{"id":"35049e75-e698-403e-9388-538c7bae683a","arxiv_id":"2502.10419","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A hybrid PSO-ACO framework is proposed for selecting edge devices and routing model updates in federated multimodal LLM training, claiming 92% accuracy and 30% lower communication cost.","lead":"This paper combines two known optimization algorithms, particle swarm optimization and ant colony optimization, to choose which edge devices train a multimodal AI model and how they send updates to a central cloud. It claims a 92% accuracy and a 30% communication cost cut, but the supporting simulation evidence is thin and internally inconsistent.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No standard FL baseline is evaluated: Section 6.2.2 compares only rule-based, CNN-LSTM, and DRL models on a UVS decision task, so the central claim of improvement over 'traditional FL methods' is unverified.","rationale":"The strongest claim is comparative: PSO-ACO beats traditional FL. That comparison is never made. Section 6.2.2's baselines are non-FL (rule-based, CNN-LSTM, DRL), and Figure 5's 92% accuracy is on a UVS decision task. The paper does describe communication rounds and client participation (Figs. 7, 9) but omits the FL protocol details needed to reproduce the setting (local epochs, aggregation, non-IID generation, seeds). Without a FedAvg-style control, any accuracy difference could come from the MLLM architecture, dataset, or simulation environment rather than the PSO/ACO mechanisms. This is more fundamental than the Eq. (16) fitness-proxy concern: even a perfect proxy would not establish the claim unless the FL experiment is present. The internal inconsistency in communication savings (30% abstract vs. 100→40 MB in Fig. 7 vs. 50 MB in Table 1) reinforces that the numbers are not auditable. I therefore keep the reader's REJECT verdict; the fix is a reproducible simulation against standard FL baselines.","tokens_in":20343,"tokens_out":4517,"duration_ms":46134,"concrete_test":"Implement the proposed PSO-ACO pipeline and a plain-FedAvg baseline with identical MLLM fine-tuning, PixSet+UrbanSound8K data, 50–200 devices, and the same non-IID split; run at least 5 seeds and report mean±std for global MLLM accuracy, per-round communication cost, and client participation. If PSO-ACO does not beat FedAvg on these metrics, the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires evidence that the PSO-ACO selection and ACO routing improve a real federated MLLM training pipeline relative to standard FL. Section 6.2.2 defines three baselines—Conventional Rule-Based, CNN-LSTM, and Deep RL—none of which is a federated learning method; no FedAvg or other FL baseline appears anywhere in Sections 6.3–6.4. Accuracy is reported on a UVS decision-making task (Fig. 5), not on the global MLLM objective or on downstream federated task performance. The convergence and communication plots (Figs. 6–7) describe rounds and clients but give no protocol for local fine-tuning, aggregation, non-IID partitioning, or client sampling, so the 92% accuracy cannot be attributed to the proposed FL-specific optimizations. The reported savings are also internally inconsistent: the abstract says 30% communication reduction, Fig. 7 implies 100→40 MB (60%), and Table 1 lists the full model at 50 MB. Because the comparison class required by the claim is absent, the experiments cannot support the headline result regardless of the fitness-proxy assumptions in Eq. (16).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid PSO-ACO framework for federated fine-tuning of multimodal large language models in edge-cloud settings: PSO selects edge devices based on a weighted fitness of energy, data relevance, and diversity, and ACO routes model updates to the cloud. The authors claim 92% accuracy, 30% communication-cost reduction, and improved client participation relative to traditional FL. They evaluate on a simulated UVS use case with Leddar PixSet and UrbanSound8K, comparing against rule-based, CNN-LSTM, and DRL baselines and an ablation.","tokens_in":20604,"tokens_out":5456,"duration_ms":46414,"significance":"If the claims were supported, the framework would be a useful contribution to resource-aware federated MLLM deployment, and the choice of PSO/ACO is well motivated for device selection and routing. However, the central claim is not supported by the experimental design: no federated learning baseline or FL training protocol is reported, the accuracy is measured on a UVS decision task rather than the global MLLM objective, and the communication-cost numbers are mutually inconsistent. The methods sections are mostly standard metaheuristic descriptions, and no code or reproducible parameter settings are provided, so the paper currently does not meet the evidentiary bar for its headline claims.","major_comments":[{"comment":"Section 6.2.2 defines the baselines as Conventional Rule-Based, CNN-LSTM, and Deep RL, none of which is a federated learning method, and Figs. 5–9 report accuracy and communication metrics on the UVS decision-making task. The abstract's claim that the framework 'significantly improves ... accuracy of 92%, reducing communication cost by 30%, and enhancing client