{"id":"05f4adc6-d4ec-4186-964b-362c4b811d1f","arxiv_id":"2501.09237","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A split fine-tuning scheme for wireless networks splits an LLM between devices and an edge server, compresses exchanged activations, and jointly optimizes compression, layer split, and bandwidth, cutting simulated fine-tuning delay by up to 80.2% and communication by 93.6%.","lead":"This paper proposes split fine-tuning, where a large AI model is split between mobile devices and an edge server, with compressed communication and joint optimization of the split point, compression, and bandwidth. In simulations on image tasks, it cuts fine-tuning delay by up to 80.2% and communication overhead by 93.6% versus baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Delay model uses max over devices while the text says server-side LoRA FP is sequential; the headline delay gains likely rest on this inconsistency.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing issue and the same likely consequence: Eq. (19) takes a max over per-device delays even though Section IV-A and Algorithm 1 describe a single shared server executing devices sequentially. This is the most direct threat to the central quantitative claim because the paper's headline delay reductions are produced by this delay model. The concern is concrete and testable, and the reader's conditional verdict already reflects the need for a fix. I see no additional objection that is more load-bearing; other issues, such as the fitted accuracy surrogate and absent code, are secondary and also do not by themselves overturn the qualitative contribution. Therefore the recommended verdict remains conditional, with no adjustment to the reader's assessment.","tokens_in":18413,"tokens_out":2331,"duration_ms":26797,"concrete_test":"Re-derive Section V-A with a serial server: for each device, express the round delay as local FP plus uplink plus cumulative server processing up to that device plus downlink plus local update, then take the maximum over devices. Rerun the Section VIII experiments, e.g., CIFAR-100 non-IID, with this corrected delay. If the reported 178.8 minutes and the 80.2% reduction change materially, the central claim is overstated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-A states that the server 'will perform the LoRA FP sequentially according to the order in which the immediate activation is received,' and Algorithm 1 loops over devices before the server executes. Yet the round delay in Eq. (19) is tau^t = max_n tau^t_n, which treats each device's server-side computation as fully overlapped. With one shared server-side model, the server-side FP/BP work must be serialized: a round's delay contains a sum over devices of tau_SC(n), not a maximum, and the per-device local and communication delays cannot all hide behind the serialized server. Since the reported 80.2% delay reduction in Section VIII is computed from this max model, the quantitative central claim is not supported unless the server is actually parallel or the delay expression is revised. This is an internal inconsistency, not merely a disagreement with field consensus, and it directly controls the headline number.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Split Fine-Tuning (SFT), a scheme for fine-tuning large language models in wireless networks by splitting the model between mobile devices and an edge server, using LoRA adapters, activation compression (sparsification, stochastic quantization, and lossless encoding), and a two-timescale resource-allocation algorithm that jointly optimizes compression rates, transformer-block allocation, and bandwidth. The authors claim that SFT reduces fine-tuning delay by up to 80.2% and communication overhead by up to 93.6% compared with federated-learning and split-learning baselines while satisfying device memory and accuracy constraints. The manuscript includes a system model, a delay/memory/communication analysis, a formulated mixed-integer optimization problem, an augmented-Lagrangian/SQP solution approach, and simulations on CIFAR-100 and Tiny-ImageNet under IID and non-IID settings.","tokens_in":18624,"tokens_out":3139,"duration_ms":33692,"significance":"If the central claims hold, the paper would make a useful contribution to edge-side LLM fine-tuning by addressing two practical bottlenecks: device memory and communication cost. The combination of split learning, LoRA, and joint compression/resource optimization is sensible and the evaluation covers several relevant baselines and data distributions. The paper gives credit for a fairly complete system model and a detailed memory-consumption analysis, and it ships a concrete algorithmic framework. However, the main delay-reduction claim is currently supported by a delay model that appears internally inconsistent with the described sequential server execution, and the accuracy feasibility claim rests on fitted surrogate models. These issues are load-bearing for the headline results, so the manuscript needs substantive revision before the claims can be accepted.","major_comments":[{"comment":"The round-delay model in Eq. (19) takes τ^t = max_n τ^t_n, which treats the server-side computation for all devices as fully overlapped. This contradicts Section IV-A, which states that the server 'will perform