{"id":"27a599e3-92fc-4bfd-bff6-7c45e15817f0","arxiv_id":"2412.20820","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"A RAG-based LLM framework is applied to MEC task offloading and reported to cut average latency by up to 86% versus DQN, DDPG, and PPO in simulation.","lead":"This paper proposes using retrieval-augmented generation (RAG) with large language models to make task offloading decisions in mobile edge computing, claiming lower latency than deep-learning baselines. The method retrieves each user's computing capability and asks an LLM to output offloading ratios, powers, and resource splits, but the retrieval is essentially a parameter lookup and no code is provided.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed latency gains are not established because LLM outputs are evaluated without checking constraint (12d), letting every user receive full server capacity β_k=1 simultaneously, which alone can manufacture large improvements.","rationale":"The central claim is that the RAG-LLM method consistently reduces average latency relative to DRL baselines. For that claim to be meaningful, all compared policies must be feasible for P1, or the comparison must account for constraint violations. The weakest point in the paper is exactly this: decisions are generated by an LLM in Eq. (15), and no constraint projection, feasibility verification, or penalty mechanism is described anywhere in Section IV or Section V. In particular, constraint (12d) is a strict coupling: the sum of the computational-resource ratios β_k(t) cannot exceed 1. If the LLM outputs β_k(t)=1 for every user, Eq. (10) treats each user as having exclusive access to the full server capacity F, which is physically impossible in the stated system model. This can, by itself, produce arbitrarily large latency reductions over any feasible policy. The energy constraint (12e) similarly has no enforcement in the evaluation, allowing power to be used freely to shorten offload time. The DRL baselines are described as having bounded action spaces (e.g., sigmoid outputs in [0,1]) but no explicit handling of the sum constraint or energy budget, so the experimental comparison is at best uncontrolled. The retrieval-tautology issue raised by the reader is real but secondary; it undermines the 'retrieval-augmented' novelty, not the latency numbers. The load-bearing technical flaw is the feasibility gap. The concrete test described above would settle whether the reported improvements survive a minimal projection onto the feasible set. Because this concern supports the reader's REJECT verdict rather than overturning it, I recommend no change to the verdict.","tokens_in":14477,"tokens_out":4761,"duration_ms":49524,"concrete_test":"Reproduce one experiment, e.g., the DUSD curve in Fig. 6, by recording the LLM's raw (α, p, β) outputs. Check Σ_k β_k(t) ≤ 1 for every time slot and Σ_t(E_k(t)+E_off_k(t)) ≤ E_max for every user. Then recompute average latency under a minimal feasible projection: β'_k(t) = β_k(t) / Σ_j β_j(t), and clip p and α so the energy budget is met by reducing offloaded data. Compare the projected RAG latency to the DRL baselines in the same figure. If the gap shrinks below the claimed 42% or reverses, the central claim is an artifact of infeasible outputs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV.B's generation module directly outputs decisions (Eq. (15)), and the evaluation section computes latency via Eq. (11) with no described projection, feasibility filter, or penalty for violating P1's constraints. The most damaging case is constraint (12d): Σ_k β_k(t) ≤ 1. If the LLM returns β_k(t)=1 for all K users, Eq. (10) gives each user an entire dedicated server of capacity F, so edge latency collapses even though the real server is K-fold oversubscribed. Similarly, without enforcing the energy budget (12e), setting p_k(t)=Pmax always reduces offloading latency with no energy penalty. The DRL baselines are described with bounded action spaces but no explicit enforcement of these coupled constraints either, so the comparison is at best uncontrolled. Unless the exact outputs used for Figs. 6-9 are shown to satisfy (12d) and (12e), or are projected onto the feasible set before latency is computed, the reported 57%, 86%, 30%, and 42% improvements are consistent with an apples-to-oranges comparison between infeasible LLM outputs and feasible or penalized baselines rather than a genuine gain.