participation compared to traditional FL methods' is therefore untested: there is no FedAvg or equivalent FL baseline, no global-model accuracy for the federated MLLM objective, and no protocol for local fine-tuning, aggregation, non-IID partitioning, or client sampling. The reported 92% cannot be attributed to the PSO/ACO selection and routing mechanism.","section":"§6.2.2, §6.3, abstract"},{"comment":"The communication-cost results are internally inconsistent. The abstract reports a 30% reduction; the text for Fig. 7 says the PSO-ACO framework reduces cost from 100 MB to about 40 MB per client (a 60% reduction); and Table 1 lists the full-model communication cost as 50 MB. These three numbers cannot all describe the same quantity, and without an explicit communication-cost model beyond Eq. (3) the claimed savings are not reproducible.","section":"§6.3, Fig. 7; Table 1"},{"comment":"The PSO fitness function is a weighted sum of energy, relevance, and diversity with weights α, β, γ, but the paper never reports the values of these weights or validates that minimizing this proxy improves the global MLLM. The 92% accuracy is reported on a UVS decision task, not on the federated MLLM objective, so the correlation between the selection proxy and model accuracy is unestablished. If the proxy is mis-specified, PSO selects the wrong devices and the central accuracy claim does not follow.","section":"§5.2, Eq. (16)"},{"comment":"The ablation study presents single values with no error bars, number of runs, or statistical tests, and it omits all PSO/ACO hyperparameters (inertia ω, acceleration coefficients c1,c2, evaporation rate ρ, and exponents α,β in Eq. (19)). The row 'Without PSO' has the same communication cost (50 MB) as the full model, so the reported differences cannot be separated from noise or parameter choice, and the ablation does not isolate the FL-specific contribution of either swarm component.","section":"§6.4, Table 1"}],"minor_comments":[{"comment":"The quantity Mi/Bi · di mixes units (model size divided by bandwidth, then multiplied by distance); the intended model should either define di as a normalized coefficient or state units explicitly.","section":"Eq. (3)"},{"comment":"The heuristic ηij for bandwidth is defined as 1/Bij while the text says higher-bandwidth paths should be preferred; as written, lower-bandwidth links receive larger heuristic values, which contradicts the stated goal. If higher bandwidth is better, ηij should be proportional to Bij (or the inverse should be used for latency only).","section":"§5.3"},{"comment":"The sentence 'Ecomm and Etrain represents the energy and time required to transmit model updates' confuses energy and time; please clarify the roles of Etrain, Ecomm, and Etime.","section":"§2.1, Eq. (1)"},{"comment":"Figures 5–9 are referenced only through their captions and accompanying text; the actual plot images are missing from the manuscript, and none of the reported metrics include error bars or confidence intervals.","section":"§6.3, Figs. 5–9"},{"comment":"The diversity term γ Σ Di is minimized together with energy and relevance; if high diversity is desired, the diversity term should enter negatively or as an inverse so that the optimization direction matches the stated non-IID objective.","section":"§5.2, Eq. (16)"},{"comment":"Section 6.3 refers to 'federated learning approaches' although Section 6.2.2 does not include any FL baseline; please align the terminology with the actual comparison class.","section":"§6.3 vs §6.2.2"}],"recommendation":"reject","confidential_remarks":"The manuscript is not suitable for publication in its current form. The algorithmic building blocks are standard and the use case is relevant, but the experiments do not test the claimed FL improvement, and the reported headline numbers are internally inconsistent. A revision would require a new experimental design with a real FedAvg baseline, a defined federated protocol, and reproducible hyperparameters; as such, the issues are beyond a local revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this paper proposes a sensible hybrid — PSO for selecting edge devices and ACO for routing model updates in federated MLLM training over edge-cloud — but the experiments do not demonstrate the claimed gains. The central claim of improving over 'traditional FL methods' is unverified because no FL baseline appears anywhere. The evaluation compares against rule-based, CNN-LSTM, and DRL on a UVS decision task, none of which is a FL method.\n\nWhat is actually new: the specific combination of PSO-based client selection and ACO-based path optimization in a federated MLLM edge-cloud setting is not in the cited literature. The problem formulation (energy, communication, non-IID) is clear, and the algorithmic descriptions are standard. That is a real but modest contribution.\n\nSoft spots are serious. No FedAvg or other FL baseline. Accuracy is measured on the UVS decision task, not on the global MLLM objective. The figures (5–9) are missing from the text, so results can't be checked. Parameters for PSO, ACO, and the fitness weights alpha, beta, gamma are not reported, so the work is not reproducible. Reported numbers are inconsistent: abstract says 30% communication reduction, text says 100→40 MB (60%), Table 1 lists 50 MB. There are no error bars. The ACO heuristic in Section 5.3 has a sign error: ηij = 1/Bij makes low-bandwidth paths more attractive, contrary to the stated intent.