{"id":"96ae6681-00fa-4674-8c9e-86362b40f562","arxiv_id":"2412.03076","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Coordinated multi-agent bandits that share rewards improve spatial reuse in simulated Wi-Fi networks, increasing mean throughput by up to 15% and minimum throughput by up to 210% over OBSS/PD SR.","lead":"This paper tests whether multiple Wi-Fi access points, coordinating through a future Wi-Fi 8 feature, can learn better interference settings using multi-armed bandit algorithms. A simulation study reports modest average throughput gains and large fairness improvements over the current spatial reuse approach.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Perfect-monitoring MAPC assumption is the load-bearing premise; if reward exchange is lossy, delayed, or consumes airtime, the reported 15%/210%/<3ms gains may not survive.","rationale":"The central claim is that coordinated MABs, rather than decentralized ones, give the reported gains. That claim depends on the coordination mechanism actually delivering the rewards. Section III-B assumes negligible and lossless communication at each Δ. The MAB update equations are exact average/min/log functions of all agents' rewards; they are not robust to missing or delayed reports. The simulation results in Figures 7-8 therefore describe an ideal coordinated system. The authors themselves relegate overheads and wireless-link impairments to future work, so this is not a hidden flaw, but it is the point where the argument is least supported. I compared this with other candidates: lack of error bars and an un-tuned baseline would affect the precision of the numbers, and the Thompson-sampling update in Eq. (3) is hard to verify from the text, but neither is as central as the coordination channel; ε-greedy AVG/PF, which drive the headline gains, do not depend on Eq. (3). A Komondor experiment with realistic MAPC message loss/delay would directly test whether coordination survives non-ideal reporting. Until that is run, the paper's quantitative claims should remain conditional, not be taken at face value.","tokens_in":8365,"tokens_out":11320,"duration_ms":117504,"concrete_test":"Extend the Komondor 9-BSS random scenario with a concrete MAPC reward-reporting channel: at the end of each Δ=0.5s iteration each agent transmits its normalized throughput in a small (e.g., 100-byte) control frame over the 20 MHz channel with a realistic packet error rate (e.g., 1% and 10%) and a bounded one-iteration delivery delay. Re-run the R=100 drops for Coord-ε-greedy-AVG and Coord-ε-greedy-PF. If mean throughput, minimum throughput, and max access delay versus the OBSS/PD SR baseline shift enough that the 15%/210%/<3ms claims no longer hold, the perfect-monitoring assumption is the binding constraint on the central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline numbers (15% mean throughput, +210% minimum throughput, <3 ms max delay) all come from simulations in which coordination is implemented as perfect monitoring: Section III-B states that \"the communication between agents for sharing the rewards is negligible and lossless, and occurs at the end of each learning iteration of fixed duration Δ.\" The MAB update in Algorithm 1 uses the shared reward exactly, so every AVG/MAX-MIN/PF value presumes that all agents receive all individual rewards, on time, every iteration. In a real IEEE 802.11bn MAPC deployment, reward reports must be transported over the same infrastructure or medium and are subject to contention, delay, and loss. A lost or stale reward report does not merely add overhead; it changes the computed reward and therefore the arm-selection statistics. For example, MAX-MIN would be computed from an incomplete set of BSS rewards, and PF could encounter log(0) if a starved BSS's report is missing. The authors explicitly defer this to future work, but that means the central quantitative claim is currently stated for an idealized coordination channel, not for the MAPC framework it invokes. Since the main contribution over prior decentralized MABs is precisely the coordination, this is the most load-bearing unsupported premise.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a coordinated multi-agent multi-armed bandit (MA-MAB) framework for spatial reuse (SR) in IEEE 802.11 networks. Each access point (AP) is an agent that tunes its packet-detect (PD) threshold and transmit power, and, using the MAPC framework, agents share rewards according to one of three criteria (AVG, MAX-MIN, PF). The authors evaluate epsilon-greedy and Thompson sampling under these reward-sharing strategies in the Komondor simulator, first on a two-BSS toy scenario and then on 100 random 9-BSS deployments. They report that coordinated MABs improve mean throughput by up to 15% and minimum throughput by up to 210% compared with OBSS/PD SR, while keeping maximum access delay below 3 ms. The paper positions the proposal as a low-complexity, learning-based alternative to explicit coordinated spatial reuse (C-SR) for Wi-Fi 8.","tokens_in":8633,"tokens_out":3346,"duration_ms":31268,"significance":"If the results hold, the work is relevant to the ongoing IEEE 802.11bn standardization effort: it offers a concrete, low-complexity