{"id":"7b296788-096a-464e-849a-210d0513ab52","arxiv_id":"2412.00090","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An edge split-learning framework with the CARD algorithm dynamically chooses the cut layer and server GPU frequency to trade off LLM fine-tuning delay against server energy.","lead":"This paper proposes an adaptive split-learning scheme for fine-tuning large language models across edge devices and a server, deciding where to cut the model and how fast to run the server's GPU. It reports 70.8% lower training delay than the device-only baseline and 53.1% lower server energy than the server-only baseline in simulations with five Jetson devices.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The per-device independence assumption in Eq. (10) omits contention for one shared edge server and one wireless channel, so the 70.8%/53.1% headline reductions are not supported as stated.","rationale":"The reader’s weakest assumption identifies exactly the point I would stress: Eq. (10) and the Section IV decomposition assume that each device’s delay is independent of the other devices. The paper provides no justification that a single edge server can serve all devices without contention, and the optimization variable f^S_{m,n} is indexed per device, which is inconsistent with a single physical GPU. This is load-bearing because the headline 70.8% and 53.1% numbers are produced from this model. I do not see a fatal internal error in the CARD derivation itself: the convexity argument for P3 is correct, Eq. (16) follows from the first-order condition, and the brute-force search over cut layers is valid if the per-device model is accepted. The absence of fine-tuning quality metrics is a separate concern, but it is secondary to the efficiency claim. The proposed concrete test—re-solving with a single-server coupling constraint—would settle whether the contention objection actually changes the reported reductions. Since the reader already made the verdict CONDITIONAL, my analysis does not move that verdict; it sharpens the condition that must be met.","tokens_in":8724,"tokens_out":4417,"duration_ms":47075,"concrete_test":"Re-run the Section V-B comparison with a shared-server model: one server GPU with a single frequency f_n per round, server-side work per device W_S(c_{m,n}) = T(η − η_D(c_{m,n}))/(δ^S σ^S) cycles, and sequential service so each device’s delay includes accumulated server cycles of previously served devices divided by f_n. Re-solve the joint cut/frequency problem under this coupling and recompute the Fig. 4 reductions. If the 70.8% delay reduction or the 53.1% energy reduction changes materially, the abstract’s headline numbers should be re-scoped to the independent-device idealization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central efficiency claim depends on solving P1 per device independently. Section IV explicitly says this is done by 'considering the independence among these devices,' and Eq. (10) defines D_{m,n} as local computation plus transmission, with no queueing or server-sharing term. Yet the system description has one edge server and one AP serving all M devices. If multiple devices send smashed data concurrently, the server GPU must time-share or frequency-switch among devices, and the wireless uplink/downlink must share bandwidth. Under such coupling, D_{m,n} for one device depends on the cut layers and transmission schedules of all other devices, so the decomposition into subproblems P2 is not valid.\n\nThe same issue is visible in the decision variables: f^S_{m,n} in Eq. (13) is indexed by device and round, as if each device had its own server GPU frequency. Fig. 3(b) shows different optimal frequencies for different devices at the same round. A single physical GPU cannot simultaneously execute at several different frequencies. If the server instead runs devices sequentially, the round delay of a later device includes the accumulated server busy time of earlier devices, and the objective in P1 is no longer separable. The reported 70.8% delay reduction relative to device-only and 53.1% server-energy reduction relative to server-only are therefore computed under a model that may be infeasible for the described single-server deployment. This is a correctness risk for the headline numerical claims, not merely a missing extension.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an energy-efficient split learning (SL) framework for fine-tuning large language models (LLMs) in edge networks. Multiple devices with heterogeneous GPUs collaborate with a single edge server; each device fine-tunes the first part of the model with LoRA adapters up to a cut layer, and the server handles the remainder. The authors formulate an optimization problem P1 that minimizes a weighted sum of normalized training delay and server energy consumption, with variables being the per-device-per-round cut layer and server GPU frequency. They propose the CARD algorithm, which decomposes P1 into independent per-device subproblems P2, solves the frequency decision in closed form via convexity (Eq. (16)), and finds the cut layer by brute-force scanning over the finite number of transformer layers. Simulations with a 1B LLaMA 3.2 model and 5 Jetson devices report a 70.8% reduction in average training delay relative to the device-only baseline and a 53.1% reduction in server energy consumption relative to the server-only baseline.","tokens_in":8969,"tokens_out":7675,"duration_ms":71164,"significance":"If the system model were faithful, the result would be a useful and simple mechanism for jointly choosing