{"id":"660a367f-81c1-4f6f-aef2-01aaeb081103","arxiv_id":"2411.15781","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Multi-user offloading of personalized diffusion models is formulated as an extended GQAP and solved by a DRL-convex hybrid with linear decision complexity.","lead":"This paper proposes a way for edge servers to run the first part of personalized image generation for many users at once using one shared model, while each user's own device finishes with their personalized model. It formulates the decisions of which users to offload and where to split as a difficult combinatorial problem, then solves it with a reinforcement learning and convex optimization hybrid.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The P1 objective uses an additive per-user reward Σα_iF(n_i), but F is fitted to the cluster-level PAI in Eq. (1), whose LPIPS term is pairwise; the optimizer may not optimize the stated metric.","rationale":"The reader's concern about representativeness of the fitted sigmoid is valid, but it presupposes that F(n_i) is the correct functional form for the objective. The more load-bearing issue is that the objective in P1 is not the metric defined in Eq. (1): PAI is cluster-level and has a pairwise LPIPS term, while P1 uses an additive per-user reward. Even an exactly correct per-user sigmoid would not fix the mismatch. This undermines the central claim because the 'optimal' split points and the DRL reward optimize a different objective than the one stated. The B&B baseline omission is also a serious comparison issue, but it is secondary. The paper's own limitations section does not mention this modeling gap, so it is an unacknowledged risk. I recommend keeping the reader's CONDITIONAL verdict, with the condition extended to require resolving the PAI objective inconsistency or clearly redefining F as a per-user surrogate and justifying the decomposition.","tokens_in":33871,"tokens_out":11425,"duration_ms":108485,"concrete_test":"Use the released models and prompts to compute, for 20 random problem instances, the true objective value with the actual PAI from Eq. (1) for the DRL-convex solution and for two alternative feasible solutions (e.g., all-local and all-offload with N̂). Compare the ranking against the ranking by Σα_iF(n_i) − L_i. If the surrogate ranking disagrees on any instance, the optimization target is not the stated metric. A cheap analytical check is also definitive: for two users with split points n_a ≠ n_b, LPIPS(m_a, m_b) depends on both n_a and n_b, so no univariate F(n_i) can represent the PAI contribution of user i alone.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. 3.4.2 defines PAI = κ·E_i[CLIP_i]·σ_P(LPIPS) with LPIPS = E_{i<i'}[LPIPS(m_i,m_i')], a cluster-level, pairwise quantity. Sec. 4 then sets the QoE objective as Σ_i α_i F(n_i) − L_i, and Sec. 5.2 solves P2 for each granted user separately using the same F(n_i). F is the sigmoid fitted in Sec. 6.1 to mean PAI values measured when all users share one split point (Fig. 7). Two problems follow: (i) if F is the cluster-level PAI, then Σ_i α_i F(n_i) is not the same metric; (ii) if F is meant per-user, the paper does not derive it from Eq. (1), and the LPIPS term depends on all users' split points, not just n_i. Since the hybrid inference in Fig. 2 allows n_i* to differ across users, the pairwise LPIPS term is not captured by a univariate sigmoid. The DRL reward (21) and the Fig. 11 evaluations inherit this misspecification, so the claimed optimality is with respect to the surrogate, not the stated PAI.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an edge-assisted offloading framework for personalized diffusion models. A cluster-wide model at the edge performs the initial denoising steps for multiple users in batches, while each user's local personalized model completes the remaining steps, with a per-user split point. The joint problem of selecting which requests to grant and choosing split points is formulated as an extended Generalized Quadratic Assignment Problem, proven NP-hard, and solved by a DRL-convex hybrid: DQN with prioritized experience replay decides grant/deny sequentially, while convex optimization determines each granted user's split point. The paper reports O(I) decision complexity and simulation improvements over baselines, and provides code, data, and models.","tokens_in":34194,"tokens_out":10444,"duration_ms":91071,"significance":"The system model is timely, and the decomposition into an outer DRL problem and an inner convex split-point selection is an interesting way to make a hard combinatorial problem tractable at scale. The NP-hardness reduction and the O(I) inference complexity