{"id":"796ffad7-1ff7-4f22-a7cc-3a7b78335ecf","arxiv_id":"2508.03345","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"AntLLM combines ant colony optimization with an unspecified LLM refinement step to place and migrate AI agents at the edge, claiming significant latency and cost reductions in small-scale experiments.","lead":"This paper proposes AntLLM, a framework that uses ant colony optimization plus an LLM refinement step to place and migrate LLM-based AI agents across edge servers. It reports 10-50% reductions in delay and resource use in tests on 2-4 Huawei Cloud servers, though with weak baselines and no error bars.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The AntLLM gains are unattributed: the LLM post-optimization step (Alg.1 line 10, Alg.2 line 13) is unspecified and never ablated, so the reported 10.31-10.64% delay reductions cannot be assigned to the proposed hybrid.","rationale":"The reader's weakest assumption is the LLM post-optimization step, and I think that is the right locus. The entire novelty of \"AntLLM\" relative to ordinary ACO is this step, yet the manuscript gives no way to reproduce it and no experiment that isolates it. Without such an ablation, the correct interpretation of Figures 2-5 is \"ACO plus an unspecified refinement beats Greedy/Polling/Random in a handful of runs\"; it is not \"LLM-based optimization helps.\" The lack of error bars compounds this, because ACO and Random are stochastic and the testbed has only 2-4 nodes/5-7 tasks, so even the direction of the comparison is not statistically established. I also note the abstract's 9.5%/11.5% numbers do not correspond to any reported experiment, which further weakens the headline. None of this imputes bad faith; an ablation and prompt disclosure would likely settle the matter. The direction of the work is reasonable and the AgentScope implementation is a real artifact, but the central claim as stated is currently not supported.","tokens_in":8313,"tokens_out":5754,"duration_ms":64268,"concrete_test":"Run the 3-server/6-task condition for 30 independent seeds in three arms: (a) AntLLM exactly as implemented, with the prompt and validation logic documented; (b) ACO-only, omitting Algorithm 1 line 10 and Algorithm 2 line 13; (c) ACO with a non-LLM random refinement of the same budget. Report per-arm mean and 95% bootstrap CI for total delay and resource consumption, and the pairwise differences. If arm (a) does not beat arm (b) by more than the CI width, the LLM component has no measurable effect and the paper's central \"AntLLM\" claim is unsupported; if arm (a) also fails to beat Greedy by more than the CI width, the headline reductions are noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the hybrid AntLLM algorithm reduces total delay by ~10.3-10.6% and resource consumption by ~38.6-49.6%. For that claim to hold, the LLM refinement step must be a well-defined, effective component of the algorithm. It is not. Algorithm 1 line 10 and Algorithm 2 line 13 read only \"Use MLLM to refine and validate S*/M*\". No prompt, no input/output schema, no validation rules, no failure behavior, and no code are provided; the post-optimization step is a black box. The paper also reports no ablation that removes the LLM step, so the measured gains cannot be attributed to the AntLLM hybrid. If the step is trivial, stochastic, or arbitrary in the authors' implementation, the numerical results are not reproducible and the central contribution is unsupported. A secondary but compounding issue is that all comparisons lack repeated trials and error bars for a randomized ACO algorithm and a Random baseline, so the 10.31%/38.56% numbers are not shown to exceed noise; the abstract's 9.5%/11.5% figures are also not traceable to any result in Section V. The most defensible version of the claim would be \"ACO with some unspecified refinement reduces delay/cost in these testbed runs,\" not \"AntLLM reduces delay/cost.