the split layer and server frequency. The derivation is self-contained: the convexity argument for Eq. (16) is correct, the brute-force search over the finite cut-layer set is valid, and the reported delay/energy comparisons are not fitted to the model (the parameters are fixed a priori). The paper also clearly identifies a real problem—static split strategies and fixed server resources are suboptimal under heterogeneity and channel dynamics. However, the central modeling assumption that devices can be optimized independently is questionable for a single shared edge server and access point, and the paper provides no validation of fine-tuning quality. These issues significantly reduce the confidence in the headline numbers.","major_comments":[{"comment":"The decomposition of P1 into independent per-device subproblems P2 is not valid for the described system with a single edge server and a single AP. Eq. (10) defines D_{m,n} as the sum of local computation, server computation, and transmission times, with no queueing or contention. If several devices transmit smashed data concurrently or wait on the same server GPU, D_{m,n} depends on the cut layers and schedules of all other devices. Moreover, the variable f^S_{m,n} in Eq. (13) implies each device has its own server GPU frequency, but one physical GPU cannot run at several different frequencies simultaneously. If the server processes devices sequentially instead, the total training delay should include waiting for earlier devices, and the objective in P1 is not separable. The reported reductions of 70.8% and 53.1% (Section V-B, Fig. 4) are therefore not directly supported by the model as stated.","section":"Section IV, Eq. (10)-Eq. (14)"},{"comment":"No fine-tuning performance (training loss, validation loss, or downstream task accuracy) is reported. The paper defines a global loss objective in Eq. (1) but the experiments in Section V only show delay and energy consumption. It is possible that dynamically changing the cut layer and server frequency affects convergence or final model quality. Without evidence that the proposed scheme actually fine-tunes the model to a satisfactory level, the claim of an \"energy-efficient fine-tuning framework\" is incomplete. A comparison of the training loss or task accuracy with the two baselines is needed.","section":"Section V"},{"comment":"The observation that the optimal cut layer is always either 32 or 0, and never an intermediate value, indicates that the proposed optimization is essentially a binary offloading decision under the assumed LLaMA model. The paper's statement that \"the optimal cut layer of each device dynamically changes\" is misleading because it changes only between the two extremes. This should be explained (e.g., by the assumed equal per-layer FLOPs and smashed-data sizes) and its implications for the generality of the CARD algorithm should be discussed.","section":"Section V-B, Fig. 3"}],"minor_comments":[{"comment":"The abstract states that the proposed approach reduces delay and server energy by 70.8% and 53.1%, but it does not specify that the delay reduction is relative to the device-only baseline and the energy reduction is relative to the server-only baseline. This should be clarified.","section":"Abstract and Section V-B"},{"comment":"The symbol \\tilde{S}(c_{m,n}) in Eq. (9) is used for the gradient data size but is not defined in the text; please define it explicitly, e.g., alongside the smashed data size S(c_{m,n}).","section":"Eq. (9)"},{"comment":"The sentence \"the optimal cut layer of each device dynamically changes with training rounds\" in Section V-B conflicts with the immediately following sentence that the optimal cut is either 32 or 0; please rephrase to avoid implying more than one intermediate cut layer is ever selected.","section":"Fig. 3 and accompanying text"},{"comment":"The notation for transmission rates R^D_{m,n} and R^S_{m,n} is introduced, but the subscript/superscript convention is not consistently used later (e.g., in Eq. (9) the uplink and downlink rates appear as R^D and R^S, which matches, but the text at the end of the paragraph uses R_{m,n} generically). Please ensure consistent notation.","section":"Section III-A-2"},{"comment":"No sensitivity analysis is provided for the weighting factor w (set to 0.2). Since w controls the trade-off between delay and energy, a short discussion of its influence (or a supplementary figure) would strengthen the evaluation.","section":"Section V-A, Table II"}],"recommendation":"major_revision","confidential_remarks":"The main technical concern is the independence assumption. The authors should either (i) reformulate the problem to include a shared-server scheduling/contention model (or a sequential-processing model with a sum-of-delays objective) or (ii) add simulations with a time-multiplexed server and shared wireless channel to verify that the reported gains persist. Without this, the headline numbers are not convincing. The absence of any model-quality metric is also a gap that should be addressed in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a clean, small optimization contribution to split learning for LLM fine-tuning, but the headline 70.8%/53.1% numbers are not supported as stated because the model treats devices as independent when they share one server GPU and one wireless channel.\n\nWhat's new: the CARD algorithm jointly selects the cut layer and server GPU frequency for LoRA-based split fine-tuning. The closed-form frequency optimum in Eq. (16) follows from a correct convexity check, and the brute-force scan over cut layers is valid. For a single device, the problem is solved correctly, and the derivation is self-contained. This is a modest but honest contribution to edge-AI resource allocation.