{"id":"8cb1db52-d40c-45af-9e98-dadf5023fb0c","arxiv_id":"2608.03852","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A serverless peer-to-peer critic-sharing framework coordinates massive-MIMO scheduling, power, and beamforming across independent RAN controllers with compressed exchange and conditional convergence guarantees.","lead":"This paper proposes FedCritic-MIMO, a distributed learning method that lets neighboring cell controllers in 6G radio access networks coordinate resource decisions by exchanging only selected critic-model parameters. The method claims better throughput, SINR, QoS, and interference efficiency than independent, centralized, and uncompressed baselines, while cutting critic traffic by about 76%.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The advertised O(T^(-1/2))+O(logT/T) critic-recursion rate is conditional on unproven tracking-error decay (77), and Lemma 1's proof gives a weaker bound than stated.","rationale":"The reader's CONDITIONAL verdict is appropriate. The paper is an honest engineering contribution: it clearly labels the convergence result as conditional, scopes it to a frozen-target critic regression in Remark 2, and acknowledges the QoS overlap with CTDE-MAPPO and the lack of a periodic-compressed ablation. My stress-test confirms the weakest point is the tracking-error decay condition (77). I sharpen it: the single lemma that would justify (77) is not merely unproved but appears to be incorrectly stated, since the Appendix A proof gives a bound of order τ_th rather than τ_th^2. This matters because the paper's headline theoretical rate inherits (77) directly, and Remark 1 concedes no proof. The empirical side is not fatally undermined: the held-out reward of 57.0 versus 52–53 for the strongest baselines, with a compact IQR, provides reasonable evidence that serverless critic exchange helps, and the 76% communication reduction is clearly relative to uncompressed distributed exchange. However, the theoretical contribution should be presented as a conditional statement whose precondition is an open question, and the evaluation should include the proposed ε_t^2 measurement or a periodic-compressed baseline. Since the paper already presents the theory as conditional and the empirical direction is plausibly correct, CONDITIONAL remains the right verdict; no change is needed.","tokens_in":22985,"tokens_out":11231,"duration_ms":99688,"concrete_test":"Re-run the Section V training protocol with the same seeds and hyperparameters (or with a lightweight instrumentation pass) and record, at each round t, ε_t^2 = (1/N) Σ_{n=1}^N ||tilde-ψ^{t+1}_n − hat-ψ^{t+1}_n||^2. Then compute S_T = (1/T) Σ_{t=0}^{T-1} ε_t^2 for T=250 and check whether S_T · T / log T remains bounded, as required by (77). If it does not, Corollary 1's O(T^{-1/2})+O(logT/T) rate does not apply to the reported experiments, and the theoretical claim must be weakened to the fixed-learning-rate neighborhood statement in Corollary 1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Corollary 1's finite-time stationarity rate is the paper's main theoretical contribution, but it only holds if 1/T Σ_{t=0}^{T-1} ε_t^2 = O(logT/T) (Eq. 77), where ε_t^2 is the mean squared discrepancy between the pre-fusion shared-critic updates and their public reconstructions. Remark 1 explicitly states that this decay condition is not proved, so the advertised rate is not a theorem about the proposed mechanism unless (77) is established elsewhere. The paper's only candidate tool, Lemma 1, is internally inconsistent: the trigger definition (45) and the Appendix A proof give, for the no-trigger case, ||e^{t+1}_n||^2 < τ_th (||hat-ψ^t_n||^2 + ε_tr), not the stated τ_th^2 (||hat-ψ^t_n||^2 + ε_tr)^2. The claimed bound is tighter by a factor τ_th (||hat-ψ^t_n||^2 + ε_tr), typically much smaller than 1, and therefore does not follow from (45). Because no other argument controls the accumulated tracking error, the finite-time guarantee is conditional on a precondition that is not shown to hold for the proposed adaptive threshold schedule or any explicit schedule. If (77) fails, the finite-time stationarity guarantee does not follow, and the simulation evidence alone carries the central coordination claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FedCritic-MIMO, a decentralized multi-agent reinforcement learning