\"","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper formulates the Edge Agent Deployment (EAD) and Edge Agent Migration (EAM) problems for LLM-based AI agents in edge intelligence systems, and proposes AntLLM, a hybrid algorithm that combines ant colony optimization with an LLM-based post-optimization step. The authors implement the system using AgentScope on Huawei Cloud servers in four regions, evaluate it on CBIR tasks, and compare it with Random, Greedy, and Polling baselines. They report average total-delay reductions of 10.31% (varying the number of servers) and 10.64% (varying the number of tasks), and resource-consumption reductions of 38.56% and 49.61%.","tokens_in":8629,"tokens_out":5023,"duration_ms":58314,"significance":"If the results are reproducible, the paper would offer a practical and timely contribution to edge AI agent lifecycle management, with a useful emphasis on agent-specific state transfer and resource-aware placement. The real distributed implementation with AgentScope is a strength, as is the explicit treatment of memory/configuration migration rather than full container migration. However, the central quantitative claims are currently not well supported: the LLM refinement step is unspecified and unablated, the stochastic experiments lack repeated trials and error bars, and the baseline set is weak. The significance therefore remains conditional on a substantial revision that pins down and validates the LLM component and the statistical evidence.","major_comments":[{"comment":"The only element that distinguishes AntLLM from a standard ant colony optimization is the post-optimization step reading 'Use MLLM to refine and validate S*/M*'. The manuscript gives no prompt template, input/output schema, validation rules, failure behavior, or cost model for the LLM calls, and the parameters of the LLM refinement (e.g., number of samples, temperature, validation criteria) are not listed. No ablation removes the LLM step, so the reported gains in Section V cannot be attributed to the AntLLM hybrid rather than to ACO alone or to run-to-run variation. A revision must specify the LLM protocol and include an ablation comparing ACO with and without the LLM refinement.","section":"§III-A, Algorithm 1 line 10; §III-B, Algorithm 2 line 13"},{"comment":"All delay and cost comparisons are reported as point averages with no error bars, confidence intervals, or significance tests, even though both AntLLM and Random are randomized procedures. The claim 'reduces the total delay by an average of 10.31%' (Section V-B) is only meaningful if the spread across repeated trials is reported; with the current data the improvement is not shown to exceed noise. The authors should provide repeated runs with distinct seeds, per-point variance, and a paired significance test or effect-size measure.","section":"§V-B, Figs. 2–5"},{"comment":"The baseline set contains only Random, Greedy, and Polling. For the small instance sizes used (2–4 servers and 5–7 tasks), an optimal or near-optimal reference (e.g., ILP or exhaustive search for the static EAD problem) is feasible and would calibrate how far the heuristic is from optimality, giving the 'significantly reduces' claim a meaningful absolute reference. Without such a baseline, the reported 10.31%/10.64% improvements are only relative to very weak competitors and do not establish that AntLLM is a strong solver.","section":"§V-A"},{"comment":"The abstract reports average reductions of 9.5% deployment latency and 11.5% migration cost, but Section V-B reports total-delay reductions of 10.31% and 10.64% and resource-consumption reductions of 38.56% and 49.61%. No experiment or table in Section V reports the 9.5%/11.5% figures, and the relationship between 'deployment latency', 'migration cost', and the metrics in Figs. 2–5 is not defined. The claims should be reconciled and tied to specific figures, or the abstract should be corrected.","section":"Abstract and §V-B"}],"minor_comments":[{"comment":"Several axis labels are garbled, e.g., 'G eedy', 'Greed', 'T a k', 'Number f N des', 'P lling', 'Rand m', 'A tLLM', 'Memo y Usage', 'VM U age', and 'VMs Usage'. Please regenerate the figures with clean labels.","section":"Figs. 2–5"},{"comment":"The text says 'The ALM algorithm is illustreted in Algorithm 2'; this should be 'illustrated'. Also, the algorithm boxes use 'MLLM' while the abstract and introduction use 'LLM'; please unify the terminology.","section":"§III-B"},{"comment":"The cost components in Eqs. (7)–(10) sum quantities with different units (seconds, MB, Mbps) into a single total cost O_r. The objective and constraint formulation would benefit from explicit normalization or weights, and the displayed statement of Problem EAD appears incomplete, with an unclosed summation in the objective. Please provide the full, well-formed optimization problem.","section":"§II-B, Eqs. (7)–(10)"},{"comment":"The manuscript does not list the numerical values of key parameters (alpha, beta, rho, Q, number of ants, number of iterations, H_th, gamma, theta_dep, and the LLM configuration), nor does it state whether code or data will be released. Without these, the experiments are not reproducible even if the algorithmic description were