\n\nThe main soft spot is the per-device independence assumption. A single physical server cannot run at different frequencies for different devices at the same time, and if devices are time-shared, later devices inherit queueing delay. The stress-test note is right: Eq. (10) has no contention term, so the reported delay reductions are computed under a model that may be infeasible for the described single-server deployment. This is a correctness risk for the headline claims, not a missing optional extension. The paper should either switch to a time-division framing where total system time is the sum of per-device slots, or add a contention-aware delay model.\n\nTwo smaller issues. There is no static-cut baseline, so it's unclear how much of the gain comes from adapting the cut layer rather than simply tuning the server frequency. And no fine-tuning quality metric is reported — no loss or accuracy — so we don't know whether the chosen split preserves model quality. For a letters paper, one of these alone is acceptable; together they weaken the evaluation.\n\nProportionate bottom line: the optimization core is sound, the writing is clear, and the problem is worth solving. The paper deserves a serious referee. I would send it back for a revision addressing the contention issue before treating the efficiency claims as robust, and I wouldn't cite it until then. I'd bring it to a reading group; the single-GPU frequency issue is a good prompt for discussing what modeling abstractions are acceptable in a short-format venue.","headline":"Clean per-device optimization, but the headline efficiency numbers ignore server contention and are not supported as stated.","tokens_in":9531,"tokens_out":3262,"would_cite":false,"duration_ms":30031,"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":"Adaptive split learning can cut LLM fine-tuning delay by 70.8 percent and server energy by 53.1 percent versus separate baselines.","keywords":["cut layer selection","fine-tuning","large language models","edge networks","split learning","LoRA","energy efficiency","resource allocation"],"falsifier":"Run two or more edge devices through the same split fine-tuning workload on one server, all with the same optimal cut layer, and measure the per-round wall-clock time as additional devices start training at the same moment. If the per-round delay rises with the number of concurrent devices, due to server GPU serialization or uplink contention, the independence-based delay model in Eq. (10) is falsified, and the 70.8 percent saving should be re-derived with a shared-resource term.","tokens_in":8493,"feed_emoji":"⚡","tokens_out":7326,"duration_ms":62874,"temperature":0.7,"pith_summary":"This paper aims to show that split learning, which runs the first layers of a large language model on mobile devices and the remaining layers on an edge server, can fine-tune models on private edge data without the full device burden of federated learning or the full server energy bill of centralized training. The proposed controller, CARD, decides two things per device and per training round: where to cut the model and how fast the server GPU should run, using a weighted, normalized cost that trades delay against server energy. Its reported result is a 70.8 percent average training-delay reduction over the device-only baseline and a 53.1 percent server-energy reduction over the server-only baseline, on a five-device testbed with a 32-layer LLM. If correct, this gives edge operators a simple per-round knob for balancing latency and energy as channels and device capabilities change.","feed_headline":"Adaptive split learning cuts LLM fine-tuning delay by 70.8%","feed_subtitle":"CARD chooses each device's cut layer and server GPU speed every round, saving server energy by 53.1 percent.","key_machinery":"The load-bearing object is the CARD algorithm, which decomposes a mixed-integer nonlinear program into two nested decisions. Given a cut layer $c_{m,n}$, the upper layer solves a convex problem in the server GPU frequency $f^S_{m,n}$, whose optimal value is the closed-form cubic root $Q = \\sqrt[3]{w(E_{\\max}-E_{\\min})/(2\\xi(1-w)(D_{\\max}-D_{\\min}))}$ clipped to the feasible frequency interval. The lower layer then brute-forces the cut layer over the $I$ transformer layers, since the transmitted-data-size functions make the cost non-convex, giving $O(I)$ complexity per device per round. The objective $U$ is a weighted sum of min-max normalized delay and energy, so CARD is explicitly selecting points on a weakly Pareto-optimal trade-off curve between latency and server power.","core_discovery":"The central claim is that jointly adapting the cut layer and the server's GPU frequency on a per-device, per-round basis makes LoRA-based split fine-tuning of LLMs substantially cheaper in both time and server energy than static splits. The paper derives a per-round delay model (device and server computation plus smashed-data, gradient, and LoRA-adapter transmission) and a server energy model cubic in GPU frequency, then minimizes a weakly Pareto-optimal weighted sum of normalized delay and energy. For a fixed cut layer the frequency subproblem is convex and yields a closed-form cubic-root solution; over the finite number of transformer layers the cut layer is found by exhaustive search. Simulations on five heterogeneous edge devices show the optimal cut is always all layers or none, with weaker devices