framework for downlink resource control in reuse-1 multi-cell massive-MIMO OFDMA networks. Each cell-level controller keeps its own actor and personalized critic head, and coordination is achieved by exchanging only a shared critic subnetwork over an interference-aware graph, using utility-aware event-triggered communication, adaptive layer-wise top-k compression with error feedback, and balanced interference-aware fusion. The authors formulate the problem as a Dec-POMDP, provide a conditional convergence analysis of the shared-critic recursion under fixed-policy and frozen-target assumptions, and evaluate the method against heuristic, independent-learning, centralized-training, and communication-ablation baselines. The simulations report improved held-out throughput, SINR, interference efficiency, and an approximately 76% reduction in critic communication relative to uncompressed distributed critic exchange, with QoS performance comparable to a centralized-training baseline.","tokens_in":23475,"tokens_out":5901,"duration_ms":58171,"significance":"If the claims hold, the paper addresses a genuinely important architectural problem: coordinating independently deployable RAN controllers without a central trainer or parameter server. The empirical study is carefully set up with common warm starts, multiple seeds, held-out evaluation, and communication-aware baselines, and the paper is unusually candid in its limitations (Remark 1 on the tracking-error condition and Remark 2 on the scope of the theory). The main significance risk is that the advertised finite-time rate depends on an unproven condition, and the stated one-step tracking-error lemma is internally inconsistent. These issues do not invalidate the empirical contribution, but they must be repaired before the theoretical claims can be taken as established.","major_comments":[{"comment":"The advertised O(T^{-1/2}) + O(log T / T) stationarity rate is conditional on the accumulated tracking-error decay 1/T Σ_{t=0}^{T-1} ε_t^2 = O(log T / T), and Remark 1 explicitly states that this condition is not proved. Since no explicit trigger-threshold schedule, compression budget, or auxiliary argument is given that would imply (77), the finite-time guarantee is not established for the proposed mechanism under its stated assumptions. The abstract and the contribution list present this rate as a central result; the paper should either prove (77) under an explicit schedule or clearly downgrade the claim to a conditional statement whose hypothesis remains unverified.","section":"Section IV-G, Corollary 1, Eq. (77), Remark 1"},{"comment":"The statement of Lemma 1 disagrees with its proof. In the no-trigger case, (45) gives Γ_n^t = ||~ψ^{t+1}_n - hatψ^t_n||^2 / (||hatψ^t_n||^2 + ε_tr) · (1 + α_Q Q̄^t_n + α_I Ī^t_n) < τ_th, so ||e^{t+1}_n||^2 < τ_th(||hatψ^t_n||^2 + ε_tr). The lemma instead claims the much tighter bound τ_th^2(||hatψ^t_n||^2 + ε_tr)^2, which does not follow from the trigger condition as written. Because Lemma 1 is the paper's only concrete tool for controlling the tracking error that enters Theorem 1 and Corollary 1, this inconsistency is load-bearing and must be corrected, along with any subsequent uses of the stated bound.","section":"Lemma 1 and Appendix A"},{"comment":"The contribution list claims improved QoS satisfaction, but the held-out results show overlapping confidence intervals with CTDE-MAPPO, and the text in Section V-F explicitly states that the figure supports comparable QoS rather than a statistically significant advantage. The abstract and contribution wording should be qualified so that the QoS claim matches the evidence; the stronger and better-supported improvements are throughput, SINR, and interference efficiency.","section":"Section V-F and Section I-B (Contributions)"}],"minor_comments":[{"comment":"The QoS-satisfaction metric should be defined precisely in one place, and the relationship between the QoS bar in Fig. 4(a) and the rate CDF in Fig. 5 should be made explicit. The paper itself flags this ambiguity; it would be better to resolve it in the text.","section":"Section V-G"},{"comment":"The comparison between Periodic-Full and Event-Uncompressed shows that the trigger remains active in most rounds, so the measured communication savings are attributed mainly to compression. Adding a periodic-compressed baseline