{"id":"934d550c-90fc-4f46-8e9e-30abd2624fdd","arxiv_id":"2504.20939","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A flexible semantic-aware resource allocation algorithm for multi-user semantic image transmission sets bandwidth, power, and compression rates so that each user's semantic similarity lands inside an acceptable range, increasing served users in simulations.","lead":"This paper proposes a wireless resource allocation scheme for semantic communication in which each user specifies a range of acceptable semantic similarity instead of a single threshold. The method tunes bandwidth, power, and compression rate to serve more users, and simulations report up to 17.1% more satisfied users than fixed-threshold baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1 never re-checks C6/C7 after the SNR update, so the reported satisfied-user counts may include infeasible allocations; the 17.1% gain has not been shown to come from the range constraint.","rationale":"Reader's weakest assumption focuses on convergence of the alternating procedure. I agree that is one gap, but the more load-bearing issue is that final feasibility is never checked even in the limit: the problem's own constraints C6 and C7 do not appear in the GP subproblem and are only handled heuristically in F2 at the SNR of that iteration. Even with a proof of convergence in objective value, the final point can violate C7 (when final SNR exceeds the SNR used to select ξ) or C6 (via the fallback). This makes the central numerical claim fragile and explains why the paper's Fig. 2 shows an unsatisfied user despite hard constraints in (4). I do not see evidence of dishonesty; the flaw is in the argument. The range-constraint idea is plausible, and the comparison to a fixed-threshold resource allocation is the right experiment, so the paper should not be rejected outright. The proposed concrete test will settle whether the concern is real: if all reported satisfied users pass C6/C7 after termination, the claim survives; if not, the claimed gains are not valid evidence. Because the concern is verifiable and the current evidence is insufficient, the reader's CONDITIONAL verdict remains appropriate; I would not move it without the feasibility check. My read agrees with the reader on the broad area of feasibility/convergence but identifies the unverified post-iteration constraints as a more specific mechanism.","tokens_in":8494,"tokens_out":9100,"duration_ms":108342,"concrete_test":"Modify the simulator to post-process every Algorithm 1 output: compute final SNRi = P_i h_i/(β_i N0), look up ξ_i(O_i, SNRi) in the UDeepSC similarity table, and compute t_i = d0,i(1−O_i)/(β_i log2(1+SNRi)). Count as satisfied only users with SNRi ≥ SNR_ith, ξ_i ∈ [ξ_i,thmin, ξ_i,thmax], and t_i ≤ τ_i,th. Rerun Fig. 1 across the bandwidth range. If the proposed method's satisfied-user count or the 17.1% improvement drops materially relative to the strict and QoE baselines, the central claim is an artifact of uncounted constraint violations.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 1's stopping test only looks at ΔF, not at feasibility of the final solution. In each F2 step, O_i and ξ_i are read from a similarity table for the current SNRi,it and are chosen to satisfy C7 and C6. Those values are then frozen while F1′ is solved as a GP over bandwidth and power with only C1–C4; the resulting SNR_i,it is recomputed and the loop repeats. No step verifies that, at termination, the actual ξ_i(O_i, SNR_i) lies in [ξ_i,thmin, ξ_i,thmax] or that t_i = d0,i(1−O_i)/(β_i log2(1+SNR_i)) ≤ τ_i,th. The explicit fallback (Algorithm 1, lines 15–21) selects ξ_i and O_i from the table at SNR_ith when no delay-feasible candidate exists, and the algorithm proceeds even if C6 is still violated; and if the GP solution lands on SNR_i > SNR_ith because of the β_i,min lower bound, the real ξ_i can exceed ξ_i,thmax even though the selected ξ_i was in range at the lower SNR. Hence the satisfied-user counts in Fig. 1 and the reported 17.1% improvement may include users whose allocations violate C6 or C7. The issue is compounded by (4): C4, C6 and C7 are stated as hard constraints for every user, yet the objective is a penalty sum and Fig. 2 explicitly leaves user 8 unsatisfied, showing that the algorithm treats satisfaction as a soft outcome without any admission-control variable. This means the empirical gain is not yet attributable to the range-constraint formulation; it may be an artifact of returning infeasible points.