are useful contributions, as is the public release of code and fine-tuned models. However, the current evidence does not establish optimality with respect to the stated Personalized Accuracy Index: the same sigmoid F(n*) fitted from a three-user dog experiment drives the split-point optimizer, the DRL reward, and the evaluation, and the additive per-user objective is not derived from the pairwise LPIPS-based PAI. With a corrected and externally validated accuracy model, the framework could be a solid contribution; in its present form the quantitative claims are only about a fitted surrogate.","major_comments":[{"comment":"The objective is not consistent with the defined metric. Eq. (1) defines PAI as κ·E_i[CLIP]·σ_P(LPIPS) with LPIPS = E_{i<i'}[LPIPS(m_i,m_i')], a cluster-level pairwise quantity. Yet (P1) maximizes Σ_i α_i F(n*_i) − L_i, and (P2) optimizes each granted user's n*_i separately using the same univariate sigmoid F. Section 6.1 fits F to mean PAI values from experiments in which all users share one split point (Fig. 7). If F is the cluster-level PAI, the additive sum is not the same metric; if F is intended to be per-user, it is not derived from Eq. (1), and the pairwise LPIPS term depends on all users' split points, not only n_i. Since the hybrid inference in Fig. 2 allows different n*_i across users, the optimization and the reward (21) can maximize a function that differs from the stated PAI.","section":"§3.4.2, §4 (P1), §5.2 (P2)"},{"comment":"The reward function double-counts the PAI term. With γ=1, the cumulative return over an episode is Σ_t α_t F(n*_t) + [Σ_i α_i F(n*_i) − Σ_i L_i] = 2Σ_i α_i F(n*_i) − Σ_i L_i, whereas the (P1) objective is Σ_i α_i F(n*_i) − Σ_i L_i. The final-step reward as written includes all PAI terms on top of the immediate PAI rewards collected at earlier steps. Unless the final reward is meant to be the difference between the total objective and the sum of previous PAI rewards, the DRL agent maximizes a different functional from (P1); if it is meant to be that difference, the formula should be written accordingly and used in training and evaluation.","section":"Eq. (21)"},{"comment":"The fitted PAI curve is the single empirical pillar of the paper, but it comes from one small experiment: three users, one dog-personalization cluster, 10 prompts, 20 images per prompt, with no error bars, no confidence intervals, and no held-out validation. The same curve (aF=0.0413, bF=71.44) is used in the convex split-point optimizer (§5.2), the DRL reward (21), and every performance comparison in §6.3, and the threshold N̂=80 is also selected from the same data. As a result, the reported gains and optimality are with respect to this self-defined surrogate; the paper does not demonstrate that decisions optimized for this F remain near-optimal for the actual PAI of Eq. (1) on other clusters, prompt sets, or user numbers.","section":"§6.1, Fig. 7, §6.3"},{"comment":"The proof of concavity relies on the assertion that 'we can select an appropriate value of N̂ to ensure F(n*_i)>0.5' on [N̂,N]. In the paper, N̂=80 is chosen in §6.1 from visual inspection before or independently of the fit, and no condition on (aF,bF,N̂) is stated or verified beyond the single fitted curve. Since the closed-form split-point solution and the reward both depend on this concavity, the paper should state the required inequality as a condition and provide a sensitivity analysis over the fitted parameters.","section":"§5.2, Proposition 2"},{"comment":"The comparison against Branch & Bound is not apples-to-apples: the B&B-based ILP baseline omits split-point optimization 'for simplicity,' while the proposed DRL-convex method includes it. The observed superiority of the proposed method over B&B may therefore reflect the additional split-point optimization rather than the quality of the DRL request-handling policy. The B&B baseline should be given the same inner split-point optimizer, or all methods should fix n* identically, before drawing conclusions about optimality.","section":"§6.3, Fig. 11c, Table 2"}],"minor_comments":[{"comment":"The label 'P13N (LPIPS)' appears to be a rendering corruption; it should read something like 'P (LPIPS)'.","section":"Fig. 6"},{"comment":"Several figures contain '/uni0000...' glyph artifacts in axis labels and legends; the figures need to be regenerated with proper fonts.","section":"Figs. 7–11"},{"comment":"Typos: 'spilt point' in the constraint C3 description, 'the the similarity' in §3.4.2, and 'with with respect to user count' in §6.3 should be corrected.","section":"Throughout"},{"comment":"The symbol κ is used both as the multiplicative constant 3 in Eq. (1) and as the factor 