\n\nThe ablation study is a nice idea but it is a table of single numbers without variance, and the removal of PSO/ACO is not tied to the FL training dynamics. The fitness proxy in Eq. (16) is unvalidated; nothing shows that the weighted combination correlates with global model accuracy.\n\nBottom line: the engineering idea is reasonable and worth exploring, but the paper in this form should not be accepted. It needs a complete rewrite of the experimental section: a real FedAvg baseline, non-IID partition protocol, hyperparameters, error bars, and consistent reporting. A serious editor might send it to review if the authors had a credible experiment, but as written the evidence is not there. I would desk reject now, and tell the authors to come back with a proper evaluation.","headline":"Reasonable hybrid idea, but the missing FL baseline and unreported experimental protocol sink the current claims.","tokens_in":21172,"tokens_out":3123,"would_cite":false,"duration_ms":29020,"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":"The paper tries to establish that a hybrid of PSO and ACO can lift federated MLLM deployment to 92% accuracy while cutting communication cost by 30%.","keywords":["Federated Learning","Multimodal Large Language Models","Swarm Intelligence","Particle Swarm Optimization","Ant Colony Optimization","Edge-Cloud Computing","Resource Optimization","Non-IID Data"],"falsifier":"Run the same federated MLLM fine-tuning task twice, once with PSO-selected devices and once with the same number of randomly selected devices matched for energy and communication budget; if random selection matches or beats PSO on global-model accuracy over 20 rounds, the fitness proxy is not doing the work.","tokens_in":20139,"feed_emoji":"🐝","tokens_out":9687,"duration_ms":82163,"temperature":0.7,"pith_summary":"The paper tries to establish that a hybrid of two swarm-intelligence algorithms can solve the two bottlenecks that keep multimodal large language models (MLLMs) out of federated edge-cloud training: deciding which resource-limited edge devices should fine-tune the model, and finding low-cost transmission paths for the updates. Particle Swarm Optimization (PSO) picks the device subset from energy, data relevance, and data diversity; Ant Colony Optimization (ACO) routes the model updates back to the cloud. If correct, federated MLLM fine-tuning could run on private, heterogeneous edge data without exhausting device batteries or swamping the network. The authors report 92% accuracy on an unmanned-vehicle decision task, about a 30% communication-cost reduction, higher client participation, and faster convergence than rule-based, CNN-LSTM, and deep-reinforcement-learning baselines.","feed_headline":"Hybrid swarm method lifts federated MLLM accuracy to 92%","feed_subtitle":"Particle swarm picks the edge devices; ant colony routes updates, cutting communication cost by 30%.","key_machinery":"The framework's load-bearing machinery is the pair of optimization loops that decide who trains and how updates travel. Particle Swarm Optimization (PSO) is a swarm search in which candidate solutions (particles) move toward their own best-known position and the swarm's best position; Ant Colony Optimization (ACO) is a path search in which ants reinforce good routes with pheromone. PSO scores each candidate device subset $S$ by a fitness function $Fitness(S)=\\alpha\\sum_{i\\in S} E_i+\\beta\\sum_{i\\in S}(1-R_i)+\\gamma\\sum_{i\\in S} D_i$ (Eq. 16), where $E_i$ is energy consumption, $R_i$ is data relevance, and $D_i$ is data diversity, with weights $\\alpha,\\beta,\\gamma$; particles update velocities and positions by the standard PSO equations (Eqs. 17-18). ACO chooses communication links with probability $P_{ij}=\\tau_{ij}^{\\alpha}\\eta_{ij}^{\\beta}/\\sum_{k\\in neighbors} \\tau_{ik}^{\\alpha}\\eta_{ik}^{\\beta}$ (Eq. 19), where $\\tau_{ij}$ is pheromone and $\\eta_{ij}$ is a heuristic favoring short or high-bandwidth links, and then evaporates and deposits pheromone (Eq. 20). The selected devices' updates are combined in the cloud as $w_{global}=\\sum_{i\\in S} \\frac{n_i}{n} w_i$ (Eq. 21). The fitness proxy is what connects device selection to model quality; the pheromone dynamics are what make routing adaptive.","core_discovery":"The paper's central claim is that the PSO-ACO framework improves federated MLLM deployment in edge-cloud systems: the cloud trains the global model, PSO selects a subset of edge devices that have enough resources and relevant, diverse local data, and ACO optimizes the communication paths over which the fine-tuned updates travel back for aggregation. The authors argue that this selection counteracts non-IID (non-independent and identically distributed) data by sampling a representative, diverse device subset each round, while the routing optimization reduces bandwidth use and latency, which are severe when MLLM updates are large. On an unmanned-vehicle-system use case combining the Leddar PixSet and UrbanSound8K