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a semantic-aware resource allocation scheme for uplink semantic communication, in which each user specifies a range for semantic similarity instead of a fixed minimum threshold. The authors formulate a mixed-integer nonlinear program (MINLP) that jointly optimizes bandwidth, power, and compression rate under SNR, bandwidth, power, delay, and semantic-similarity constraints, and decompose it into a geometric-programming subproblem over bandwidth and power and a lookup-table subproblem over the compression rate. An alternating algorithm is presented, and simulations with the UDeepSC model on CIFAR-10 report up to 17.1% more satisfied users and 14% higher total network semantic similarity compared with QoE-aware and strict-similarity baselines. The paper concludes that flexible similarity ranges enable the base station to serve more users at a slight average-similarity cost.","tokens_in":12,"tokens_out":6561,"duration_ms":127132,"significance":"If the reported gains were established for feasible solutions, the idea of replacing a point constraint on semantic similarity with an acceptable range would be a useful contribution to semantic-communication resource allocation, because it directly addresses the trade-off between user satisfaction and resource scarcity. The paper is also constructive in shipping a tractable decomposition: F1' is a correctly formed geometric program once each ξi is fixed, and the use of a measured lookup table from a pretrained UDeepSC model avoids requiring a closed-form semantic-similarity expression. The main limitation is that the alternating algorithm is not shown to return feasible points of the original MINLP, so the central numerical claim is currently unsupported; the approach is nevertheless promising and repairable within the manuscript's scope.","major_comments":[{"comment":"The algorithm selects ξi and Oi from the similarity table based on the current SNRi,it, solves F1′ over bandwidth and power, and then recomputes SNRi,it; the termination condition checks only ΔF. Because C6 and C7 are hard constraints in problem (4), and no step verifies that the actual ξi(Oi,SNRi) lies in [ξi,thmin, ξi,thmax] or that t_i = d0,i(1−Oi)/(βi log2(1+SNRi)) ≤ τi,th after the GP update, the returned allocation may violate C6 or C7. Consequently, the satisfied-user counts in Fig. 1 include users whose allocations are not known to be feasible, which undermines the reported 17.1% improvement.","section":"Section III, Algorithm 1 (lines 8–24)"},{"comment":"The fallback for a user with no delay-feasible candidate selects ξi and Oi from the table at SNRith. This fallback is triggered precisely when no candidate satisfying C6 was found, so the algorithm continues while fixing a pair that is known to violate C6; it then solves F1′ with that pair as a constant. In addition, if the GP solution raises SNRi above SNRith, the true ξi(Oi,SNRi) can exceed ξi,thmax even though the selected ξi satisfied C7 at the lower SNR. Problem (4) contains no admission-control or slack variable to accommodate such cases, so the algorithm may return an infeasible point as if it were a solution.","section":"Section III, Algorithm 1 (lines 15–21)"},{"comment":"No convergence or feasibility proof is provided for the alternating procedure. The stopping criterion ΔF measures only the penalty objective; because F1 and F2 optimize different variable sets, a monotone decrease of the objective between alternating updates is not established, and objective convergence would not imply convergence to a feasible point of the MINLP. The claim in Section III that Algorithm 1 reaches a suboptimal solution of (4) therefore requires a fixed-point or monotonicity argument that is absent.","section":"Section III, Eqs. (4)–(6)"},{"comment":"The objective in (4) is a smooth penalty, Σ (SNRith/SNRi)^a (ξi,thmin/ξi)^a, not a count of satisfied users. Minimizing this penalty does not generally maximize the number of users satisfying C4 and C7, since a user that remains infeasible still contributes a decreasing penalty as its ratios decrease, and no mechanism drops or repairs such a user. Figure 2 itself shows user 8 unsatisfied in the proposed method, which is consistent with this issue but conflicts with the paper's framing of the range constraint as a feasibility enhancement. The 17.1% gain in Fig. 1 is thus not yet attributable to the range-constraint formulation; it may be an artifact of returning points outside the feasible set.","section":"Section IV, Figs. 1–2"}],"minor_comments":[{"comment":"The summation index in C1 is written as n while the term is β_i; it should be Σ_{i=1}^N β_i ≤ M. Similar index inconsistency appears in the objective of (4)–(5), where 'NP n=1' should be a standard summation.","section":"Eq. (4)"},{"comment":"The phrase 'our method without similarity rang' in the paragraph after Eq. (8) should read 'range'.","section":"Section IV"},{"comment":"The transmission