the LoRA FP sequentially according to the order in which the immediate activation is received,' and Algorithm 1 (lines 5–18) shows a per-device loop in which the server executes FP/BP before moving to the next device. With a single shared server-side model, the server-side work must be serialized, so the round delay should include a sum over devices of τ_SC(n) (or at least the serialized server component), not a pure maximum. Because the reported 80.2% delay reduction in Section VIII is computed from the max model, this inconsistency directly affects the central quantitative claim. The authors should either justify how the server-side computation is actually parallelized (e.g., separate server replicas or pipelining) or revise the delay model and rerun the simulations.","section":"§IV-A and §V-A, Eq. (19)"},{"comment":"The accuracy constraint in Eq. (27b) uses A(ρ, E), which is a third-order polynomial fitted to offline data in Section IV-C (Fig. 3), and the compression-strategy mapping in Section IV-B is a piecewise linear function fitted to offline data. Thus the 'optimal' compression rates and the feasibility of the accuracy constraint are determined by fitted surrogate models, not by actual measured fine-tuning accuracy. If the delay results in Section VIII enforce Eq. (27b) through these surrogates, then the claim 'while satisfying ... model accuracy constraints' is not established for the true model. The authors should validate the optimized (ρ*, E*, l*) on actual fine-tuning runs and provide sensitivity analysis with respect to the surrogate fitting error; reporting the fit MSE is not sufficient.","section":"§IV-C and §VI, Eq. (27b)"}],"minor_comments":[{"comment":"Eq. (12) writes τ^t_CC(l) on the left-hand side but the right-hand side depends on device n via f_n, C^u_n, and D^u_n; the device index should appear in the left-hand side's notation.","section":"§V-A, Eq. (12)"},{"comment":"Eq. (14) uses the notation f^s_n for the server-side frequency, but the server is a single shared resource; please clarify whether the server frequency is device-dependent or whether f^s_n denotes something else, and update the notation consistently.","section":"§V-A, Eq. (14)"},{"comment":"The delay expressions in Eqs. (13) and (15) use β as the compression ratio, while the sparsification rate is defined as ρ in Eq. (10); the relationship between β, ρ, the quantization level E, and the lossless encoding should be stated explicitly to avoid ambiguity.","section":"§IV-B and §V-A, Eqs. (10), (13), (15)"},{"comment":"The constraint enumeration contains typos: 'for q = 1' should be 'p = 1', and the lower-bound constraint on l is listed twice; please correct the indexing and remove the duplicate.","section":"Algorithm 2"},{"comment":"Table II lists the system bandwidth B as 30 MHz, but Section VIII-A states that 'We consider a system bandwidth of 5 MHz' and Figure 9 discusses a 'constrained 5 MHz bandwidth'; please align these values and clarify which bandwidth is used for the reported results.","section":"§VIII-A and Table II"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is relevant to the journal's scope and the authors have made a serious effort in system modeling and evaluation. The main risk is not the novelty of the architecture but the validity of the delay model and the reliance on fitted surrogates for the accuracy constraint. If the authors can resolve the sequential/parallel server inconsistency and validate the accuracy constraint on actual training runs, the paper would likely become acceptable. I would recommend requiring these revisions before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper integrates parallel split learning with LoRA and joint compression for LLM fine-tuning in wireless networks. The actual contribution is the specific combination plus the two-timescale resource allocation, which is legitimate systems work, though none of the components are new on their own.\n\nWhat it does well: the memory analysis is careful and correctly shows why LoRA alone doesn't solve the memory problem (activations dominate). The compression pipeline, while standard, is evaluated in a reasonable way, and the simulations cover IID and non-IID settings across two datasets, showing large communication savings. The figures on per-round delay under different bandwidth allocations are also informative.\n\nThe main soft spot is the delay model. Equation (19) defines the round delay as max over devices, but Section IV-A and Algorithm 1 describe the server running LoRA forward/backward sequentially for each device using one shared server-side model. Under that description, the server-side delay should be a sum over devices, not a max. The headline 80.2% delay reduction is computed from the max model, so it is likely overstated. This is not a minor presentational issue; it directly controls the quantitative central claim. The authors need to either justify parallel server-side execution (e.g., multiple workers) or revise the delay expression and re-run the simulations.\n\nA second concern is that the accuracy constraint A(rho,E) is a fitted third-order polynomial, and the compression strategy mapping is also a piecewise linear fit to offline data. This means the 'optimal' compression rates and the stated accuracy feasibility are only as good as those fits. The paper reports an MSE below 0.26% but no confidence intervals or sensitivity analysis, and the simulations lack error bars across runs. That limits how strongly the accuracy and delay results can be stated.