way to realize MAPC-based SR without the joint optimization overhead of explicit C-SR, and it systematically compares reward-sharing criteria. The paper also provides a useful demonstration that coordinated bandits can avoid the conservative equilibrium that arises with selfish bandits. However, the central quantitative claims rest entirely on simulations that assume an ideal coordination channel, and the statistical evidence is presented as summary statistics over 100 random drops without confidence intervals or hypothesis tests. The headline numbers in the abstract should therefore be read as best-case simulation results for an idealized MAPC, not as validated performance guarantees for a realistic 802.11bn deployment. The strengths are the clarity of the problem formulation, the breadth of compared strategies, and the use of a publicly known simulator; the weaknesses are the idealized coordination assumption and the lack of statistical rigor in the evaluation.","major_comments":[{"comment":"The perfect-monitoring assumption is load-bearing for the central claim. The text states that 'the communication between agents for sharing the rewards is negligible and lossless, and occurs at the end of each learning iteration of fixed duration Δ.' Every reward-sharing strategy, especially MAX-MIN and PF, is computed from the full set of individual rewards. In a real MAPC deployment, reward reports must be transported over a wireless medium and are subject to loss, delay, and contention; a single missing report changes the computed reward and hence the action-selection statistics. Since the paper explicitly defers overhead and imperfect-information analysis to future work, the abstract's 15%/210%/<3 ms numbers describe an idealized coordination channel, not the MAPC framework as it would actually operate. The authors should either quantify the sensitivity to lossy/delayed reward exchange or substantially soften the claims.","section":"Section III-B"},{"comment":"The evaluation over 100 random drops reports only the mean, minimum, and maximum performance across BSSs, with no confidence intervals, error bars, or statistical significance tests. Given the large shadowing factor (σ = 9.5 dB) and the small number of BSSs (9), it is unclear whether the reported differences (e.g., 15% mean throughput, 210% minimum throughput) are statistically distinguishable from noise. The authors should provide per-drop distributions, confidence intervals, and paired significance tests (e.g., bootstrap or Wilcoxon signed-rank) for the key comparisons against OBSS/PD SR.","section":"Section IV-B, Figure 7"},{"comment":"The Thompson sampling reward update appears to contain a typo: as printed, \\(\\hat{r}_k \\leftarrow \\hat{r}_k N_k + r_k/(r_k + 2)\\) is dimensionally inconsistent and cannot be the sample-mean update described in the text. The standard update is \\(\\hat{r}_k \\leftarrow (\\hat{r}_k N_k + r_k)/(N_k + 1)\\), or an equivalent incremental form. Since the Thompson sampling results in Figures 4, 5, and 7 depend on this update, the formula should be corrected and the simulations checked against the corrected update.","section":"Equation (3)"},{"comment":"The abstract presents best-case numbers as if they are general: 'mean throughput increases by 15%, fairness is improved by increasing the minimum throughput across the network by 210%, while the maximum access delay is kept below 3 ms.' In Section IV-B these are described as 'up to' values for specific algorithm/reward combinations (Coord-ε-greedy-AVG for mean throughput, Coord-ε-greedy-PF for minimum throughput), and the maximum-delay claim applies only to the MAB implementations, not to OBSS/PD SR. The claims should be scoped to the specific configurations that achieve them, and the abstract should reflect that these are simulation results for an idealized MAPC.","section":"Abstract and Section IV-B"},{"comment":"The conclusion states that coordinated MABs offer 'a significant advantage over the OBSS/PD SR operation and uncoordinated MABs,' but the comparison against uncoordinated bandits appears only in the two-BSS toy scenario (Section IV-A, Table II and Figure 3). In the random 9-BSS scenarios, Figure 7 compares only OBSS/PD SR with the coordinated MAB variants. The claim of advantage over uncoordinated MABs in general deployments is therefore not supported by the reported experiments. Either add the uncoordinated-bandit baseline to the random-scenario evaluation or restrict the conclusion to the scenarios actually tested.","section":"Conclusions and Section IV-B"}],"minor_comments":[{"comment":"The ordering of lines 3 and 4 is confusing: line 3 says to update the reward based on performance observed in t+1, while line 4 selects the arm. In the text, the selection happens first and the reward is received in the next iteration. The algorithm should match the textual description.","section":"Algorithm 1"},{"comment":"For Thompson sampling, the columns AVG and PF report identical percentages in all rows, which is suspicious and may indicate a data transcription error or a mislabeled column. Please verify the table entries.","section":"Table II"},{"comment":"The sentence 'Thompson sampling addresses