0.05 in the α_i sampling interval in §6.2.1; please disambiguate these two uses.","section":"Eq. (1) and §6.2.1"},{"comment":"The sentence about increasing 'the exponent of the total latency term' is unclear and should be rewritten.","section":"§5.2, Remark 1"},{"comment":"The learning curve for the specific-application model is reported from a single seed; multiple seeds or confidence bands would make the convergence claim more robust.","section":"Fig. 9(c)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is likely to become acceptable after a substantive revision. The main risks are the surrogate-objective circularity and the reward mismatch; if the authors redefine the objective in terms of per-user PAI curves with proper validation, or reformulate to include pairwise LPIPS, and fix the reward, the contribution could be solid. The public code and models are a strength. Please also ensure that the comparison with Branch & Bound is made fair and that the single-experiment F fit is not the sole empirical basis for the claimed gains."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWhat should you know first: this is a genuinely new problem framing, and the algorithmic decomposition is sensible. The paper casts multi-user offloading of personalized diffusion models as an extended Generalized Quadratic Assignment Problem, with a shared cluster-wide model at the edge and per-user split points, then solves it with a DRL outer loop over grant/deny decisions and a convex inner loop for split points. That combination hasn't appeared in the cited literature, and the MDP transformation—cyclic shifting, reward backpropagation, local-global state design—is careful and mostly clear. The claimed O(I) decision complexity is plausible for an online solution.\n\nThe main contribution is the formulation, not the empirical validation. The experiments show the hybrid approach outperforms three baselines, but those baselines are weak: all offload with or without split-point optimization, or full local inference. The Branch & Bound comparison omits split-point optimization entirely, so the gap is partly built into the setup.\n\nThe real soft spot is the objective. The Personalized Accuracy Index in Eq. (1) is defined as a cluster-level, pairwise LPIPS metric, but the optimization objective in P1 is a sum of per-user terms α_i F(n_i), and F is fitted to mean PAI values from an experiment where all users share a single split point (Fig. 7). Once split points differ across users, the optimizer is not maximizing the stated cluster-level PAI. If F is meant to be per-user, it is not derived from Eq. (1). The fitted sigmoid also comes from one three-user dog experiment with no error bars, and N_hat = 80 is selected post hoc from the same data that produced the fit. The convexity proof in Section 5.2 additionally relies on F(n*) > 0.5 over the domain, which is only guaranteed by that choice of N_hat.\n\nSo the claimed optimality is relative to a self-defined surrogate, not an externally validated metric. That does not mean the framework is wrong—the latency model and the GQAP reduction are plausible, and the authors are honest about limitations such as the shared-size cluster model and unexplored clustering. But the evaluation needs to either fix the PAI definition so the objective matches the metric, or provide a per-user PAI model fitted under heterogeneous split points.\n\nFor a reader in edge computing or AIGC, this problem formulation is worth knowing. I would send it to a serious referee with the expectation of major revision focused on the PAI mismatch and stronger baselines. If the authors can show the objective matches a validated metric, this could become a solid contribution.","headline":"The DRL-convex decomposition is smart and the problem is new, but the objective rests on a fitted cluster-level PAI that may not be the metric actually optimized once split points diverge.","tokens_in":34691,"tokens_out":4775,"would_cite":false,"duration_ms":38395,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C27","90B80","68M20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that multi-user personalized diffusion inference can be offloaded in near real time by decomposing a joint offload-and-split decision into a DRL layer and a convex split-point layer.","keywords":["diffusion models","edge offloading","personalized content synthesis","generalized quadratic assignment problem","deep reinforcement learning","convex optimization","hybrid inference","batching"],"falsifier":"Measure PAI for a number of users and prompts (different subjects, breeds, styles, and devices) across split points from 80 to 200 and compare with the fitted