\n\nMinor issues: reference [1] and [23] are the same paper, and there are notation inconsistencies (e.g., tau_SC is written as a function of f_s^n but the text uses f_s; Table II has parameters that don't match the narrative). No code or data is provided, which makes verification harder.\n\nOverall, this is a solid systems paper for the edge-AI community, not a breakthrough. The core idea is sound and the memory/communication analysis is valuable. The delay model inconsistency needs to be resolved before the quantitative claims can be accepted. I would send it to serious peer review with a request for major revision, and I'd want the revised version to address the serial/parallel gap and provide more robust evidence for the fitted surrogates.","headline":"A plausible wireless split-fine-tuning system with a genuine integration contribution, but the headline delay numbers rest on a serial/parallel inconsistency in the delay model.","tokens_in":19100,"tokens_out":2353,"would_cite":false,"duration_ms":25166,"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":"This paper claims that split fine-tuning, combining parallel split learning with LoRA and three-stage activation compression, can fine-tune large language models on memory-constrained mobile devices in wireless networks, cutting…","keywords":["split learning","LoRA","large language model fine-tuning","activation compression","wireless edge computing","resource allocation","mixed-integer optimization","split fine-tuning"],"falsifier":"In the reported eight-device setup, log the server's GPU busy time per fine-tuning round. If the server processes devices sequentially, the total server time will grow linearly with the number of devices; recomputing the round delay with server delay summed over devices rather than maximized would show whether the 80.2% reduction still holds.","tokens_in":18230,"feed_emoji":"📡","tokens_out":9904,"duration_ms":88367,"temperature":0.7,"pith_summary":"This paper proposes Split Fine-Tuning (SFT), a scheme for fine-tuning large language models on mobile devices that cannot hold the full model in memory. The model is cut at a chosen transformer block; the device runs the lower blocks with local LoRA adapters, the edge server runs the upper blocks on one shared model, and compressed activations travel over the wireless link. The paper claims this combination of parallel split learning, a three-stage compression pipeline (sparsification, stochastic quantization, lossless encoding), and two-timescale resource allocation cuts fine-tuning delay by up to 80.2% and communication overhead by 93.6% relative to federated and split-learning baselines, while keeping device memory and accuracy within constraints. If true, personalized fine-tuning of LLMs becomes feasible on memory-constrained mobile hardware in wireless networks.","feed_headline":"Split fine-tuning cuts LLM fine-tuning delay by up to 80%","feed_subtitle":"Accuracy stays within 2% of baseline, device memory fits mobile GPUs, and wireless overhead drops 93.6%.","key_machinery":"The load-bearing mechanism is the transformer cut: the $L$ transformer blocks are split at layer $l$, giving each device a small device-side model (embedding plus $l$ blocks with LoRA) and leaving the remaining $L-l$ blocks on the server, so devices send only compressed intermediate activations instead of model updates. The round-delay model $\\tau^t = \\max_{n \\in \\mathcal{N}} \\tau_n^t$ sums seven phases per device (distribution, device compute, activation upload, server compute, gradient download, device update, LoRA upload) and then takes the maximum over devices, which is what the two-timescale optimizer minimizes. The optimizer separates variables by timescale: the augmented Lagrangian method chooses sparsification rate $\\rho$, quantization level $E$, and cut layer $l$ in the large timescale, while sequential quadratic programming allocates bandwidth $b_n$ in the small timescale. A third-order fitted accuracy function $A(\\rho, E)$ acts as the constraint that ties compression choices to fine-tuning quality.","core_discovery":"The central claim is that fine-tuning an LLM in a wireless network does not require the full model on any device: splitting the transformer at layer $l$, keeping only $l$ blocks plus LoRA adapters on the device, and letting all devices share one server-side model reduces device memory by about 58.2% relative to federated LoRA training while preserving accuracy. The paper further claims that compressing the intermediate activations with Top-K sparsification, stochastic quantization, and run-length lossless encoding keeps accuracy loss within 2% at up to 80% sparsity and 3-bit quantization, while shrinking transmitted data to roughly 6.8% of the original. The delay minimization chooses sparsification rate, quantization level, cut layer, and per-device bandwidth; the reported result is that this joint optimization reduces fine-tuning delay by up to 80.2% and communication overhead by 93.6% compared with the baselines.","pith_inferences":["A natural extension is to apply the same parallel-split structure