would isolate the contribution of event triggering; the paper notes this limitation, but the missing baseline weakens the ablation story.","section":"Section V-E"},{"comment":"The proof of Lemma 1 uses a strict inequality in the no-trigger case while the lemma statement uses a non-strict inequality; trivial, but the statement and proof should be consistent after the bound is corrected.","section":"Appendix A"},{"comment":"Equation (79) defines the directed logical payload with the factor |B_n|, but the reported overhead in Fig. 3 should state explicitly whether this factor is included or whether physical broadcast is assumed. The text explains both options, but the figure caption does not say which one was used.","section":"Section IV-H"},{"comment":"The claim that repeatedly solving (17) steers the system toward satisfying the long-term average-rate constraints would benefit from a brief justification or a pointer to the standard drift-plus-penalty argument, since this link is used to motivate the surrogate objective.","section":"Section II-B"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the paper is within the journal's scope and the empirical evaluation is substantial. The main risk is the theoretical overclaim: Corollary 1's rate is conditional on an unproven tracking-error condition, and Lemma 1 contains a definite bound error. Both are fixable in revision. I would also encourage the editors to ask the authors for a reproducibility statement or access to code/data, since the simulation results carry much of the paper's contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the short version. FedCritic-MIMO is a practical, well-executed system for coordinating interference-coupled RAN controllers by exchanging compressed shared critic parameters peer-to-peer, no central trainer. The empirical case is the real contribution: on a reuse-1 massive-MIMO OFDMA setup, it beats independent, CTDE, full-exchange, and event-uncompressed baselines on throughput, SINR, and interference-per-bit, and cuts critic communication by ~76% versus uncompressed exchange. The baselines are well chosen and the paper is honest about limitations—Remark 2 explicitly scopes the theory to a fixed-policy, frozen-target critic regression, and the authors do not oversell convergence of the full actor-critic system.\n\nWhat's new is the combination: utility-aware event triggering driven by critic innovation, queue urgency, and interference, layer-wise top-k sparsification with error feedback, and symmetric interference-aware fusion. That combination is distinct from the cited gossip/compression literature, and the analysis of the balanced compressed P2P recursion is a genuine extension over their prior FedCritic. The system modeling—Dec-POMDP, virtual queues, structured beamforming—is also careful.\n\nThe soft spots are real but not fatal. First, the advertised rate O(T^{-1/2})+O(logT/T) in Corollary 1 is conditional on Eq. (77), an accumulated-tracking-error decay condition that Remark 1 says is not proved. The paper states this openly, but it means the headline theory is not a theorem for the proposed mechanism as analyzed. Second, Lemma 1's stated no-trigger bound squares the threshold and the reconstruction error, while the proof in Appendix A gives only the linear bound. That's an internal inconsistency, and since Lemma 1 is the only tool offered to bound the tracking error, the statement should be corrected to match the proof (or a sharper argument supplied). Third, the QoS advantage over CTDE-MAPPO is within overlapping confidence intervals; the paper's own text admits this. And the 76% communication gain is mostly compression, not event triggering—Periodic-Full and Event-Uncompressed are nearly identical—so a periodic-compressed baseline is needed to isolate the contributions. No code or data, which limits external verification.