curve $F(n^*) = 1/(1+\\exp(-0.0413(n^*-71.44)))$; if the measured curve shifts or flattens for any group, the convexity premise breaks and the claimed optimal split points would need re-derivation, while large across-user variation would falsify the single-curve assumption.","tokens_in":41,"feed_emoji":"⚡","tokens_out":5499,"duration_ms":103827,"temperature":0.7,"pith_summary":"The paper tries to establish that an edge server can decide, user by user, whether to offload part of a personalized diffusion model's denoising steps to a shared cluster-wide model, and how many steps to offload, in near real time. It models the joint decision as an extended Generalized Quadratic Assignment Problem, which is NP-hard, and then splits it into a deep-reinforcement-learning layer that grants or denies offload requests and a convex-optimization layer that sets each user's split point. The claim is that this decomposition recovers near-optimal system performance with decision complexity linear in the number of users, beating all-offload, fixed-split, and local-only baselines. A sympathetic reader would care because it offers a path to personalized text-to-image services on heterogeneous devices without replicating every user's model on the server.","feed_headline":"Solver picks offload and split points in linear time","feed_subtitle":"A DRL layer grants or denies offloads while a convex layer sets each user's personalization-latency split.","key_machinery":"The machinery is the Personalized Accuracy Index (PAI), a product of CLIP-based prompt-image similarity and a sigmoid-modulated LPIPS personalization score, summarized by the fitted function $F(n^*) = 1/(1+\\exp(-a_F(n^*-b_F)))$ with $a_F = 0.0413$ and $b_F = 71.44$, plus a linear model $L_e(b,G) = k_e(b/G) + h_e$ for the latency of one batched edge denoising step. These two empirical curves make the inner split-point problem concave, so each user's split point can be computed by a scalar equation root, and they feed the MDP reward that the DRL agent uses to learn grant/deny policies. The outer-inner nesting, with request handling dominant over split-point selection, is what converts an extended GQAP into a tractable decision sequence.","core_discovery":"On the paper's own terms, the central discovery is that the coupled choice of offload requests and split points for a cluster of users can be recast as a nested optimization: once the binary grant/deny decisions are fixed, each user's optimal split point is found by a concave maximization whose solution depends only on that user's device speed, preference weight, and the edge server's batch latency. Because the personalized accuracy index follows an S-shaped curve that the paper fits as a sigmoid with parameters aF=0.0413 and bF=71.44, the objective is strictly concave on the legal split range, so the inner problem has a unique optimum. The outer grant/deny problem is then treated as a sequential Markov decision process and solved with a DQN using prioritized experience replay, cyclic state shifting, and reward backpropagation. Simulations with three DreamBooth-finetuned personalized models and a cluster-wide model show the hybrid solver outperforming all-offload, fixed-split, local-only, branch-and-bound, and genetic-heuristic baselines in average objective value, with complexity O(I).","pith_inferences":["Beyond the paper, the same split-point optimizer would apply to any personalization task whose PAI curve is sigmoidal, so faces, styles, and other subjects might reuse the machinery after a one-time fit.","A testable extension is to replace the single fitted sigmoid with per-user or per-cluster PAI curves and check whether convexity still holds; the paper's claims would weaken if curves vary strongly across users.","The reward-backpropagation design suggests the DRL layer could be retrained online as the user population shifts, an adaptive variant the paper itself does not evaluate."],"forward_implications":["An edge server with a single trained Q-network can serve a varying number of users with heterogeneous local GPUs without re-solving the combinatorial problem at decision time.","The per-user split point becomes an online by-product of the grant/deny choice, so users receive individualized latency-accuracy trade-offs rather than a fixed offload depth.","The complexity comparison implies the scheme scales to user sets where branch-and-bound becomes prohibitive; the stated cost is O(I) against O(2^{2I}) for branch-and-bound.","Because only the cluster-wide model is stored at the edge, the approach directly addresses the storage burden of replicating each user's personalized model on the server."],"supporting_citations":[{"why":"Single-user