complete.","section":"§V-A"}],"recommendation":"major_revision","confidential_remarks":"The paper's novelty claim of being 'the first systematic deployment and management solution' should be checked carefully against recent literature on LLM-agent placement and edge resource scheduling; the reference list leans on older container-migration work. The central empirical claim will remain unreliable unless the LLM ablation and statistical rigor are added; if the authors cannot specify the LLM post-optimization step in concrete terms, I would move toward rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper has a credible problem statement and one genuinely interesting design choice—treating an AI agent's state as memory plus configuration, so migration moves a few MBs, not the whole runtime. That's the right way to think about agent mobility at the edge. The ACO-based placement and migration framework is also standard but cleanly formulated. If the only claim were 'ACO works reasonably for agent placement in a small testbed,' I'd be fine with it.\n\nThe trouble is the claim is about AntLLM, the hybrid with LLM post-optimization. The paper never specifies what the LLM does in Algorithms 1 and 2 ('Use MLLM to refine and validate S*'), no prompt, no output schema, no validation logic, and no ablation removing it. So the 10.31-10.64% delay reduction and 38.56-49.61% resource savings cannot be attributed to the hybrid. The LLM could be doing nothing or could be adding noise. This is the load-bearing flaw.\n\nCompounding: the experiments have no error bars or significance tests, the baselines are only Random, Greedy, and Polling, the testbed is 2-4 servers and 5-7 tasks, and the abstract's 9.5%/11.5% improvements don't match the body's numbers. Those last points are fixable in a revision, but they make the current manuscript a hard reject as-is.\n\nWhat's genuinely good: the system implementation on AgentScope is real engineering, the resource and latency model is sensible, and the idea that agent migration is cheaper than container migration because only memory/config need to move is worth stating explicitly. The paper also cites the relevant prior work on container and service migration, though the references are heavily self-citations—not a problem in itself, but worth a glance.\n\nWho is this for: people working on edge AI orchestration. It's a reasonable system prototype, but not yet a reliable algorithmic contribution. The LLM step needs to be either specified and ablated or dropped. If the authors add a rigorous evaluation with repeated trials, confidence intervals, and a proper ablation, the core idea could be salvageable.\n\nRecommendation: send it to peer review—the topic is timely and the direction is not wrong—but tell the authors clearly that the current evidence doesn't support the central claim. A serious refereeing process could force the missing ablation and statistical rigor. Without that, I wouldn't rely on any of the numbers.","headline":"The lightweight-migration idea is sensible, but the paper's key novelty—the LLM refinement step—is unvalidated and the experiments are too thin to support the claimed gains.","tokens_in":9153,"tokens_out":3850,"would_cite":false,"duration_ms":40748,"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":"AntLLM cuts edge AI agent deployment latency by ~9.5% and migration cost by ~11.5% by combining ant-colony search with an LLM refinement step.","keywords":["AI agent placement","edge intelligence","agent migration","ant colony optimization","LLM-based agents","quality of service","resource allocation","distributed systems"],"falsifier":"Run the exact same CBIR workloads on the same four edge servers with the LLM post-optimization step removed, leaving only the ant-colony search; if the delay and resource-consumption improvements over Greedy, Polling, and Random are unchanged, the LLM refinement is not responsible for the reported gains and the distinguishing LLM part of AntLLM is unsupported. Conversely, if deleting the LLM step erases the margin, the LLM is doing the work and its prompt and validation logic would need to be reported for the result to be reproducible.","tokens_in":8104,"feed_emoji":"🐜","tokens_out":5691,"duration_ms":60372,"temperature":0.7,"pith_summary":"The paper asks how LLM-based AI agents, which must plan, call tools, and remember state, should be placed on edge servers and moved as users travel. It claims that a hybrid algorithm, AntLLM, solves both the placement and migration problems better than greedy, polling, or random assignment. The algorithm first uses ant colony optimization to construct low-cost deployment and