delay t_i is defined in prose as d0,i(1−Oi)/(βi log2(1+SNRi)), but C6 in (4) states only t_i ≤ τi,th; include the explicit expression in the constraint listing for self-containedness.","section":"Section II-A"},{"comment":"The text says user 8 was not served in the proposed method, but it is not stated whether an unsatisfied user receives zero resources or receives an allocation that simply lies outside its semantic-similarity range; making this convention explicit would help interpret the satisfied-user counts.","section":"Fig. 2 and Section IV"},{"comment":"The complexity expression O(it(kN + (7N + 3)^3)) should be justified: the GP has 2N decision variables, so the origin of the (7N+3)^3 term is unclear and should be derived or referenced.","section":"Section III, Complexity"}],"recommendation":"major_revision","confidential_remarks":"The conceptual contribution is interesting and within the journal's scope, but the current numerical claims depend on a feasibility guarantee that is not established. I recommend major revision, not rejection, because the algorithm can be repaired with a post-GP verification step and an admission-control variable; however, the experimental comparisons in Section IV must be re-run after such a repair, since the reported gains may shrink."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing to know: the paper introduces a genuine generalization—per-user semantic similarity interval [ξmin, ξmax] instead of a fixed floor—and shows a plausible way to trade a little similarity for more served users. The simulations report up to 17.1% more satisfied users than a QoE-based baseline. If the algorithm actually delivered feasible points, this would be a useful engineering result.\n\nWhat is new and good: the range constraint is a real departure from [1,4,5,9,11], and the decomposition into a GP over bandwidth/power plus a lookup-table step for compression rate is a workable recipe. The GP subproblem F1′ is correctly formed as a geometric program once ξ_i is fixed; C1′–C4′ are posynomial or monomial. The lookup table is measured from UDeepSC, not fabricated. The baseline comparison targets the right class.\n\nBut the stress-test concern lands. Algorithm 1 never re-checks C6 or C7 after the SNR update. It selects ξ_i and O_i for the current SNRi,it, solves the GP with those frozen, then recomputes SNR—and the loop stops on ΔF, not on feasibility. If the GP lands at a higher SNR than the table row used, the actual ξ_i can exceed ξ_i,thmax; the fallback (lines 15–21) picks a ξ_i at SNR_ith even when the delay constraint is still violated. So reported satisfied-user counts may include infeasible allocations. The 17.1% gain is not yet attributable to the range constraint specifically; it could be an artifact of counting users whose C6/C7 are violated.\n\nAlso, problem (4) claims to maximize the number of satisfied users, but the objective is a penalty sum over reliability ratios. No admission-control variable exists; Fig. 2 explicitly leaves user 8 unsatisfied while resources are still allocated. The paper treats satisfaction as a soft outcome, not as the stated hard objective. The core idea survives, but the evidence needs repair.\n\nFor whom: people working on semantic communication resource allocation, and 6G system designers who want a flexible-threshold formulation. The paper is worth a serious referee—it should go to review, not desk reject—but the referee will need to push for a fixed-point feasibility analysis, an explicit dropping or admission-control mechanism, error bars, and ideally released code.\n\nI wouldn't cite the numerical claims, but I'd cite the range-constraint idea.","headline":"Genuine range-constraint idea, but Algorithm 1 may count infeasible users; worth a serious referee.","tokens_in":9449,"tokens_out":2181,"would_cite":true,"duration_ms":22496,"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":"Relaxing a fixed semantic-similarity threshold into a per-user interval lets a base station serve up to 17.1% more users in a semantic-communication uplink.","keywords":["Semantic communication","Resource allocation","Semantic similarity","Compression rate","Geometric programming","User satisfaction","Uplink scheduling"],"falsifier":"Run Algorithm 1 on a random instance with the paper's simulation parameters and record the offered similarity $\\xi_i$ and transmission delay $t_i$ after each iteration; if any user's terminal allocation has $\\xi_i$ outside $[\\xi_{i,\\mathrm{thmin}}, \\xi_{i,\\mathrm{thmax}}]$ or $t_i > \\tau_{i,\\mathrm{th}}$, the claimed satisfaction gain is an artifact of the loop's unchecked fallback rather than a property of the interval formulation.","tokens_in":8269,"feed_emoji":"📶","tokens_out":7919,"duration_ms":67833,"temperature":0.7,"pith_summary":"Semantic