{"id":"6c16553a-42de-4146-b6b3-1539a515974f","arxiv_id":"2505.03139","paper_version":1,"verdict":"REJECT","confidence":"LOW","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"The paper proposes a collaborative edge deployment framework for large AI models, but it is a roadmap with only one narrow case study and no validated core result.","lead":"This paper outlines a framework for deploying large AI models across edge devices for IoT, using federated fine-tuning and microservice-based inference. It includes one small case study on chain-of-thought reasoning, but the framework's central claims are not yet validated.","discovery_kind":"unclear","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Case study filters to 'correctly solving reasoning problems', so reported memory/latency gains never test whether microservice decomposition preserves output quality; central edge-LAM claim is unsupported.","rationale":"The reader identified the modular-decomposition equivalence as the weakest assumption; I agree. The paper's only quantitative evidence is Section III-C, and its correctness filter makes the efficiency comparison conditional on the very property under test. No derivation, code, or full-system validation supports the training or inference claims, so the central claim cannot be accepted as stated. A full-set accuracy and end-to-end latency comparison would settle whether the framework delivers edge LAM intelligence or merely a faster degraded approximation.","tokens_in":9523,"tokens_out":3318,"duration_ms":34003,"concrete_test":"Run Section III-C again on the full GSM8K test partition (or a documented random subset) with no correctness filter, comparing monolithic Qwen2.5-7B-Instruct against the microservice/CoT deployment. Measure exact-match accuracy and end-to-end wall-clock latency including inter-device communication, not just per-device computation latency. If accuracy is preserved within statistical error and total latency is lower, the concern is resolved; if accuracy drops or communication-dominated latency rises, the reported 70.8%/59.6% gains do not transfer to real edge-LAM service.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim — virtualizing LAM functional modules as edge microservices improves resource utilization and reduces latency — requires that decomposed models preserve output quality. The only evaluation (Section III-C) cannot establish this: it reports memory and computation latency only 'when correctly solving reasoning problems', discarding all incorrect trajectories. This selection removes exactly the cases where decomposition or cross-device routing may have broken reasoning, so the 70.8% memory and 59.6% latency reductions are computed on a survivor subset. Section V itself lists 'Generation Quality' as an open challenge, confirming that output-quality preservation is not settled. The paper also never specifies how CoT steps are extracted into microservices, how the path-graph constraint is enforced during decoding, or whether inter-device communication latency is included in the reported 'computation latency'. Without an accuracy comparison on the full task distribution, the efficiency numbers do not show that edge LAMs, rather than a degraded approximation, are delivered.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a collaborative deployment framework for large AI models at the network edge ('edge LAMs'), with two components: a federated fine-tuning and unlearning framework for heterogeneous wireless edge networks, and a microservice-based inference framework that virtualizes functional modules of Mixture-of-Experts and Chain-of-Thought architectures. It also sketches IoT applications in intelligent transportation and industrial fault diagnosis and lists challenges and future directions. The paper claims that these techniques reduce communication and computation overhead during fine-tuning and improve resource utilization and inference latency. The only quantitative evaluation is a case study in Section III-C using Qwen2.5-7B-Instruct on GSM8K, reporting that a microservice-based CoT architecture reduces memory consumption by 70.8% and computation latency by 59.6% relative to a monolithic baseline.","tokens_in":9701,"tokens_out":4552,"duration_ms":44324,"significance":"The problem addressed is timely and important: deploying large generative models at the network edge could enable low-latency, privacy-preserving IoT intelligence, and the modular-decomposition ideas are well aligned with current trends in edge inference. The manuscript also usefully identifies heterogeneous computation, data modalities, and wireless dynamics as central challenges. However, the significance of the specific claims is not yet established. There is no derivation or full-system evaluation for the claimed communication/computation overhead reductions, the training sections are qualitative, and the sole experiment is filtered to correctly solved problems, so it cannot support the claim that the framework preserves output quality while improving efficiency. The paper would be a useful roadmap if its claims were appropriately scoped, but as written it overstates what has been demonstrated.","major_comments":[{"comment":"The only quantitative evidence for the inference framework is based on trajectories filtered to those 'correctly solving reasoning problems' (Section III-C). This filters out exactly the cases in which microservice decomposition or cross-device routing may have degraded output quality, so the reported reductions of 70.8% in memory and 59.6% in