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a retrieval-augmented generation (RAG) framework for mobile edge computing (MEC), where a large language model (LLM) generates task offloading decisions—offloading ratio, transmit power, and server resource allocation—to minimize average latency. A vector knowledge base stores each user's computing capability, a Bi-encoder retrieves the user's own configuration, and GPT-4o/Qwen2.5-turbo generate decisions from the retrieved context. The authors evaluate the method on four data sets and report latency improvements of 57%, 86%, 30%, and 42% over DQN, DDPG, and PPO baselines. The central claim is that retrieval-augmented LLM decisions outperform deep reinforcement learning in average latency.","tokens_in":14685,"tokens_out":3567,"duration_ms":33159,"significance":"If established, the result would be a notable demonstration of LLM-based resource allocation in MEC. The paper has strengths: it does not fit free parameters to the test set, it evaluates several LLM and embedding configurations, and it formulates a clear constrained optimization problem with explicit constraints. These strengths, however, are outweighed by the absence of any feasibility guarantee or verification for the LLM outputs, a retrieval mechanism that only returns the user's own stored parameter, and a lack of statistical or optimality baselines. The reported gains are therefore not supported by the presented evidence. The paper is best viewed as a preliminary position that could become a serious contribution only after substantial reworking of both the method and the evaluation.","major_comments":[{"comment":"The generation module directly outputs decisions, but the paper does not describe any projection onto the feasible set of problem P1. In particular, constraint (12d), Σ_k β_k(t) ≤ 1, and the energy budget (12e) are not enforced at inference time. If the LLM returns β_k(t)=1 for all k, Eq. (10) gives each user the full server capacity F, artificially reducing edge latency in Figs. 6–9. The authors must either verify that the actual stored LLM outputs satisfy (12a)–(12e) and report the verification, or add a projection/penalty step and recompute all latency comparisons.","section":"Section IV.B, Eq. (15); Section III.D, Eqs. (12a)–(12e)"},{"comment":"The retrieval step is self-referential: it encodes each user's own f_k into the vector database and then retrieves that same configuration using the user's identity K. Consequently, Table II's hit rate and MRR measure the trivial task of retrieving exactly the item just stored, not the utility of retrieved context for decision quality. No experiment compares the full RAG pipeline against the same LLM without retrieval or with retrieval of a different user's configuration, so the paper does not establish that retrieval contributes to the reported latency gains. This is a central claim of the title and abstract, and the current experimental design cannot support it.","section":"Section IV.B, Eqs. (13)–(14); Table II"},{"comment":"No variance, confidence intervals, or multiple-seed results are reported, and no classical optimization baseline (e.g., exhaustive search on a coarse grid, convex relaxation, or greedy allocation) is included. Because LLM generation is stochastic, the reported 57%, 86%, 30%, and 42% improvements cannot be distinguished from sampling noise. Without an optimality reference, the claim that the method 'optimizes' latency is unquantified and unverifiable.","section":"Section V.D, Figs. 6–9"},{"comment":"The notation L_k = LLM(P) is ambiguous: the text says the module 'output optimization decisions,' but the equation writes L_k, which was already defined in Eq. (11) as the latency objective. The manuscript must specify exactly which quantities the LLM returns (α_k, β_k, p_k) and how they are parsed from the generated text before the numerical evaluation can be reproduced or trusted.","section":"Section IV.B, Eq. (15)"}],"minor_comments":[{"comment":"The heading 'Genrative Learning for Computation Offloading' contains a typo and should read 'Generative Learning for Computation Offloading.'","section":"Section II.D heading"},{"comment":"The text says 'Let f_m denote the computing cycles per second of the kth MU,' but Eq. (1) uses f_k. The notation should be made consistent throughout the paper.","section":"Section III.A and Eq. (1)"},{"comment":"Eq. (13) stores v_k = Encode(f_k), while Table I expresses user computing capability as ϕ/f_k in seconds per bit. The relationship between f_k and ϕ/f_k should be clarified, and the