migration plans, then hands the best plan to an LLM for refinement and validation. On a distributed testbed spanning four edge locations, the paper reports an average reduction of about 10.3-10.6% in total task delay and 38.6-49.6% in resource consumption compared with baselines, with the abstract citing 9.5% lower deployment latency and 11.5% lower migration cost. The contribution is a systematic cost/latency model for agent placement and migration plus an algorithm that exploits the lightweight nature of agent migration, which transfers only memory and configuration rather than a full codebase.","feed_headline":"Ant-colony placement cuts edge AI agent delay by ~10%","feed_subtitle":"LLM-based agents can migrate with only memory and config, following mobile users across edge servers while saving resources.","key_machinery":"The load-bearing mechanism is the AntLLM algorithm, which pairs ant colony optimization with a final LLM-based post-optimization step. Each ant constructs a candidate placement or migration plan by probabilistically choosing servers according to a pheromone matrix and a heuristic built from the EAD/EAM cost models; the best plan is then passed to an LLM, marked in the pseudocode as 'Use MLLM to refine and validate S*' (or M*), before becoming the final decision. For migration, the algorithm uses a net-gain objective: expected latency improvement minus migration time and overhead minus dependency damage, and only triggers migration when the user has moved beyond a hop threshold or the hosting server's resources fall below a safety level. Agent migration is kept lightweight by transferring memory and configuration rather than the full agent codebase.","core_discovery":"The central claim is that the Edge Agent Deployment (EAD) and Edge Agent Migration (EAM) problems can be formulated as resource-constrained optimization problems and solved by an ant-colony heuristic whose best solution is then refined by an LLM. The paper defines total task latency as transmission plus migration plus agent-initiation plus processing latency, and total cost as computing plus storage plus communication cost, and minimizes these subject to memory, storage, and communication capacity constraints. The migration decision compares the latency benefit of moving an agent closer to the user against migration time, overhead cost, and damage to inter-agent dependency latencies. The paper's experiments, run with content-based image retrieval tasks on edge servers distributed across four regions, show AntLLM outperforming Greedy, Polling, and Random baselines on delay and resource metrics, and the authors conclude that their system autonomously places and migrates agents to maintain quality of service in dynamic edge environments.","pith_inferences":["An obvious next test the paper does not run is an ablation that removes the LLM post-optimization step; without it, a reader cannot tell whether the reported gains come from the ant-colony search, the LLM refinement, or the interaction.","The same post-optimization pattern could transfer to other combinatorial resource-allocation problems in edge computing, such as service caching or container placement, but the transfer is speculative until the LLM step is specified and benchmarked.","Because the LLM refinement presumably incurs its own latency and API cost, a fair comparison would need to include the overhead of calling the LLM inside the decision loop; the current measurements appear to report only the resulting deployment and migration outcomes.","The lightweight migration assumption rests on memory and configuration being the only agent state; agents with large vector stores, tool caches, or local model adapters would need a richer state model before the same gains could be expected."],"forward_implications":["If the reported results hold, LLM-based agents can be kept close to mobile users at roughly 10% lower total task delay than greedy, polling, or random placement, without sacrificing quality of service.","The 38.6-49.6% resource-consumption reductions imply that the same edge infrastructure can host substantially more agents, or the same agents on cheaper servers, than baseline deployment policies.","The paper's cost/latency model gives later work a concrete objective for joint placement and migration of agents that carry memory and tool dependencies.","The migration policy, which triggers on user movement or resource bottlenecks and migrates only memory and configuration, offers a directly implementable rule for sustaining agent state continuity in edge networks."],"supporting_citations":[{"why":"Supplies