, for instance when a prompt-based class-continual method is moved to a domain-drift setting.","The survey's own placement of CFeD under both class and domain continual learning hints that scenario boundaries are not sharp; a matrix of 'what changes by how the method prevents forgetting' might represent the literature more faithfully than three disjoint buckets.","The benchmark discussion's emphasis on blurry task boundaries implies that real deployments mix class and domain drift, so extending the taxonomy to explicitly cover mixed-drift tasks would be a natural next step.","FCL with foundation models could make the class/domain distinction less central: when a frozen pretrained backbone is combined with prompts or adapters, what is forgotten is task-specific parameterization, not the shared representation."],"forward_implications":["Federated class continual learning can be attacked at the data, model, or algorithm level, and method choice should follow from whether intra-task or inter-task forgetting dominates.","Because task identity is absent in class and domain scenarios, methods that need task-specific parameter components are less natural there; task identity supplied in advance is what makes architecture-based and task-aware replay methods viable.","Domain-drift solutions sit on a privacy-generalization spectrum: data supplementation risks leakage, knowledge distillation adds computational cost, model enhancement may not transfer, and weight aggregation mainly improves known domains.","The nine application areas share the same three task characteristics, so the taxonomy gives practitioners a common language for choosing and comparing FCL approaches across domains.","The open directions identified by the survey—benchmarks, explainability, algorithm-hardware co-design, and foundation models—are where the taxonomy and its coverage assumptions will be tested next."],"supporting_citations":[{"why":"Supplies the federated learning setup of server-client aggregation with local private data that FCL extends to continual learning.","marker":"[6]"},{"why":"Defines continual learning and catastrophic forgetting, the core problem that all three FCL scenarios address.","marker":"[13]"},{"why":"Provides the three-types-of-incremental-learning framing that underlies the class/domain/task task-characteristic taxonomy.","marker":"[21]"},{"why":"Supplies the replay-based, regularization-based, and parameter-isolation method families that the survey adapts throughout.","marker":"[22]"},{"why":"The earlier FCL survey from a knowledge-fusion perspective that this paper positions itself against.","marker":"[24]"},{"why":"Supplies the tri-level data-centric, model-centric, and algorithmic division used to structure the FCCL chapter.","marker":"[25]"},{"why":"Anchors the parameter-decomposition family in both the class and task continual learning literature.","marker":"[40]"},{"why":"A representative method reviewed under both FCCL and FDCL, marking a boundary case for the taxonomy.","marker":"[15]"}],"fun_headline_variants":["Class, domain, task: the FCL taxonomy for edge AI","Survey maps forgetting prevention in federated edge learning","Edge AI's continual learning gets a three-part breakdown","Three task types shape federated continual learning","How federated learning handles new classes, domains, tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The taxonomy works only if every federated continual learning method can be assigned to exactly one of the three task scenarios, and the survey does not justify that the categories are exhaustive or non-overlapping.","fun_headline_variants_meta":{"raw":{"variants":["Class, domain, task: the FCL taxonomy for edge AI","Survey maps forgetting prevention in federated edge learning","Edge AI's continual learning gets a three-part breakdown","Three task types shape federated continual learning","How federated learning handles new classes, domains, tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000597,"raw_usage":{"total_tokens":2807,"prompt_tokens":974,"completion_tokens":1833,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":1756}},"tokens_in":590,"tokens_out":1833,"duration_ms":13692,"temperature":1.0,"reasoning_tokens":1756,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:56:05.986976+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a coverage test over the surveyed corpus: assign each method to exactly one of federated class, domain, or task continual learning; if methods such as CFeD must be placed in more than one category, or if newly published FCL methods fall outside all three, then the taxonomy's partition claim fails.","supporting_citations":[{"cited_title":"Federated Continual Learning via Knowledge Fusion: A Survey","cited_arxiv_id":"2312.16475","evidence_quote":"The earlier FCL survey from a knowledge-fusion perspective that this paper positions itself against."}],"review_version":1}