latency are conditional on a survivor subset. The paper must report accuracy on the full GSM8K split and analyze the errors before claiming that the proposed architecture preserves model quality while improving efficiency.","section":"III-C, Fig. 5"},{"comment":"The abstract's claim of reduced communication and computation overhead during fine-tuning is not supported by any derivation or evaluation. Sections II-A and II-B describe the proposed mechanisms in qualitative, future-tense terms ('we will develop', 'we aim to develop') and provide no optimization problem, algorithm, convergence analysis, or simulation for the LoRA/zero-padding/truncation, knowledge-distillation, or federated-unlearning components. These components are load-bearing for the central training claim and need at least a formal problem statement and a proof-of-concept evaluation.","section":"II-A, II-B"},{"comment":"The microservice inference proposals are underspecified for a performance claim. The paper does not state how MoE experts or CoT steps are isolated into microservices, how the path-graph constraint on CoT microservice flows is enforced during decoding, or whether inter-device communication latency is included in the reported 'computation latency' in Fig. 5. Without these details, the claimed latency and resource-utilization gains cannot be verified or reproduced.","section":"III-A, III-B"},{"comment":"Section V lists 'Generation Quality' as an open challenge and proposes only future projection and retrieval-augmented generation modules to handle hallucinations and domain-constraint violations. This is inconsistent with the abstract's implication that the proposed framework already delivers reliable real-time edge LAM inference. The paper should either provide evidence that output quality is preserved by microservice decomposition or explicitly reframe the contribution as a roadmap with output-quality guarantees left to future work.","section":"V, Generation Quality"}],"minor_comments":[{"comment":"The sentence says the paper will focus on 'three critical domains: intelligent transportation systems, and industrial fault diagnosis' but only two domains are enumerated; either add the third domain or revise the sentence.","section":"Section IV opening"},{"comment":"GSM8K is described as containing '253 logical reasoning problems'; the standard GSM8K test split has 1319 problems, so the dataset source and split should be clarified.","section":"III-C"},{"comment":"The figure reports no error bars, number of runs, or statistical significance, and 'computation latency' is not defined; these details should be added, along with code/data availability if possible.","section":"Fig. 5"},{"comment":"Reference [2] contains a typo ('specfic'), and the text repeatedly uses 'We shall propose' and 'we will develop' where a completed contribution is claimed; the tense should be made consistent and the scope of what is actually implemented should be stated precisely.","section":"References and style"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads more like a vision/roadmap paper than a technical contribution with validated mechanisms. The central quantitative claim rests on a single filtered case study, and the training sections are qualitative. If the venue intends to publish position papers, the authors should be asked to remove or substantially soften the quantitative claims and to clearly mark the proposed methods as research agenda; otherwise, substantial additional validation is required. The reliance on the authors' own reference [6] is not problematic in itself, but the text should separate inherited components from new contributions more clearly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a roadmap paper, not a validated research result. It assembles established building blocks—federated LoRA fine-tuning, knowledge distillation for data heterogeneity, federated unlearning, MoE microservices, CoT decomposition, and graph diffusion for deployment—under the 'edge LAM' umbrella. The framing is useful and the taxonomy of challenges is sensible, but the central efficiency claims are not backed by derivation or a full-system evaluation.\n\nWhat the paper does well: Section II gives a coherent picture of heterogeneous federated fine-tuning, with zero-padding/truncation to handle varying LoRA ranks and a shared-module KD mechanism for modality skew. The unlearning section, using orthogonal gradient projection and noise injection over fading channels, is a plausible extension of existing ideas. Section III's move to virtualize MoE experts and CoT steps as microservices is a reasonable research direction, and the cited prior work is relevant. The self-citation to [6] is not a problem: it anchors the FedFT discussion, but the proposed framework stands apart from that earlier paper.\n\nThe soft spots are real and load-bearing. The case study in Section III-C uses Qwen2.5-7B on GSM8K, reports only memory and computation latency, and—critically—only 'when correctly solving reasoning problems.' That selection discards exactly the cases where decomposition or cross-device routing may have broken reasoning, so the 70.8% memory and 59.6% latency reductions are computed on a survivor subset. The paper itself lists 'Generation Quality' as an open challenge, conceding that output-quality preservation is not settled. Details are also missing on how CoT steps are extracted into microservices, how the path-graph constraint is enforced during decoding, and whether inter-device communication latency is included in the reported 'computation latency.' The training and unlearning parts are described in future tense with no experiments. There is no code or data, and no accuracy comparison on the full task distribution.