{"id":"6eab7c46-43e7-4f8f-a587-4448ea73f2f8","arxiv_id":"2501.01637","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A two-tier edge network jointly optimizes knowledge sharing, semantic extraction, and resource allocation to maximize a generalized effective semantic transmission rate under accuracy and delay constraints.","lead":"This paper proposes a collaborative knowledge sharing mechanism for semantic-bit communication networks that maximizes a generalized effective semantic transmission rate by jointly optimizing knowledge sharing, semantic extraction ratios, and base station/subchannel allocations. It may be useful to engineers designing task-oriented semantic communication systems where transmitter and receiver knowledge bases are mismatched.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported θ parameters violate Eq. (15)'s non-negativity constraints, so the semantic accuracy function is non-monotone and exceeds 1; the ξ-threshold derivation and all GESTR simulations rest on this invalid surrogate.","rationale":"Good-faith reading: the paper's main algorithmic idea is a two-stage decomposition—for each fixed (i,j,k), solve the joint knowledge-sharing/extraction subproblem, then choose the best BS/subchannel assignment via weighted matching. This decomposition is structurally sound because subchannels are orthogonal, so subproblems decouple and the K-M stage handles constraints (5)–(6) exactly. I therefore do not attack the architecture of the proof. The load-bearing flaw is in the semantic-accuracy surrogate, which sits underneath both the objective and the constraints. The reader's weakest assumption flags Eq. (15) as an unvalidated fitted curve; the stronger finding here is that the specific parameter tuple used in the simulations contradicts the model's own sign constraints, making ε(ξ) non-monotone and > 1. This is not merely an external-validity concern waiting for more data; it is an internal inconsistency in the manuscript as written. The discretized ξ search and lack of convergence verification noted by the reader are secondary: even if those were fixed, the reported numerical comparisons would still rest on an invalid accuracy function. The appropriate action is a major revision with corrected parameters and rerun experiments, so the verdict remains CONDITIONAL; I leave it unchanged relative to the reader.","tokens_in":10070,"tokens_out":8646,"duration_ms":85317,"concrete_test":"Evaluate Eq. (15) at ξ=0 and ξ=1 using the parameter tuple reported in Section IV: (−6.205e−8, 16.45, 0.9228, −0.06917). If ε(0) > 1 or ε(0) < ε(1), the monotonicity and the ξ^th derivation are contradicted; then re-run the simulations with parameter values satisfying θ_i ≥ 0 (or with the surrogate actually fitted in [9]) and check whether the proposed scheme still outperforms the baselines at low SNR.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The numerical claims are not established because the semantic-accuracy surrogate used in the simulations is internally inconsistent. Equation (15) defines ε = −θ1 e^{θ2(1−ξ)} + θ3 e^{−θ4(1−ξ)} with θ1, θ2, θ3, θ4 ≥ 0, and the text states that ε is monotonically increasing in ξ, so constraint (16) yields a lower bound ξ ≥ ξ^th. Section IV instead uses (θ1, θ2, θ3, θ4) = (−6.205e−8, 16.45, 0.9228, −0.06917), violating θ1 ≥ 0 and θ4 ≥ 0. With these values ε(0) ≈ 1.85 and ε(1) ≈ 0.92, so accuracy exceeds 1 and decreases with ξ, opposite to the monotonicity premise. The derivation of ξ^th in Section III-A therefore fails, and all GESTR values in Figures 3–4 are computed from an invalid accuracy model. The claim that the proposed FP-BnB/K-M solution achieves the optimum and outperforms baselines at low SNR is unsupported as written. The two-stage decomposition itself is plausible—subproblems decouple over (i,j,k) and the allocation is a weighted matching—but solving a mis-specified objective cannot validate the performance claims.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a two-tier semantic-bit communication network in which mobile devices (MDs) with mismatched knowledge bases (KBs) can share missing knowledge either by uploading it to the serving base station or, for small-cell associations, downloading it from the macro base station, with remaining data carried by bit communication. The authors define a generalized effective semantic transmission rate (GESTR) that includes semantic accuracy and knowledge-sharing overhead, formulate a mixed-integer nonlinear program (MINLP) to maximize total GESTR over knowledge-sharing decisions, extraction ratios, and BS/subchannel allocations, and propose a decomposition: per-association subproblems solved by a fractional-programming branch-and-bound (FP-BnB) method, followed by a modified Kuhn-Munkres (K-M) matching