communication promises to save spectrum by transmitting only the meaning of a message, but most resource-allocation schemes still treat each user's semantic quality as a fixed bar: fall below it and the user is unsatisfied. This paper argues that the fixed bar is the problem and replaces it with a per-user acceptable interval, bounded below by the user's minimum semantic similarity and above by the user's maximum willingness to spend computing and energy. The base station then jointly chooses bandwidth, power, and compression rate so that as many users as possible land inside their interval, and the simulations report up to 17.1% more satisfied users and 14% higher total network semantic similarity than a QoE-aware single-channel baseline. If the interval constraint is what enlarges the feasible region, then treating semantic quality as negotiable is a general lever for denser multi-user semantic networks.","feed_headline":"Relaxing semantic similarity bounds serves up to 17.1% more users","feed_subtitle":"An interval instead of a fixed similarity target lets the scheduler admit more users for the same bandwidth.","key_machinery":"The central object is the semantic-similarity range $[\\xi_{i,\\mathrm{thmin}}, \\xi_{i,\\mathrm{thmax}}]$ substituted for the conventional point constraint $\\xi_i \\ge \\xi_{i,\\mathrm{th}}$, together with a lookup table that encodes $\\xi_i(O_i,\\mathrm{SNR}_i)$ by running the UDeepSC encoder–decoder over a grid of compression rates and SNRs. The range turns the semantic requirement into a negotiating interval: the base station can reduce the offered similarity to a level still inside the user's acceptance zone and use the saved bandwidth and power to admit another user. The argument then moves through a two-step decomposition—F1, a geometric program over bandwidth $\\beta_i$ and power $P_i$ at fixed similarity values, and F2, a table search choosing the highest feasible similarity and its compression rate $O_i$ for the current SNR—iterated until the objective stops improving. The lookup table is load-bearing: every feasibility check of the delay constraint (C6) and the similarity constraint (C7) passes through its entries.","core_discovery":"The paper's claim is that the feasible region of a multi-user semantic uplink grows when the semantic-similarity constraint is an interval $[\\xi_{i,\\mathrm{thmin}}, \\xi_{i,\\mathrm{thmax}}]$ rather than a single lower bound. In problem (4), the base station minimizes a penalized product of transmission and semantic reliability ratios subject to total-bandwidth, per-user power, minimum-SNR, compression-rate, delay, and similarity-interval constraints. Because the semantic similarity $\\xi_i(O_i,\\mathrm{SNR}_i)$ has no closed form, the paper decomposes the MINLP into a geometric-programming subproblem for bandwidth and power (F1) and a lookup-table subproblem for compression rate (F2), solved alternately by Algorithm 1 until a stopping condition. Simulations using the UDeepSC encoder–decoder on CIFAR-10 show the algorithm serving up to 17.1% more users than the QoE-aware strict-similarity baseline and raising total network semantic similarity by 14%, at the price of a slightly lower average similarity per user because satisfied users are served near their minimum acceptable similarity rather than their maximum.","pith_inferences":["The interval-relaxation idea should transfer to other resource dimensions—user association, number of semantic symbols, task selection—wherever the quality metric is monotone in the resource; the paper demonstrates it only for bandwidth, power, and compression rate.","The 17.1% figure depends on the specific UDeepSC lookup table and the chosen interval widths; a natural testable extension is that widening the intervals keeps raising the served-user count until the delay or SNR constraints bind, a scaling law the paper does not state.","Because Algorithm 1 has no convergence proof, a practical failure mode is oscillation between F1 and F2 where a user is alternately admitted and dropped; checking for cycles on random channel realizations would tell whether the reported gains survive outside the simulated instances.","The fallback step that picks the best table entry at $\\mathrm{SNR}_{i,\\mathrm{th}}$ when no delay-feasible entry exists can return a similarity outside the user's interval; adding a feasibility check inside the loop would make the satisfaction guarantee conditional on the lookup table, a small change the paper leaves implicit."],"forward_implications":["Using a similarity interval instead of a fixed threshold means the same total bandwidth can satisfy a larger user population, because resources are spent only