to decoder-only LLMs, where the compressed hidden states are token sequences rather than patch embeddings; the delay model would need an autoregressive generation factor.","The offline third-order accuracy fit could be replaced by an online accuracy estimator, allowing the optimizer to handle datasets and tasks not seen during fitting.","The downlink gradient transmission could be compressed with the same sparsification-quantization-encoding pipeline, which would shrink the gradient-transmission term in the round-delay sum beyond what the paper reports."],"forward_implications":["Devices with roughly 4-8 GB of memory can participate in LLM fine-tuning by holding only a few transformer blocks and their LoRA adapters, while the server holds the rest.","The three-stage compression scheme can cut activation uploads to about 6.8% of their original size with accuracy loss kept within 2%.","Allocating bandwidth and choosing the cut layer jointly at different timescales reduces per-round communication delay by up to 53.1% even at a 5 MHz system bandwidth.","Total fine-tuning delay drops by up to 80.2% and communication overhead by 93.6% versus federated fine-tuning and split-learning baselines."],"supporting_citations":[{"why":"Supplies the ViT-base architecture and patch/token dimensions used in the simulations and in the activation-volume estimate.","marker":"[15]"},{"why":"Defines the LoRA initialization scheme (zero matrix B, Gaussian matrix A) that the SFT fine-tuning procedure adopts.","marker":"[33]"},{"why":"Provides the activation-memory estimation formula used to derive the device-side memory constraint.","marker":"[38]"},{"why":"Serves as the federated-learning baseline (FL-based FT) for the delay and overhead comparisons.","marker":"[41]"},{"why":"Serves as the split-learning baseline (SL-based FT) in the performance comparison.","marker":"[42]"},{"why":"Supplies the run-length encoding used for lossless compression of the sparse activation mask.","marker":"[36]"},{"why":"Supplies the third-order fitting method used to model fine-tuning accuracy as a function of sparsification rate and quantization level.","marker":"[37]"},{"why":"Motivates the sparsification-quantization-encoding pipeline applied to intermediate activations.","marker":"[34]"}],"fun_headline_variants":["Split fine-tuning slashes LLM delay by 80% in wireless nets","Split LLM fine-tuning cuts delay 80%, overhead 93.6%","Edge-server split fine-tuning fits LLMs on mobile, cuts time 80%","Wireless split fine-tuning: 80% delay drop, 93.6% comm cut","Split fine-tuning: 80% faster LLM training in wireless nets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The delay model takes the maximum over devices, which assumes the server can process all devices' tasks at the same time, but the paper also states the server runs those tasks in the order activations arrive; if the server-side work is serial rather than parallel, the reported delay reductions would be smaller.","fun_headline_variants_meta":{"raw":{"variants":["Split fine-tuning slashes LLM delay by 80% in wireless nets","Split LLM fine-tuning cuts delay 80%, overhead 93.6%","Edge-server split fine-tuning fits LLMs on mobile, cuts time 80%","Wireless split fine-tuning: 80% delay drop, 93.6% comm cut","Split fine-tuning: 80% faster LLM training in wireless nets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001047,"raw_usage":{"total_tokens":4445,"prompt_tokens":1032,"completion_tokens":3413,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":648,"completion_tokens_details":{"reasoning_tokens":3307}},"tokens_in":648,"tokens_out":3413,"duration_ms":24311,"temperature":1.0,"reasoning_tokens":3307,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:08:32.843845+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the reported eight-device setup, log the server's GPU busy time per fine-tuning round. If the server processes devices sequentially, the total server time will grow linearly with the number of devices; recomputing the round delay with server delay summed over devices rather than maximized would show whether the 80.2% reduction still holds.","supporting_citations":[{"cited_title":"Resource allocation for stable LLM training in mobile edge computing,","cited_arxiv_id":null,"evidence_quote":"Provides the activation-memory estimation formula used to derive the device-side memory constraint."},{"cited_title":"Distributed learning of deep neural network over multiple agents,","cited_arxiv_id":null,"evidence_quote":"Serves as the split-learning baseline (SL-based FT) in the performance comparison."},{"cited_title":"Run-length encodings (corresp.),","cited_arxiv_id":null,"evidence_quote":"Supplies the run-length encoding used for lossless compression of the sparse activation mask."},{"cited_title":"Adap- tive digital twin-assisted 3C management for QoE-driven MSVS: A GAI-based DRL approach,","cited_arxiv_id":null,"evidence_quote":"Supplies the third-order fitting method used to model fine-tuning accuracy as a function of sparsification rate and quantization level."},{"cited_title":"Fast: Fidelity-adjustable semantic transmission over heterogeneous wireless networks,","cited_arxiv_id":null,"evidence_quote":"Motivates the sparsification-quantization-encoding pipeline applied to intermediate activations."}],"review_version":1}