units in Table I should be stated explicitly.","section":"Section IV.B, Eq. (13); Table I"},{"comment":"The DDPG action space is described as 'constrained to [0, 1],' but no mechanism is described that enforces the coupled constraint Σ_k β_k(t) ≤ 1. To ensure a fair comparison, the baseline implementations should be described with the same feasibility treatment as the proposed method.","section":"Section V.C"},{"comment":"The figure captions do not state the units of the y-axis (latency) or the averaging procedure used to produce each curve, which reduces reproducibility.","section":"Section V.D, Figs. 6–9"}],"recommendation":"reject","confidential_remarks":"The core experimental evidence does not support the feasibility or optimality of the LLM outputs, and the retrieval mechanism is tautological. A revision would require a new retrieval design and a full feasibility-corrected evaluation, which is beyond the scope of a standard revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is the first RAG-for-MEC offloading paper I know of, but the headline gains are not believable because the LLM outputs are never checked against the constraints of P1, and the retrieval step reduces to looking up each user's own f_k. The central comparison is apples-to-oranges.\n\nWhat is genuinely new: the authors actually run LLM-based decision-making (GPT-4o, Qwen2.5-Turbo, three bge embeddings) across four parameter scenarios, and they frame it as RAG for MEC. The problem formulation is standard and clearly written, and the citation coverage looks reasonable. The retrieval scoring table at least shows that the embeddings can identify the intended user configuration. This is not a lazy paper; it just does not support its own claims.\n\nThe soft spots are serious. Section IV.B's generation module (Eq. 15) outputs α, p, and β directly. The evaluation then computes latency with Eq. (11), but the paper never says the outputs satisfy (12a)-(12e). In particular, (12d) says Σ_k β_k(t) ≤ 1. If the LLM returns β_k=1 for every user, each user gets the entire server capacity F, and offloading latency collapses. The reported 57%, 86%, 30%, and 42% improvements are exactly what you would expect from that inflation. The stress-test note is correct. The same issue affects the DRL baselines: DDPG and PPO use per-action sigmoids that bound each β to [0,1] but do not enforce the sum constraint, so the comparison is uncontrolled on both sides. There are also no error bars, no classical or optimal baseline (e.g., a convex relaxation or a simple feasible greedy solution), and no code or data artifact. The retrieval is tautological: storing each user's own f_k and retrieving it by user identity is prompt construction, not retrieval-augmented generation in any meaningful sense.\n\nThe paper is coherent and well written, but the central claim is unsupported. I would not send this to a serious referee in its current form. A resubmission with a feasibility projection, a feasible optimal/classical baseline, and a demonstration that the exact outputs in Figs. 6-9 satisfy (12d) and (12e) would be a modest but legitimate applied contribution. As is, it deserves a desk reject with a clear resubmission path.","headline":"The RAG-for-MEC hook is timely, but the reported latency gains look like artifacts of unconstrained LLM outputs, and the 'retrieval' is just a per-user parameter lookup.","tokens_in":15215,"tokens_out":2903,"would_cite":false,"duration_ms":31153,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A retrieval-augmented LLM can cut mobile edge offloading latency by up to 86 percent, beating deep RL baselines.","keywords":["retrieval-augmented generation","mobile edge computing","task offloading","resource allocation","large language models","latency minimization","deep reinforcement learning","Bi-encoder retrieval"],"falsifier":"Run the proposed RAG pipeline on a small instance (e.g., two users, one server, one time slot), verify whether the returned $\\alpha_k$, $p_k$, and $\\beta_k$ satisfy constraint (12d) and the energy constraint (12e), and compare the achieved average latency to an exhaustive-search solution of P1. If feasibility fails or the gap to the optimum exceeds the reported margins, the central claim collapses.","tokens_in":1702,"feed_emoji":"⚡","tokens_out":2273,"duration_ms":68950,"temperature":0.7,"pith_summary":"This paper tries to show that retrieval-augmented generation (RAG) can replace deep reinforcement learning as the decision engine for computation