to move each user's similarity to the bottom of their interval, not to a prescribed point.","The strict-similarity variant of the scheme, which keeps a fixed lower bound, serves fewer users; the 17.1% gain is therefore specifically the work of the interval constraint, not of the solution algorithm alone.","Serving users near their minimum acceptable similarity intentionally lowers the network's average similarity, but the paper reports total network semantic similarity rising by 14% because more users are admitted.","The QoE-based single-channel baseline underperforms partly because it cannot split a user's traffic across channels, so its assigned-but-unsatisfied users waste spectrum that the flexible allocation can reuse.","Even with no closed-form expression for semantic similarity, a precomputed table of $\\xi(O,\\mathrm{SNR})$ supports a polynomial-complexity allocation loop, since the geometric-programming subproblem is convex and the table search is bounded."],"supporting_citations":[{"why":"QoE-based semantic-aware resource allocation for multi-task networks: the strict-similarity single-channel baseline the proposed method is compared against and outperforms by 17.1% in satisfied users.","marker":"[11]"},{"why":"Unified Multi-Task Semantic Communication System (UDeepSC): supplies the encoder–decoder model whose SNR/compression-rate-to-similarity lookup table is used inside subproblem F2.","marker":"[17]"},{"why":"Interior Point Polynomial Methods in Convex Programming: justifies the polynomial-time solution of the geometric-programming subproblem F1.","marker":"[18]"},{"why":"Geometric programming for communication systems: the framework showing F1 is a geometric program solvable by standard tools such as CVX/MOSEK.","marker":"[19]"}],"fun_headline_variants":["Semantic range constraints lift served users by 17.1%","Interval semantic bounds serve 17.1% more users","Flexible semantic ranges admit up to 17.1% extra users","Semantic range scheduling boosts user satisfaction 17.1%","Relaxed similarity intervals serve 17.1% more users"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The alternating loop that switches between choosing compression rates and allocating bandwidth and power is assumed to settle at a solution where every user still meets their delay and similarity requirements, but the paper does not prove convergence or show that the fallback selection preserves those requirements.","fun_headline_variants_meta":{"raw":{"variants":["Semantic range constraints lift served users by 17.1%","Interval semantic bounds serve 17.1% more users","Flexible semantic ranges admit up to 17.1% extra users","Semantic range scheduling boosts user satisfaction 17.1%","Relaxed similarity intervals serve 17.1% more users"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000591,"raw_usage":{"total_tokens":2793,"prompt_tokens":989,"completion_tokens":1804,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":605,"completion_tokens_details":{"reasoning_tokens":1715}},"tokens_in":605,"tokens_out":1804,"duration_ms":12339,"temperature":1.0,"reasoning_tokens":1715,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:15:36.387623+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 1 on a random instance with the paper's simulation parameters and record the offered similarity $\\xi_i$ and transmission delay $t_i$ after each iteration; if any user's terminal allocation has $\\xi_i$ outside $[\\xi_{i,\\mathrm{thmin}}, \\xi_{i,\\mathrm{thmax}}]$ or $t_i > \\tau_{i,\\mathrm{th}}$, the claimed satisfaction gain is an artifact of the loop's unchecked fallback rather than a property of the interval formulation.","supporting_citations":[{"cited_title":"QoE-based Semantic-Aware Resource Allocation for Multi-Task Networks","cited_arxiv_id":"2305.06543","evidence_quote":"QoE-based semantic-aware resource allocation for multi-task networks: the strict-similarity single-channel baseline the proposed method is compared against and outperforms by 17.1% in satisfied users."},{"cited_title":"Zhang, Q","cited_arxiv_id":null,"evidence_quote":"Unified Multi-Task Semantic Communication System (UDeepSC): supplies the encoder–decoder model whose SNR/compression-rate-to-similarity lookup table is used inside subproblem F2."},{"cited_title":"Nesterov and A","cited_arxiv_id":null,"evidence_quote":"Interior Point Polynomial Methods in Convex Programming: justifies the polynomial-time solution of the geometric-programming subproblem F1."},{"cited_title":"Geometric programming for communication systems","cited_arxiv_id":null,"evidence_quote":"Geometric programming for communication systems: the framework showing F1 is a geometric program solvable by standard tools such as CVX/MOSEK."}],"review_version":1}