for allocation. Simulations compare the proposed solution with no-collaboration and no-knowledge-sharing baselines, reporting gains that are larger at low SNR and low bandwidth.","tokens_in":10367,"tokens_out":12561,"duration_ms":123589,"significance":"If the correctness issues were resolved, the paper would make a moderate contribution: the decomposition into independent per-association joint subproblems plus a weighted bipartite matching is natural and mostly well executed, and the explicit treatment of knowledge-sharing overhead in the semantic rate is a useful modeling step. The Dinkelbach-style fractional programming inside a branch-and-bound framework is standard but is applied here to a relevant problem, and the reduction of the final allocation to maximum-weight matching is clean. However, the paper does not provide machine-checked proofs or code, and the numerical validation as written rests on an internally inconsistent semantic-accuracy model, so the empirical claims are not currently established.","major_comments":[{"comment":"The simulation parameters (θ1, θ2, θ3, θ4) = (−6.205e−8, 16.45, 0.9228, −0.06917) violate the non-negativity constraints θ1 ≥ 0 and θ4 ≥ 0 stated in Eq. (15). With these values, ε(0) ≈ 1.85 and ε(1) ≈ 0.92, so the semantic accuracy exceeds 1 and decreases with ξ, directly contradicting the monotonicity premise used in Section III-A to derive the lower bound ξ ≥ ξ^th and to define the search range [ξ^th, 1]. Consequently, constraint (16) is not enforced as written, and all GESTR values reported in Figs. 3 and 4 are computed from an invalid accuracy model. The authors must either supply fitted parameters that satisfy θ1, θ2, θ3, θ4 ≥ 0 and are consistent with a monotone increasing ε(ξ), or correct Eq. (15) to match the sign convention of the source [9], and then re-run all simulations.","section":"Section IV, Eq. (15)"},{"comment":"The claimed global optimality of the joint solution is not established for the continuous variable ξ. The manuscript replaces ξ ∈ [ξ^th, 1] by M equal segments and searches only the discrete points ξ^(m), but the objective γ_{i,j,k} depends on ξ nonlinearly through ε(ξ), the semantic transmission time tS ∝ ξ, and the computing time tR ∝ ξ^{−ρ}; no monotonicity or convexity argument is given to guarantee that an optimum occurs at a grid point, and no error bound is provided for finite M. The value of M is also never reported in the simulation setup. The authors should either prove that the discretization is exact, provide a rigorous approximation guarantee, or revise the claim from 'joint optimum' to an approximate solution, and they should report M.","section":"Section III-A and Section IV, Table I"},{"comment":"The two baseline schemes, 'no collaboration' and 'no knowledge sharing', are only identified by citations to [10] and [15]; the optimization problems solved for these baselines are not stated. Since the central performance claim is that the proposed mechanism outperforms these baselines, the comparison is only meaningful if the reader can verify which variables are optimized and which constraints are retained. Please specify the baseline formulations explicitly, or state clearly that they are taken verbatim from the cited papers with their original parameter settings.","section":"Section IV, baselines"}],"minor_comments":[{"comment":"The entry 'f C n [4, 2] G Hz' appears to be a typographical artifact; please clarify the computing speed f_C^j and its units.","section":"Table I"},{"comment":"The sentence 'Each MD m has a minimum semantic accuracy requirement' uses the symbol m, but MDs are indexed by i; please correct the notation.","section":"Section II-C, after Eq. (16)"},{"comment":"The phrase 'one-dimension l inear search' contains a missing space and should read 'one-dimensional linear search'.","section":"Section III-A"},{"comment":"The substitution b̃ = ab is acceptable because Y(a,b) and T(a,b) depend on b only through a and ab; it would improve readability to state this explicitly when introducing b̃.","section":"Section III-A, Eq. (20)-(21)"},{"comment":"The objective allows δi,j,k = 0 for all k for some MDs, so the 'total GESTR of all MDs' may be maximized by not serving some MDs; if the intent is to serve every MD, a minimum-service constraint should be added.","section":"Problem (18)"}],"recommendation":"major_revision","confidential_remarks":"The most serious problem is the parameter inconsistency in the semantic-accuracy model: the values used in the simulation directly violate the constraints stated in Eq. (15), and the resulting ε(ξ) is neither monotone nor bounded by 1. This is fixable in principle by re-fitting the surrogate or correcting Eq. (15), but the current numerical results cannot be used. Please also ask the authors to report M and to describe the baselines precisely; without those details, the claimed performance gains are not verifiable even after the accuracy model is repaired."