offloading in mobile edge computing. It formulates a latency-minimization problem over the offloading ratio, transmit power, and server resource allocation, then proposes a three-module pipeline: a vector database stores each user's computing capability, a retrieval module matches the current query to the most relevant stored configurations, and an LLM generates the offloading decisions from a prompt that combines retrieved context with current system state. The reported experiments claim consistent latency reductions over DQN, DDPG, and PPO baselines, with improvements of 57%, 86%, 30%, and 42% under varying user computing capability, server capability, transmit power, and data volume. If true, the value is practical: an LLM-driven controller could adapt to new conditions by retrieval and re-prompting instead of retraining.","feed_headline":"LLM retrieval cuts edge offloading latency up to 86%","feed_subtitle":"No training, just retrieval: an LLM picks offloading ratios and beats deep RL on average latency.","key_machinery":"The central object is the RAG pipeline of three modules. The storage module encodes each user's computing cycles per second $f_k$ into a vector $v_k = \\text{Encode}(f_k)$; the retrieval module computes cosine similarity $\\text{Sim}(q,v_k) = q \\cdot v_k / (\\|q\\|\\|v_k\\|)$ between a query built from user identities and data volumes and the stored vectors; the generation module concatenates the retrieved computing capability $C$, the edge server's capability $F$, the data volumes $D$, bandwidth $B$, and noise power $\\sigma^2$ into a prompt $P = (K\\|D\\|C\\|F\\|B\\|\\sigma^2)$ from which the LLM outputs decisions $L_k = \\text{LLM}(P)$. This pipeline is what lets the system convert a stored-knowledge retrieval problem into a text-generation problem, and it carries the paper's entire argument.","core_discovery":"The paper claims that the latency-minimization problem P1 for a multi-user MEC system can be solved by a retrieval-augmented generation pipeline: store each user's local computing capability as a vector, retrieve the best-matching configuration for the current query, concatenate it with the current data volumes, server capability, bandwidth and noise into a prompt, and let an LLM output the offloading ratio, transmit power, and server allocation ratio. In experiments, this pipeline reports 57%, 86%, 30%, and 42% lower average latency than DQN, DDPG, and PPO baselines under varying user computing capability, server capability, transmit power, and data volume, respectively.","pith_inferences":["The paper's approach effectively replaces constrained optimization with a single shot of LLM prompting; adding a constraint-projection or feasibility-check layer after generation would likely improve reliability when the method is scaled to larger systems.","The retrieval module only indexes user computing capability, so the reported gains may be dominated by the retrieval of that one parameter; retrieving other time-varying context such as channel gains, server load history, or energy budgets could matter more in truly dynamic scenarios.","Because the LLM is used to output continuous-valued decisions, the method depends on the model's numerical compliance; a cheap testable improvement would be to reject and re-prompt when outputs violate constraints.","The comparison is only against deep RL baselines; comparing against a classical optimization heuristic or a convex relaxation would isolate whether the gains come from the RAG/LLM approach or from the problem structure itself."],"forward_implications":["A retrieval-augmented LLM can produce offloading decisions without training, beating deep RL baselines in average latency across varying user computing capability, server capability, transmit power, and data volume.","The method's reported latency improvements are 57%, 86%, 30%, and 42% on the four test datasets, suggesting the benefit holds across different sources of system heterogeneity.","Retrieval hit rates between 0.855 and 0.948 and MRR values between 0.846 and 0.948 indicate that stored computing-capability configurations are matched to users with high accuracy, supporting the claim that retrieved context is relevant.","Because adaptation happens through retrieval and prompting rather than gradient updates, the approach can respond to changing system parameters without retraining.","Since RAG provides a traceable path from retrieved information to the generated decision, the offloading choices are more interpretable