\n\nStill, the central claim—that serverless critic exchange can coordinate controllers without centralized training—is supported by the evidence, and the conditional theory is a reasonable starting point. This deserves a serious referee: I'd accept it with requests for a corrected Lemma 1, a demonstration or at least a tighter argument for (77), a periodic-compressed baseline, and ideally code/data. It's a good read for anyone working on distributed MARL for wireless resource control.","headline":"A solid engineering contribution with an honest but conditional theory; the central coordination claim is plausible and worth refereeing, but the headline rate depends on an unproved tracking-error condition and Lemma 1's statement outruns its proof.","tokens_in":23814,"tokens_out":3416,"would_cite":true,"duration_ms":29392,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Serverless exchange of compressed critic parameters lets interference-coupled 6G controllers coordinate without a central trainer, cutting critic traffic by roughly 76% while matching centralized QoS.","keywords":["AI-native 6G RAN","Open RAN","federated reinforcement learning","multi-agent reinforcement learning","serverless learning","massive MIMO","resource allocation","communication-efficient learning"],"falsifier":"Measure the mean squared tracking error $\\varepsilon_t^2 = \\frac{1}{N}\\sum_{n=1}^N \\mathbb{E}\\|e_t^n\\|_2^2$ between each controller's local shared-critic parameters and the public reconstruction its neighbors hold, in the same 7-cell reuse-1 setup; if the running average $\\frac{1}{T}\\sum_{t=0}^{T-1}\\varepsilon_t^2$ does not decay at the $O(\\log T/T)$ rate, Corollary 1's advertised stationarity rate does not follow from the theory and the method's practical value rests on the simulation operating point alone. A second, cheaper check is to run a periodic-compressed baseline at the same sparsification budget $\\rho_c = 0.20$; the paper's own discussion predicts the event trigger contributes little at this operating point, because the trigger fires in most training rounds.","tokens_in":22778,"feed_emoji":"📡","tokens_out":14323,"duration_ms":98465,"temperature":0.7,"pith_summary":"The paper argues that coordinating a dense, interference-limited 6G radio access network does not require a central trainer, a parameter server, or shared policies: neighboring base-station controllers can coordinate by exchanging only the shared part of their critic (value-estimation) networks, sent only when the update matters and compressed per layer. It proposes FedCritic-MIMO, which combines an event trigger driven by critic change, queue pressure, and interference intensity; layer-wise top-$k$ sparsification with error feedback; and balanced fusion that weights strongly coupled neighbors while preserving the network average of the shared critic. Under a fixed-policy, frozen-target critic-regression model, the paper derives conditional finite-time stationarity and consensus bounds for the shared-critic recursion, with a stationarity rate of $O(T^{-1/2})+O(\\log T/T)$ when the accumulated tracking error decays as $O(\\log T/T)$. In reuse-1 massive-MIMO OFDMA simulations the method achieves the best performance-communication tradeoff among the considered baselines, raising held-out throughput and mean SINR while cutting critic-communication overhead by about 76% relative to uncompressed distributed critic exchange. A sympathetic reader would take the central claim to be that critic-only, sparsely exchanged value information is sufficient to coordinate otherwise independent controllers.","feed_headline":"Sparse critic sharing coordinates 6G cells with 76% less traffic","feed_subtitle":"Peer-to-peer exchange of sparse critic bits matches a centralized trainer on QoS while cutting inter-cell signaling.","key_machinery":"The load-bearing object is the balanced, compressed peer-to-peer critic recursion $\\Psi^{t+1} = (\\Psi^t - \\eta_c G^t)(W^t)^T + R^t$, where $\\Psi^t$ stacks the shared critic parameters across base stations, $G^t$ stacks local critic gradients, $W^t$ is a symmetric doubly stochastic fusion matrix built from two-sided interference-relevance scores, and $R^t$ is the perturbation caused by event triggering and top-$k$ compression. The key identity is Lemma 2: because $W^t$ is doubly stochastic and symmetric, the network average satisfies $\\bar{\\psi}^{t+1} = \\bar{\\psi}^t - \\eta_c \\bar{g}^t$, exactly the update of a centralized gradient method, while the perturbation obeys $\\frac{1}{N}\\mathbb{E}\\|R^t\\|_F^2 \\le 4\\varepsilon_t^2$, with $\\varepsilon_t^2$ the mean squared distance between local shared-critic parameters and the public reconstructions neighbors hold. The disagreement recursion $Z^{t+1} = (Z^t - \\eta_c G^t\\Pi)(A^t)^T + R^t$ with $\\|A^t\\|_2 \\le \\lambda_W < 1$ then yields the