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe interesting part of this paper is the collaborative knowledge-sharing mechanism: when an MD associates with an SBS, mismatched KB classes can be uploaded from the MD or downloaded from the MBS via backhaul, and unshared classes fall back to bit communications. That is a sensible extension of the KB-mismatch literature, and the GESTR metric that includes sharing overhead is a reasonable way to evaluate it. The problem formulation is detailed, and the decomposition into per-link joint subproblems plus a weighted matching allocation is structurally sound.\n\nBut there is a load-bearing flaw in the numerical section. Equation (15) defines semantic accuracy as ε = −θ1 e^{θ2(1−ξ)} + θ3 e^{−θ4(1−ξ)} with θ1, θ2, θ3, θ4 ≥ 0 and states ε is monotonically increasing in ξ. Yet Section IV adopts (θ1, θ2, θ3, θ4) = (−6.205e−8, 16.45, 0.9228, −0.06917), which violates θ1 ≥ 0 and θ4 ≥ 0. With those values ε(0) ≈ 1.85 and ε(1) ≈ 0.92, so accuracy exceeds 1 and decreases with ξ, the opposite of the stated premise. The threshold ξ^th derived from monotonicity is therefore meaningless, and the GESTR curves in Figures 3–4 are computed from an invalid model. This is not a subtle issue; it is an internal contradiction in the paper's own equations.\n\nThe rest of the paper has soft spots too. The transformation from (20) to (21) using the auxiliary variable ~b is plausible but the recovery of binary b* from ~b* is not clearly specified. The FP-BnB algorithm is claimed to find the global optimum, but no convergence or complexity analysis specific to this non-convex problem is provided. Those are secondary; the surrogate invalidates the main performance claims.\n\nWho is this for? Researchers working on resource allocation in semantic communications. The collaborative sharing idea deserves attention, and the framework could be useful after the surrogate is fixed and simulations re-run with a valid accuracy model (or with real semantic codecs). I would send this to a serious referee, but only with the expectation of major revision. As it stands, the paper is not publishable without redoing the numerical results.\n\nRecommendation: engage with the work, but insist on replacing or re-validating the semantic accuracy model and re-running the experiments.","headline":"The collaborative knowledge-sharing framework is a legitimate extension, but the simulations rest on a self-contradictory accuracy model, so the performance claims don't hold as written.","tokens_in":10883,"tokens_out":3219,"would_cite":false,"duration_ms":27578,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper establishes that a two-tier edge network with collaborative knowledge sharing—uploading or downloading mismatched knowledge and using bit fallback—achieves the exact maximum of a generalized effective semantic transmission rate…","keywords":["semantic communications","knowledge base mismatch","knowledge sharing","effective semantic transmission rate","resource allocation","branch and bound","Kuhn-Munkres algorithm","fractional programming"],"falsifier":"A brute-force enumeration of all feasible associations, transmission modes, sharing manners, and extraction ratios on a small network instance would settle the optimality claim: if any enumerated solution has total GESTR higher than the FP-BnB plus modified Kuhn-Munkres output, the claimed global optimum is false.","tokens_in":9844,"feed_emoji":"📡","tokens_out":7264,"duration_ms":66038,"temperature":0.7,"pith_summary":"The paper addresses knowledge base mismatch in task-oriented semantic communications, where transmitters and receivers must share the same knowledge to exchange semantics and misalignment degrades performance. It proposes collaborative sharing in a two-tier edge network: mismatched knowledge is uploaded from mobile devices or downloaded from the macro base station, with unshared knowledge transmitted as ordinary bits. To measure performance, it defines the generalized effective semantic transmission rate (GESTR), which combines semantic accuracy with the overhead of knowledge sharing. The core claim is that the resulting mixed-integer nonlinear program can be solved to global optimality by splitting it into per-association subproblems solved by fractional-programming branch-and-bound and an assignment subproblem solved by a modified Kuhn-Munkres algorithm. Simulations show this design outperforms baselines without collaboration or knowledge sharing, with the largest gains when signal-to-noise ratios are low.","feed_headline":"Optimal knowledge-sharing plan maximizes semantic