than those of deep RL policies."],"supporting_citations":[{"why":"Introduces retrieval-augmented generation, the architectural foundation the paper adapts to MEC offloading.","marker":"[20]"},{"why":"Presents a model-free RL offloading algorithm that learns long-term offloading strategies, serving as one of the baselines.","marker":"[24]"},{"why":"Proposes deep RL for online computation offloading in wireless powered MEC, one of the DRL baselines the proposed method is compared against.","marker":"[28]"},{"why":"Develops deep RL for energy-efficient computation offloading in multi-user MEC, another baseline in the evaluation.","marker":"[31]"},{"why":"Applies RAG to engineering design knowledge, informing the design of RAG for non-NLP tasks.","marker":"[43]"},{"why":"Provides packed resources for general embeddings, supporting the Bi-encoder retrieval component used for dense retrieval.","marker":"[44]"},{"why":"Defines the mean reciprocal rank and hit-rate metrics used to evaluate retrieval quality in the paper.","marker":"[46]"}],"fun_headline_variants":["RAG-LLM slashes MEC offloading latency by 86%","LLM retrieval beats deep RL in edge task offloading","Retrieval-augmented LLM cuts edge latency by 86%","Edge offloading: LLM retrieval outperforms DL, 86% less"],"cache_read_input_tokens":17408,"weakest_assumption_plain":"The load-bearing premise is that the LLM's raw output already satisfies the constraints of problem P1 (the sum of server shares below 1 and each user's energy budget, among others) and is close to optimal, even though the pipeline performs no feasibility check, projection, or optimality verification.","fun_headline_variants_meta":{"raw":{"variants":["RAG-LLM slashes MEC offloading latency by 86%","LLM retrieval beats deep RL in edge task offloading","Retrieval-augmented LLM cuts edge latency by 86%","Edge offloading: LLM retrieval outperforms DL, 86% less"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000759,"raw_usage":{"total_tokens":3345,"prompt_tokens":890,"completion_tokens":2455,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":2375}},"tokens_in":506,"tokens_out":2455,"duration_ms":15221,"temperature":1.0,"reasoning_tokens":2375,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:09:03.604977+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the proposed RAG pipeline on a small instance (e.g., two users, one server, one time slot), verify whether the returned $\\alpha_k$, $p_k$, and $\\beta_k$ satisfy constraint (12d) and the energy constraint (12e), and compare the achieved average latency to an exhaustive-search solution of P1. If feasibility fails or the gap to the optimum exceeds the reported margins, the central claim collapses.","supporting_citations":[{"cited_title":"Retrieval-augmented generation for knowledge-intensive NLP tasks,","cited_arxiv_id":null,"evidence_quote":"Introduces retrieval-augmented generation, the architectural foundation the paper adapts to MEC offloading."},{"cited_title":"Learning for com- putation offloading in mobile edge computing,","cited_arxiv_id":null,"evidence_quote":"Presents a model-free RL offloading algorithm that learns long-term offloading strategies, serving as one of the baselines."},{"cited_title":"Deep reinforcement learning for online computation offloading in wireless powered mobile-edge computing networks,","cited_arxiv_id":null,"evidence_quote":"Proposes deep RL for online computation offloading in wireless powered MEC, one of the DRL baselines the proposed method is compared against."},{"cited_title":"Deep reinforcement learning for energy-efficient computation offloading in mobile-edge computing,","cited_arxiv_id":null,"evidence_quote":"Develops deep RL for energy-efficient computation offloading in multi-user MEC, another baseline in the evaluation."},{"cited_title":"Retrieval augmented generation using engi- neering design knowledge,","cited_arxiv_id":null,"evidence_quote":"Applies RAG to engineering design knowledge, informing the design of RAG for non-NLP tasks."},{"cited_title":"C- pack: Packed resources for general Chinese embeddings,","cited_arxiv_id":null,"evidence_quote":"Provides packed resources for general embeddings, supporting the Bi-encoder retrieval component used for dense retrieval."},{"cited_title":"Evaluating retrieval quality in retrieval- augmented generation,","cited_arxiv_id":null,"evidence_quote":"Defines the mean reciprocal rank and hit-rate metrics used to evaluate retrieval quality in the paper."}],"review_version":1}