edge-cloud split inference for Stable Diffusion that this work generalizes to many users with a shared cluster-wide model.","marker":"[13]"},{"why":"Collaborative distributed diffusion-based AIGC in wireless networks, establishing the split-inference setting and wireless transmission assumptions.","marker":"[18]"},{"why":"Supplies the linear batching latency model L(b)=kb+h and the multiuser edge inference resource-allocation formulation used in the latency expressions.","marker":"[21]"},{"why":"Used in the NP-hardness reduction to convert the linear objective term into the diagonal of the quadratic form, yielding a pure GQAP.","marker":"[33]"},{"why":"Source of the generalized assignment problem and its multiplier-adjustment method, cited as evidence that GQAP is NP-hard.","marker":"[34]"},{"why":"Classic NP-hardness reference for assignment-type problems, supporting the claim that the reduced problem is NP-hard.","marker":"[35]"},{"why":"DreamBooth is how the cluster-wide and per-user personalized models are fine-tuned in the experiments that produce the PAI curve.","marker":"[39]"},{"why":"DQN algorithm that the policy network builds on for the outer grant/deny decisions.","marker":"[36]"},{"why":"Prioritized experience replay used to stabilize and speed training of the DRL agent.","marker":"[38]"}],"fun_headline_variants":["Hybrid DRL + convex solver hits linear-time offloading","Offload and split decisions optimized in linear time","One-shot offloading: DRL and convex merge for diffusion","Multi-user diffusion offloading now linear-time","Efficient offloading: DRL gates, convex splits"],"cache_read_input_tokens":36864,"weakest_assumption_plain":"Everything rests on one fitted sigmoid curve, trained on three users' dog photos, standing in for every user's personalization-accuracy behavior; if that curve is wrong for a user or prompt, the \"optimal\" split points and rewards are optimal only for the curve, not for the real system.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid DRL + convex solver hits linear-time offloading","Offload and split decisions optimized in linear time","One-shot offloading: DRL and convex merge for diffusion","Multi-user diffusion offloading now linear-time","Efficient offloading: DRL gates, convex splits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00022,"raw_usage":{"total_tokens":1503,"prompt_tokens":1056,"completion_tokens":447,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":672,"completion_tokens_details":{"reasoning_tokens":369}},"tokens_in":672,"tokens_out":447,"duration_ms":4603,"temperature":1.0,"reasoning_tokens":369,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:54:06.898982+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure PAI for a number of users and prompts (different subjects, breeds, styles, and devices) across split points from 80 to 200 and compare with the fitted curve $F(n^*) = 1/(1+\\exp(-0.0413(n^*-71.44)))$; if the measured curve shifts or flattens for any group, the convexity premise breaks and the claimed optimal split points would need re-derivation, while large across-user variation would falsify the single-curve assumption.","supporting_citations":[{"cited_title":"Exploring collaborative distributed diffusion- based ai-generated content (aigc) in wireless networks,","cited_arxiv_id":null,"evidence_quote":"Collaborative distributed diffusion-based AIGC in wireless networks, establishing the split-inference setting and wireless transmission assumptions."},{"cited_title":"Some remarks on quadratic programming with 0-1 variables,","cited_arxiv_id":null,"evidence_quote":"Used in the NP-hardness reduction to convert the linear objective term into the diagonal of the quadratic form, yielding a pure GQAP."},{"cited_title":"A multiplier adjustment method for the generalized assignment problem,","cited_arxiv_id":null,"evidence_quote":"Source of the generalized assignment problem and its multiplier-adjustment method, cited as evidence that GQAP is NP-hard."},{"cited_title":"P-complete approximation problems,","cited_arxiv_id":null,"evidence_quote":"Classic NP-hardness reference for assignment-type problems, supporting the claim that the reduced problem is NP-hard."},{"cited_title":"Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation,","cited_arxiv_id":null,"evidence_quote":"DreamBooth is how the cluster-wide and per-user personalized models are fine-tuned in the experiments that produce the PAI curve."},{"cited_title":"Human-level control through deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"DQN algorithm that the policy network builds on for the outer grant/deny decisions."}],"review_version":1}