\n\nThese are not cosmetic flaws; they undercut the abstract's claim that the framework 'reduces communication and computation overheads' while preserving the LAM's capabilities. Still, the paper is honest about its exploratory nature—it repeatedly says 'we shall propose'—so it reads as a proposal or survey, not a finished system. There are no fitted parameters and no circular reasoning; the issue is missing evidence, not internal inconsistency.\n\nWho is this for? A reader who wants a quick conceptual map of edge-LAM research directions, especially the microservice-based inference framing. I would not cite it for any quantitative claim or use it as evidence that decomposed edge LAMs preserve quality. If the authors substantially implemented and evaluated even one component—say, MoE microservices on more than one model with full accuracy metrics—or reframed this as an explicit survey, it could be a useful magazine-level contribution. In current form, I would not send it to a strict research journal for full peer review.","headline":"A coherent roadmap for edge LAMs that recombines known techniques, but the only experimental result is a filtered case study that cannot support the headline efficiency claims.","tokens_in":10197,"tokens_out":2298,"would_cite":false,"duration_ms":23390,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that a large AI model can be decomposed into modular pieces and run collaboratively across edge devices, cutting fine-tuning communication, inference memory, and latency enough to make real-time IoT services practical.","keywords":["edge large AI models","federated fine-tuning","low-rank adaptation (LoRA)","mixture of experts","chain-of-thought","microservice orchestration","Internet of Things","wireless edge networks"],"falsifier":"Run the same reasoning benchmark on a monolithic deployment and on the chain-of-thought microservice decomposition, then compare solution accuracy; if the split system solves markedly fewer problems correctly at the reported memory and latency savings, the efficiency claim does not establish deployability.","tokens_in":1413,"feed_emoji":"🧩","tokens_out":1956,"duration_ms":54674,"temperature":0.7,"pith_summary":"The paper tries to establish that large AI models do not have to live in a cloud or on one powerful machine: by decomposing them functionally and coordinating edge devices, they can be trained and served at the network edge. It argues that the two barriers to edge large AI models are heterogeneity in devices and data, plus redundant computation in inference, and it proposes a federated fine-tuning framework and a microservice-based inference framework to address these barriers. If the framework works, IoT applications such as smart traffic control and industrial fault diagnosis could run real-time, privacy-preserving large-model intelligence. The paper is a proposal backed by one small case study; its claims about preserving output quality under decomposition are assumed rather than demonstrated.","feed_headline":"Splitting large AI models across edge devices cuts memory 70%","feed_subtitle":"A microservice-based scheme also trims reasoning latency by 59.6% against monolithic deployment—if output quality holds.","key_machinery":"The enabling objects are LoRA adapters for parameter-efficient federated fine-tuning and microservice virtualization of model internals. LoRA freezes the pre-trained weights and trains only low-rank matrices, so devices exchange small matrices rather than full gradients, and truncation with zero-padding standardizes matrices of different ranks at the edge server. On the inference side, mixture-of-experts experts become microservices scheduled by the edge server via Lyapunov optimization, while chain-of-thought sub-processes become microservices whose placement is learned by a graph diffusion model. The reported efficiency numbers come from the chain-of-thought microservice decomposition.","core_discovery":"The paper's central claim is that a large AI model can be split along functional lines rather than parameter lines: during fine-tuning, trainable low-rank adapters are sized per device, truncated or zero-padded at the server to reconcile heterogeneous capabilities, and aggregated over wireless links; during inference, mixture-of-experts layers and chain-of-thought reasoning steps are virtualized as microservices placed on different edge devices. It claims that this approach cuts communication overhead compared with full-model federated tuning and, in a reasoning experiment with a 7-billion-parameter instruction-tuned model on a grade-school mathematics benchmark, cuts total memory consumption by 70.8% and computation latency by 59.6% relative to a monolithic deployment while correctly solving the tested reasoning problems. Accuracy preservation across the decomposition is assumed, not measured.","pith_inferences":["Editorial extension: the accuracy-preservation assumption could be tested cheaply by running the same reasoning benchmark on monolithic inference and on the chain-of-thought microservice split and comparing exact-match rates; the paper does not report this comparison.","Editorial extension: the decomposition principle resembles model parallelism, so the framework's real novelty is making the split adaptive to wireless channel state rather than the split itself.","Editorial extension: if low-rank