offloading more to the server, and report the 70.8 percent and 53.1 percent savings against the two baselines.","pith_inferences":["If multiple devices share one edge server in the same round, the delay model needs a queueing or serialization term for the server GPU and uplink; without it, the reported 70.8 percent figure is best read as a per-device bound rather than a system-level guarantee.","The all-or-nothing structure of the optimal cut suggests the cost may be monotone in the cut layer when every transformer layer has equal FLOPs and equal smashed-data size, which would let a threshold rule replace the exhaustive scan.","The same two-level CARD machinery transfers to split inference or to other parameter-efficient fine-tuning methods, provided per-layer computation and transmitted-data sizes are known.","A direct stress test is to increase the number of participating devices while fixing total data volume; the model predicts unchanged per-round delay, whereas contention in a real deployment would make delay grow."],"forward_implications":["A per-device, per-round adaptation rule lets the system track wireless channel changes without solving a global joint problem.","Weak devices will systematically offload the entire transformer stack, and the server will assign them higher GPU frequencies, reducing the impact of device heterogeneity on fine-tuning time.","The weighting factor $w$ gives an operator a direct knob: raise it to favor low delay, lower it to favor low server energy, without changing the algorithm.","On a 32-layer model the optimal cut collapses to layer 0 or layer 32, so the practical control decision becomes a simple all-or-nothing offload rule in the simulated regime."],"supporting_citations":[{"why":"Introduces split learning and its resource-management problem, motivating the SL framework.","marker":"[8]"},{"why":"Supplies the device-edge cooperative fine-tuning setting this paper builds on.","marker":"[9]"},{"why":"Provides the FLOP-based computation delay model used in Eqs. (7)-(8).","marker":"[10]"},{"why":"Establishes the split LoRA fine-tuning precedent with device/server model partition.","marker":"[11]"},{"why":"Defines the SNR-to-MCS/CQI mapping used to compute transmission rates.","marker":"[12]"},{"why":"Supplies the weakly Pareto-optimal cost formulation used in Eq. (12).","marker":"[13]"},{"why":"Underlies the 32-layer LLaMA model used in the simulations.","marker":"[14]"}],"fun_headline_variants":["Adaptive split learning cuts LLM fine-tune delay by 70.8%","Per-device split learning slashes server energy by 53.1%","CARD adapts cut layer and GPU frequency to cut cost","Cut-layer and GPU tuning saves 70.8% time in LLM edge fine-tuning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a device's per-round training delay is only its own local computation plus its own transmission time, with no queueing or contention for the shared server GPU or uplink; if several devices train simultaneously, that independence breaks and the reported delay reduction no longer follows directly.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive split learning cuts LLM fine-tune delay by 70.8%","Per-device split learning slashes server energy by 53.1%","CARD adapts cut layer and GPU frequency to cut cost","Cut-layer and GPU tuning saves 70.8% time in LLM edge fine-tuning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001349,"raw_usage":{"total_tokens":5427,"prompt_tokens":841,"completion_tokens":4586,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":457,"completion_tokens_details":{"reasoning_tokens":4501}},"tokens_in":457,"tokens_out":4586,"duration_ms":27579,"temperature":1.0,"reasoning_tokens":4501,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:19:33.166228+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run two or more edge devices through the same split fine-tuning workload on one server, all with the same optimal cut layer, and measure the per-round wall-clock time as additional devices start training at the same moment. If the per-round delay rises with the number of concurrent devices, due to server GPU serialization or uplink contention, the independence-based delay model in Eq. (10) is falsified, and the 70.8 percent saving should be re-derived with a shared-resource term.","supporting_citations":[{"cited_title":"Split learning over wireless networks: Parallel design an d resource management,","cited_arxiv_id":null,"evidence_quote":"Introduces split learning and its resource-management problem, motivating the SL framework."},{"cited_title":"Device-edge cooperative fin e-tuning of foundation models as a 6G service,","cited_arxiv_id":null,"evidence_quote":"Supplies the device-edge cooperative fine-tuning setting this paper builds on."},{"cited_title":"Resource allocation for stable LLM t raining in mobile edge computing,","cited_arxiv_id":null,"evidence_quote":"Provides the FLOP-based computation delay model used in Eqs. (7)-(8)."},{"cited_title":"Federated Fine-Tuning for Pre-Trained Foundation Models Over Wireless Networks","cited_arxiv_id":"2407.02924","evidence_quote":"Establishes the split LoRA fine-tuning precedent with device/server model partition."},{"cited_title":"NR; Physical layer procedures for data,","cited_arxiv_id":null,"evidence_quote":"Defines the SNR-to-MCS/CQI mapping used to compute transmission rates."},{"cited_title":"D elay-aware microservice coordination in mobile edge computing: A rein forcement learning approach,","cited_arxiv_id":null,"evidence_quote":"Supplies the weakly Pareto-optimal cost formulation used in Eq. (12)."}],"review_version":1}