transmission rate","feed_subtitle":"FP-BnB plus Kuhn-Munkres solving beats no-collaboration baselines, with biggest gains in poor channels.","key_machinery":"The driving object is the generalized effective semantic transmission rate (GESTR), a ratio objective that counts successfully delivered semantic information in suts per second, with knowledge-sharing time treated as overhead in the denominator. The solution mechanism has three parts: a fractional-programming transform (the parametric subtractive form of the ratio, updated iteratively) that turns each relaxed branch-and-bound subproblem into a linear program; the linearization of the product of binary transmission-mode and sharing-manner variables via an auxiliary variable $\\tilde{b} = a b$; and branch-and-bound over the relaxed binaries whose bounds prune the search. Once these joint subproblems are solved, the remaining assignment of devices to subchannels is a bipartite matching on the computed GESTR weights, solved by a modified Kuhn-Munkres algorithm. The key identity that makes the decomposition work is that, for fixed association, the objective and constraints factor per device, so each device's joint subproblem can be solved independently before matching.","core_discovery":"The central claim is that a two-tier semantic-bit network's total generalized effective semantic transmission rate is maximized by jointly choosing, for every device, whether each mismatched knowledge class is shared via an MD-to-BS upload or an MBS-to-SBS download, how much semantic data is extracted (the extraction ratio), and which base station and subchannel serve the device. The paper shows that given an association, the remaining subproblem separates across devices and can be solved optimally by relaxing the binary transmission-mode and sharing-manner variables, applying fractional programming to the ratio objective, and branching over fractional solutions with branch-and-bound. The optimal subproblem values are then fed to a bipartite matching solved by a modified Kuhn-Munkres algorithm to assign base stations and subchannels. The result is a claimed global optimum for the original MINLP, and the paper demonstrates that the collaborative knowledge-sharing mechanism yields higher GESTR than no-collaboration and no-knowledge-sharing baselines, particularly in low-SNR conditions.","pith_inferences":["Editorial inference: the FP-BnB-plus-matching pipeline is not specific to GESTR; any ratio objective that separates given an assignment, with binary variables entering linearly after a product linearization, can use the same decomposition, so the method likely transfers to other joint mode-selection and resource-allocation problems.","Editorial inference: the paper's optimality is relative to the fitted semantic-accuracy curve; a natural extension is to replace the closed-form surrogate with a lookup table or a differentiable neural accuracy model and keep the same optimization machinery, or to close the loop by fine-tuning knowledge-sharing decisions from measured accuracy feedback.","Editorial inference: the constraint that one device gets at most one subchannel and one subchannel is used by at most one device makes the allocation a simple matching; in multi-carrier or OFDMA settings with frequency reuse, the allocation becomes a harder generalized assignment, and the exactness of the Kuhn-Munkres step would be lost, though the FP-BnB subproblems would remain valid.","Editorial inference: the low-SNR advantage suggests collaborative sharing acts as a form of diversity; an experimental study varying the MBS-to-SBS backhaul capacity or the ratio of stored KB classes could quantify when downloading from the MBS ceases to help."],"forward_implications":["The joint decomposition yields the exact optimum, so it supplies a performance upper bound against which heuristic or learning-based resource allocators in semantic networks can be benchmarked.","The collaborative scheme shows that when device-to-SBS channels are weak, downloading mismatched knowledge from the MBS to the SBS is the better sharing manner, which transfers the burden from power-limited devices to the wired backhaul.","The GESTR metric, including knowledge-sharing overhead and semantic accuracy, offers a single objective that trades off semantic fidelity against the cost of alignment; this can serve as a design target for future hybrid semantic-bit systems.","Since bit communications act as a fallback for unshared knowledge, the formulation naturally covers mixed semantic and bit transmission without forcing every knowledge class to be shared, which matters for delay-constrained tasks."],"supporting_citations":[{"why":"provides the closed-form nonlinear semantic accuracy model $\\epsilon'(\\xi|\\theta_1,\\theta_2,\\theta_3,\\theta_4)$ and the fitted parameter values used to enforce the accuracy constraint and compute GESTR in all simulations.","marker":"[9]"},{"why":"is the no-collaboration baseline scheme that the proposed collaborative mechanism is compared against in the simulations.","marker":"[10]"},{"why":"supplies the convergence proof of the iterative fractional programming update used inside the branch-and-bound to solve the relaxed joint subproblems.","marker":"[14]"},{"why":"is the matching algorithm the authors modify into the Kuhn-Munkres procedure that optimally solves the BS/subchannel allocation subproblem.","marker":"[13]"},{"why":"is the no-knowledge-sharing baseline scheme used for comparison, isolating the contribution of knowledge sharing to GESTR.","marker":"[15]"},{"why":"provides the channel model and link-gain parameters (path loss and Rayleigh fading) and semantic extraction model settings adopted in the numerical setup.","marker":"[16]"},{"why":"is the nonlinear least-squares fitting algorithm used to obtain the parameters of the semantic accuracy surrogate.","marker":"[11]"}],"fun_headline_variants":["Collaborative knowledge sharing maximizes semantic transmission rate","Joint sharing and allocation achieve global semantic rate optimum","Semantic-bit rate maximized by collaborative knowledge sharing","Maximize semantic rate via knowledge sharing and allocation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire optimization and all numerical results rest on the closed-form semantic-accuracy surrogate $\\epsilon'(\\xi)$ in equation (15), whose four parameters are fitted by nonlinear least squares to data from reference [9] rather than derived from the actual neural semantic codecs used for the tasks; if that curve does not match the real accuracy-versus-extraction-ratio relation, the optimal decisions computed from it may not be optimal in practice.","fun_headline_variants_meta":{"raw":{"variants":["Collaborative knowledge sharing maximizes semantic transmission rate","Joint sharing and allocation achieve global semantic rate optimum","Semantic-bit rate maximized by collaborative knowledge sharing","Maximize semantic rate via knowledge sharing and allocation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001679,"raw_usage":{"total_tokens":6651,"prompt_tokens":931,"completion_tokens":5720,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":5661}},"tokens_in":547,"tokens_out":5720,"duration_ms":43239,"temperature":1.0,"reasoning_tokens":5661,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:21:39.440504+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A brute-force enumeration of all feasible associations, transmission modes, sharing manners, and extraction ratios on a small network instance would settle the optimality claim: if any enumerated solution has total GESTR higher than the FP-BnB plus modified Kuhn-Munkres output, the claimed global optimum is false.","supporting_citations":[{"cited_title":"Adaptable semantic compres- sion and resource allocation for task-oriented communicat ions,","cited_arxiv_id":null,"evidence_quote":"provides the closed-form nonlinear semantic accuracy model $\\epsilon'(\\xi|\\theta_1,\\theta_2,\\theta_3,\\theta_4)$ and the fitted parameter values used to enforce the accuracy constraint and compute GESTR in all simulations."},{"cited_title":"Task-oriented and semantic-aware heterogeneous network s for artificial intelligence of things: Performance analysis and optimiza tion,","cited_arxiv_id":null,"evidence_quote":"is the no-collaboration baseline scheme that the proposed collaborative mechanism is compared against in the simulations."},{"cited_title":"Joint computatio n offloading and radio resource allocations in wireless cellular networ ks,","cited_arxiv_id":null,"evidence_quote":"is the matching algorithm the authors modify into the Kuhn-Munkres procedure that optimally solves the BS/subchannel allocation subproblem."},{"cited_title":"Knowledge base aware semantic communication in vehicular networks,","cited_arxiv_id":null,"evidence_quote":"is the no-knowledge-sharing baseline scheme used for comparison, isolating the contribution of knowledge sharing to GESTR."},{"cited_title":"Semantic extraction mode l selection for IoT devices in edge-assisted semantic communications,","cited_arxiv_id":null,"evidence_quote":"provides the channel model and link-gain parameters (path loss and Rayleigh fading) and semantic extraction model settings adopted in the numerical setup."},{"cited_title":"The levenberg-marquardt algorithm for no nlinear least squares curve-fitting problems,","cited_arxiv_id":null,"evidence_quote":"is the nonlinear least-squares fitting algorithm used to obtain the parameters of the semantic accuracy surrogate."}],"review_version":1}