consensus bound, so both the theory and the practical benefit hinge on keeping the tracking error small relative to the spectral gap $1-\\lambda_W$.","core_discovery":"FedCritic-MIMO's central claim is that compatible shared critic parameters exchanged peer-to-peer over an interference-aware graph can coordinate RAN controllers without centralized trajectory collection, parameter-server aggregation, or actor homogenization. Each base station keeps its own actor, local experience, and personalized critic head; only a designated shared critic subnetwork $\\psi^n_t$ is transmitted, and only when a utility score $\\Gamma^n_t$ combining critic innovation, queue urgency, and interference intensity crosses a threshold. Messages are layer-wise top-$k$ compressed with error feedback, and fusion uses symmetric interference-relevance weights $W_t$ that are doubly stochastic, so the network average of the shared critic parameters evolves exactly like a centralized gradient recursion while a perturbation controlled by the tracking error $\\varepsilon_t^2$ enters the stationarity and consensus bounds (Theorem 1). Under the decay condition $\\frac{1}{T}\\sum_{t=0}^{T-1}\\varepsilon_t^2 = O(\\log T/T)$, the randomized-iterate stationarity rate is $O(T^{-1/2})+O(\\log T/T)$ (Corollary 1); Remark 1 states explicitly that this decay condition is not proved. In simulations over seven reuse-1 massive-MIMO cells, the framework attains the highest held-out reward, a mean SINR about 1.7–2.5 dB above the strongest competing learning baselines, QoS comparable to the centralized-training baseline, the lowest interference cost per delivered bit, and about 76% lower critic-communication overhead than uncompressed distributed critic exchange.","pith_inferences":["A direct empirical check of whether $\\frac{1}{T}\\sum_{t=0}^{T-1}\\varepsilon_t^2$ decays as $O(\\log T/T)$ in the same 7-cell setup would reveal whether the advertised rate is realized in practice or whether the method's value rests on simulations alone; the paper does not perform this check.","The same trigger-compress-fuse machinery should transfer to asynchronous controllers, time-varying interference graphs, and imperfect CSI, which the paper lists as future work; the doubly stochastic fusion property and the average-preservation identity would need revisiting when exchange is asymmetric or graph edges change.","The paper's own note that the QoS bar and the user-rate CDF must come from the same samples if both measure the same quantity invites a re-derivation of the reported satisfaction ratio directly from the held-out rate samples as a consistency check.","Sweeping the compression budget $\\rho_c$ and the trigger floor $\\tau_{th,\\min}$ should trace a communication-performance frontier shaped by $\\bar{\\varepsilon}^2/(1-\\lambda_W)^2$; a reader could test whether the tradeoff curve bends as the theory predicts."],"forward_implications":["Critic-only, sparsely exchanged value information is sufficient for coordination: controllers improve throughput, user-rate distribution, mean SINR, QoS satisfaction, and interference efficiency while keeping actors and personalized critic heads private.","The serverless architecture matches the centralized-training baseline (CTDE-MAPPO) on QoS and beats it on interference cost, so removing the central trainer need not sacrifice coordination quality.","The advertised $O(T^{-1/2})+O(\\log T/T)$ stationarity rate holds only for the fixed-policy, frozen-target critic-regression phase and only if the accumulated tracking error decays as $O(\\log T/T)$; it does not certify the full actor-critic process or global optimality.","At the reported operating point, the roughly 76% communication saving comes mainly from layer-wise sparse exchange with error feedback rather than from fewer communication rounds, since the event trigger fires in most training rounds."],"supporting_citations":[{"why":"Supplies the fully decentralized networked-agent MARL paradigm on which peer-to-peer critic exchange is built.","marker":"[22]"},{"why":"Provides the finite-time decentralized actor-critic analysis whose tools the shared-critic recursion and consensus bounds extend.","marker":"[23]"},{"why":"Gives the gossip-based decentralized optimization and compressed-communication foundation for serverless model exchange.","marker":"[32]"},{"why":"Establishes