datasets, the framework achieves 92% accuracy, an F1 score of 91%, median 88% accuracy under non-IID data, communication cost falling from 100 MB to around 40 MB per client over 20 rounds, and higher client participation than the baselines. The ablation study attributes the gains to all three components: removing PSO drops accuracy to 85%, removing ACO raises communication cost by roughly 30%, and removing edge-cloud integration drops accuracy to 83%.","pith_inferences":["A direct testable extension is to replace the hand-set weights $\\alpha,\\beta,\\gamma$ with an online learned weighting or a per-device marginal-contribution estimate; the paper gives no evidence the fixed proxy is optimal.","The same hybrid decomposition (metaheuristic selection plus metaheuristic routing) should transfer to other large-model federated settings, such as parameter-efficient fine-tuning adapters on phones, where the discrete-continuous split of decisions is the same.","Because the 92% accuracy is measured on the UVS decision task rather than on the federated aggregation objective, the framework's benefit to the global MLLM itself would be established more directly by comparing global-model loss under PSO selection against random selection at equal communication budget."],"forward_implications":["If the claim holds, federated MLLM training no longer requires every device to participate: a swarm-selected subset can sustain accuracy while conserving battery and network resources.","The roughly 30% communication-cost reduction would directly address the biggest practical obstacle to MLLM federated learning, since model updates are orders of magnitude larger than for small classifiers.","The reported stable accuracy under non-IID data (median around 88%) suggests device selection on relevance and diversity can work without data-sharing or reweighting strategies.","The ablation results imply the three components are complementary: the accuracy gain comes mainly from PSO and edge-cloud integration, while the communication gain comes mainly from ACO."],"supporting_citations":[{"why":"It supplies a precedent for using particle-swarm-based client selection in federated learning.","marker":"Supriya & Gadekallu (2023)"},{"why":"It supports applying swarm intelligence to energy- and resource-constrained IoT contexts.","marker":"Xing et al. (2023)"},{"why":"It motivates PSO for multi-criteria device selection in UAV swarm networks.","marker":"Cao et al. (2024)"},{"why":"It motivates ACO for communication path optimization in dynamic UAV networks.","marker":"Jiang et al. (2024b)"},{"why":"It frames swarm intelligence as a solution to IoT resource and communication challenges.","marker":"Abualigah et al. (2023)"},{"why":"It provides a survey establishing swarm intelligence algorithms for IoT communication optimization.","marker":"Sun et al. (2020)"},{"why":"It supplies the DQN deep-reinforcement-learning baseline the framework is compared against.","marker":"Mnih et al. (2013)"},{"why":"It supplies the PPO deep-reinforcement-learning baseline the framework is compared against.","marker":"Schulman et al. (2017)"}],"fun_headline_variants":["PSO+ACO swarm boosts federated MLLM accuracy to 92%","Hybrid swarm cuts federated MLLM comms by 30%","Edge-cloud swarm method hits 92% accuracy in federated MLLM","Ant colony routing trims federated MLLM bandwidth by 30%","Particle swarm selects edge devices for smart MLLM training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the weighted sum of energy use, data relevance, and data diversity in the PSO fitness function predicts which edge devices actually improve the global multimodal model's accuracy.","fun_headline_variants_meta":{"raw":{"variants":["PSO+ACO swarm boosts federated MLLM accuracy to 92%","Hybrid swarm cuts federated MLLM comms by 30%","Edge-cloud swarm method hits 92% accuracy in federated MLLM","Ant colony routing trims federated MLLM bandwidth by 30%","Particle swarm selects edge devices for smart MLLM training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000697,"raw_usage":{"total_tokens":3182,"prompt_tokens":1010,"completion_tokens":2172,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":626,"completion_tokens_details":{"reasoning_tokens":2072}},"tokens_in":626,"tokens_out":2172,"duration_ms":12983,"temperature":1.0,"reasoning_tokens":2072,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T13:52:01.600655+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same federated MLLM fine-tuning task twice, once with PSO-selected devices and once with the same number of randomly selected devices matched for energy and communication budget; if random selection matches or beats PSO on global-model accuracy over 20 rounds, the fitness proxy is not doing the work.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies a precedent for using particle-swarm-based client selection in federated learning."},{"cited_title":"M., Alharbi, A","cited_arxiv_id":null,"evidence_quote":"It supports applying swarm intelligence to energy- and resource-constrained IoT contexts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It frames swarm intelligence as a solution to IoT resource and communication challenges."}],"review_version":1}