adapters of different ranks are aggregated by zero-padding, the server's aggregated matrix may be dominated by the largest rank, so communication savings could degrade as rank heterogeneity grows.","Editorial extension: the graph diffusion deployment model, if it generalizes across edge topologies, would allow edge systems to re-plan microservice placements online as devices join and leave the network."],"forward_implications":["If the framework is correct, fine-tuning a large model across heterogeneous edge devices becomes communication-feasible because only low-rank matrices travel over the wireless network.","If the framework is correct, mixture-of-experts edge inference can avoid redundant expert deployments by activating shared expert microservices on demand.","If the framework is correct, chain-of-thought reasoning latency can be traded against the number of participating devices, with per-device token allocation as the control knob.","If the framework is correct, the same decomposition extends to multimodal IoT tasks whose sensor data is first mapped into token representations.","If the framework is correct, privacy improves because raw data stays on edge devices and only adapters or projected gradients are exchanged."],"supporting_citations":[{"why":"Supplies the federated fine-tuning method for pre-trained foundation models over wireless networks, which the paper adapts to heterogeneous LoRA ranks.","marker":"[6]"},{"why":"Establishes the edge-AI-for-6G vision and constraints that motivate deploying large models at the edge.","marker":"[4]"},{"why":"Provides the mixture-of-experts architecture in mobile edge networks whose layers the paper virtualizes as microservices.","marker":"[11]"},{"why":"Surveys mobile edge intelligence for large language models and motivates the chain-of-thought microservice decomposition.","marker":"[12]"},{"why":"Supplies microservice-oriented service placement in mobile edge computing, which the paper reformulates for chain-of-thought inference.","marker":"[8]"},{"why":"Provides the federated edge learning foundations on which the proposed collaborative fine-tuning framework builds.","marker":"[5]"}],"fun_headline_variants":["Splitting AI models by function cuts edge memory 70.8%","Microservices-based edge AI trims memory 70.8%, latency 59.6%","Functional decomposition of AI models cuts edge memory 70.8%","Edge AI split cuts memory 70.8% but accuracy assumed","Large AI models split by function for edge cut memory 70.8%"],"cache_read_input_tokens":12544,"weakest_assumption_plain":"The load-bearing premise is that splitting a large model into independently executed modules, such as experts or reasoning steps, across separate devices leaves the model's output quality unchanged; the case study assumes this by reporting only results for problems the split system already solves correctly.","fun_headline_variants_meta":{"raw":{"variants":["Splitting AI models by function cuts edge memory 70.8%","Microservices-based edge AI trims memory 70.8%, latency 59.6%","Functional decomposition of AI models cuts edge memory 70.8%","Edge AI split cuts memory 70.8% but accuracy assumed","Large AI models split by function for edge cut memory 70.8%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001143,"raw_usage":{"total_tokens":4728,"prompt_tokens":917,"completion_tokens":3811,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":3710}},"tokens_in":533,"tokens_out":3811,"duration_ms":27705,"temperature":1.0,"reasoning_tokens":3710,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:58:05.197735+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same reasoning benchmark on a monolithic deployment and on the chain-of-thought microservice decomposition, then compare solution accuracy; if the split system solves markedly fewer problems correctly at the reported memory and latency savings, the efficiency claim does not establish deployability.","supporting_citations":[{"cited_title":"Federated fine-tuning for pre-trained foundation models over wireless networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the federated fine-tuning method for pre-trained foundation models over wireless networks, which the paper adapts to heterogeneous LoRA ranks."},{"cited_title":"Edge artificial intelligence for 6G: Vision, enabling technologies, and applications,","cited_arxiv_id":null,"evidence_quote":"Establishes the edge-AI-for-6G vision and constraints that motivate deploying large models at the edge."},{"cited_title":"Toward scalable generative AI via mixture of experts in mobile edge networks,","cited_arxiv_id":null,"evidence_quote":"Provides the mixture-of-experts architecture in mobile edge networks whose layers the paper virtualizes as microservices."},{"cited_title":"Mobile edge intelligence for large language models: A contemporary survey,","cited_arxiv_id":null,"evidence_quote":"Surveys mobile edge intelligence for large language models and motivates the chain-of-thought microservice decomposition."},{"cited_title":"Microservice-oriented service placement for mobile edge computing in sustainable internet of vehicles,","cited_arxiv_id":null,"evidence_quote":"Supplies microservice-oriented service placement in mobile edge computing, which the paper reformulates for chain-of-thought inference."},{"cited_title":"Federated edge learning for 6G: Foundations, methodolo- gies, and applications,","cited_arxiv_id":null,"evidence_quote":"Provides the federated edge learning foundations on which the proposed collaborative fine-tuning framework builds."}],"review_version":1}