error feedback as the mechanism that corrects top-k compression bias, used in the exchange rule.","marker":"[33]"},{"why":"Is the authors' earlier FedCritic framework that this paper extends to massive-MIMO scheduling, power allocation, and beamforming.","marker":"[34]"},{"why":"Supports the event-triggered communication design, specifically the choice of a decreasing threshold versus a positive floor.","marker":"[36]"},{"why":"Justifies treating the tracking-error term as the communication-induced component of the convergence behavior in wireless federated learning.","marker":"[37]"}],"fun_headline_variants":["Peer-only critic sharing cuts 6G overhead 76%","No central trainer: peer critic sharing cuts 76% traffic","76% less critic traffic via peer-only sparse exchange","Serverless critic sharing cuts 76% overhead with same QoS","Sparse peer critic exchange: 76% less signaling and same QoS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The finite-time stationarity rate is conditioned on a decay property of the accumulated tracking error — the combined error from skipped communications and compressed critic updates — that the paper explicitly does not prove; if that error does not shrink at least as fast as $O(\\log T/T)$, the guarantee collapses to the simulation results alone.","fun_headline_variants_meta":{"raw":{"variants":["Peer-only critic sharing cuts 6G overhead 76%","No central trainer: peer critic sharing cuts 76% traffic","76% less critic traffic via peer-only sparse exchange","Serverless critic sharing cuts 76% overhead with same QoS","Sparse peer critic exchange: 76% less signaling and same QoS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001446,"raw_usage":{"total_tokens":5945,"prompt_tokens":1182,"completion_tokens":4763,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":798,"completion_tokens_details":{"reasoning_tokens":4677}},"tokens_in":798,"tokens_out":4763,"duration_ms":30094,"temperature":1.0,"reasoning_tokens":4677,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:45:44.928395+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the mean squared tracking error $\\varepsilon_t^2 = \\frac{1}{N}\\sum_{n=1}^N \\mathbb{E}\\|e_t^n\\|_2^2$ between each controller's local shared-critic parameters and the public reconstruction its neighbors hold, in the same 7-cell reuse-1 setup; if the running average $\\frac{1}{T}\\sum_{t=0}^{T-1}\\varepsilon_t^2$ does not decay at the $O(\\log T/T)$ rate, Corollary 1's advertised stationarity rate does not follow from the theory and the method's practical value rests on the simulation operating point alone. A second, cheaper check is to run a periodic-compressed baseline at the same sparsification budget $\\rho_c = 0.20$; the paper's own discussion predicts the event trigger contributes little at this operating point, because the trigger fires in most training rounds.","supporting_citations":[{"cited_title":"Emergent communication in multi-agent reinforcement learning for future wireless networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the fully decentralized networked-agent MARL paradigm on which peer-to-peer critic exchange is built."},{"cited_title":"A collaborative m ulti-agent deep reinforcement learning-based wireless power allocation w ith centralized training and decentralized execution,","cited_arxiv_id":null,"evidence_quote":"Provides the finite-time decentralized actor-critic analysis whose tools the shared-critic recursion and consensus bounds extend."},{"cited_title":"Fully decentralized multi-agent reinforcement learnin g with networked agents,","cited_arxiv_id":null,"evidence_quote":"Is the authors' earlier FedCritic framework that this paper extends to massive-MIMO scheduling, power allocation, and beamforming."},{"cited_title":"Fully decentralized coopera tive multi-agent reinforcement learning: a survey,","cited_arxiv_id":null,"evidence_quote":"Supports the event-triggered communication design, specifically the choice of a decreasing threshold versus a positive floor."},{"cited_title":"Learn-to-share: A decentralized multi-agent spectrum sharing framework for heterogeneous networks in the 6G era,","cited_arxiv_id":null,"evidence_quote":"Justifies treating the tracking-error term as the communication-induced component of the convergence behavior in wireless federated learning."}],"review_version":2}