the multi-agent platform used to implement the distributed agent deployment and migration system.","marker":"[5]"},{"why":"Establishes edge intelligence as the setting where latency-sensitive AI workloads should be processed.","marker":"[6]"},{"why":"Provides the container migration algorithm and QoS logic that the agent migration design extends.","marker":"[3]"},{"why":"Supplies migration modeling and learning methods for fog and edge environments that ground the migration latency formulas.","marker":"[4]"},{"why":"Informs the collaborative service placement formulation that the EAD problem builds on.","marker":"[7]"},{"why":"Motivates LLM QoS optimization across cloud and edge, the direct application domain of this work.","marker":"[10]"},{"why":"Describes AI-augmented edge computing challenges that justify agent-aware resource management.","marker":"[17]"},{"why":"Models migration in dynamic vehicular edge networks, supporting the adaptive migration trigger design.","marker":"[2]"}],"fun_headline_variants":["Ant+LLM cuts edge agent latency and migration costs","Ant colony + LLM trims edge AI agent delay","LLM-ant hybrid cuts edge agent latency and cost","Ant+LLM edge agents migrate with only state","Ant colony plus LLM shrinks edge agent delay"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the LLM post-optimization step actually improves the ant-colony solution; the paper gives no prompt, no output format, no validation logic, and no ablation that isolates this step, so if the LLM call is arbitrary or unreliable the reported 10.31-10.64% delay and 38.56-49.61% resource gains could come from the ant-colony algorithm alone.","fun_headline_variants_meta":{"raw":{"variants":["Ant+LLM cuts edge agent latency and migration costs","Ant colony + LLM trims edge AI agent delay","LLM-ant hybrid cuts edge agent latency and cost","Ant+LLM edge agents migrate with only state","Ant colony plus LLM shrinks edge agent delay"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000984,"raw_usage":{"total_tokens":4161,"prompt_tokens":917,"completion_tokens":3244,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":3166}},"tokens_in":533,"tokens_out":3244,"duration_ms":28090,"temperature":1.0,"reasoning_tokens":3166,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T04:29:29.614596+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the exact same CBIR workloads on the same four edge servers with the LLM post-optimization step removed, leaving only the ant-colony search; if the delay and resource-consumption improvements over Greedy, Polling, and Random are unchanged, the LLM refinement is not responsible for the reported gains and the distinguishing LLM part of AntLLM is unsupported. Conversely, if deleting the LLM step erases the margin, the LLM is doing the work and its prompt and validation logic would need to be reported for the result to be reproducible.","supporting_citations":[{"cited_title":"Agentscope: A flexible yet robust multi- agent platform,","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-agent platform used to implement the distributed agent deployment and migration system."},{"cited_title":"Edge intelligence: Paving the last mile of artificial intelligence with edge computing,","cited_arxiv_id":null,"evidence_quote":"Establishes edge intelligence as the setting where latency-sensitive AI workloads should be processed."},{"cited_title":"Migration mod- eling and learning algorithms for containers in fog computing,","cited_arxiv_id":null,"evidence_quote":"Supplies migration modeling and learning methods for fog and edge environments that ground the migration latency formulas."},{"cited_title":"Collaborative service placement for edge computing in dense small cell networks,","cited_arxiv_id":null,"evidence_quote":"Informs the collaborative service placement formulation that the EAD problem builds on."},{"cited_title":"Enhancing LLM QoS through Cloud-Edge Collaboration: A Diffusion-based Multi-Agent Reinforce- ment Learning Approach ,","cited_arxiv_id":null,"evidence_quote":"Motivates LLM QoS optimization across cloud and edge, the direct application domain of this work."},{"cited_title":"Ai augmented edge and fog computing: Trends and challenges,","cited_arxiv_id":null,"evidence_quote":"Describes AI-augmented edge computing challenges that justify agent-aware resource management."},{"cited_title":"Adaptive digital twin migration in vehicular edge computing and networks,","cited_arxiv_id":null,"evidence_quote":"Models migration in dynamic vehicular edge networks, supporting the adaptive migration trigger design."}],"review_version":1}