exploration-exploitation in a different way than ε-greedy: that it samples arms based on their probability of being optimal' is grammatically incomplete; 'in that' or 'namely' would fix the clause.","section":"Section II"},{"comment":"The text says 'NA V time' with an inconsistent space; it should be 'NAV time' throughout.","section":"Section IV"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of a networking journal and addresses a timely standardization topic. The main risk is that the headline results are presented for an idealized coordination channel, while the paper itself identifies the overhead analysis as future work. I would encourage the editor to require the authors to either add a sensitivity analysis for imperfect reward exchange or clearly restrict the claims to the idealized setting. The lack of statistical testing is also a recurring issue in this paper and should be addressed before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read: this is a worthwhile extension of the authors' own decentralized-MAB work. The new piece is coordination: agents exchange rewards through MAPC and optimize PD and transmit power with AVG, MAX-MIN, or proportional-fairness rewards. The toy scenario cleanly shows the coordination benefit over selfish bandits, which get stuck in the conservative action, and the 9-BSS random deployment supports the direction. The paper is honest that coordination is modeled as perfect monitoring, and the evaluation uses their own Komondor simulator, which is established in this niche.\n\nSoft spots. First, the abstract states best-case numbers (15%, 210%, <3 ms) as if general. These come from one 9-BSS setup with 100 drops and no error bars or statistical tests, and the minimum and maximum curves in Figure 7 are across BSSs per drop, not across drops. Confidence intervals for the mean would help. Minor: Eq. (3) looks wrong dimensionally—r_k divided by r_k + 2? That is likely a typo in the reward update. Second, the central load-bearing assumption is perfect, lossless, zero-cost reward sharing. That is in Section III-B, and the authors defer overhead analysis. Since the claimed gains are precisely the coordination gains, this is the main unsupported premise. It may still be a reasonable idealization for a first study, but the paper should either model MAPC messaging cost/loss or explicitly scope the claims. Third, the baseline is a single OBSS/PD threshold (-72 dBm); a tuned or multi-threshold baseline would be a fairer comparison. The citation pattern is fine; prior work is cited and the incremental contribution is clear.\n\nWho this is for: people working on 802.11bn MAPC and AI/ML-native Wi-Fi. As a simulation study it is useful but not definitive; the numbers should be treated cautiously. I would send it to review, with the expectation that the authors address error bars, the Eq. (3) typo, and the coordination-channel assumption. The central argument holds up as a proof-of-concept; it just needs honest scoping.","headline":"A competent simulation study showing coordinated MABs can beat uncoordinated OBSS/PD SR in dense Wi-Fi, but the headline numbers assume a lossless, free coordination channel that the paper itself defers.","tokens_in":9181,"tokens_out":1714,"would_cite":true,"duration_ms":16542,"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":"Coordinated multi-agent bandits improve Wi-Fi spatial reuse by 15 percent in simulation, with the minimum throughput rising 210 percent and maximum access delay staying below 3 ms.","keywords":["multi-agent bandits","spatial reuse","Wi-Fi","MAPC","IEEE 802.11bn","reward sharing","throughput fairness","access delay"],"falsifier":"Simulate the same 9-BSS deployments while injecting realistic MAPC message loss, delay, or a nonzero cost for reward exchange; if the minimum-throughput and delay gains over OBSS/PD spatial reuse shrink or disappear as loss or delay grows, the central claim fails for real deployments.","tokens_in":8195,"feed_emoji":"📶","tokens_out":3650,"duration_ms":32575,"temperature":0.7,"pith_summary":"This paper argues that future Wi-Fi spatial reuse can be driven by coordinated multi-agent bandits rather than by explicit coordinated spatial reuse. When access points share reward information over the Multi-Access Point Coordination (MAPC) framework, bandit learning improves mean throughput by 15% and raises the minimum throughput by 210% compared with current OBSS/PD spatial reuse, while keeping the maximum access delay below 3 ms. The authors position this as a low-complexity, AI-native alternative to static coordinated spatial reuse proposals. A sympathetic reader should care because Wi-Fi 8 is actively considering coordinated spatial reuse, and this work shows a learning-based route to similar network-wide benefits without centrally solving a joint optimization.","feed_headline":"Coordinated bandits lift Wi-Fi spatial reuse by 15 percent","feed_subtitle":"Reward-sharing agents raise the worst-case throughput 210% and cap access delay under 3 ms in dense networks.","key_machinery":"The central object is a multi-agent multi-armed bandit system in which each access point is an agent that chooses an action pairing a packet-detect threshold and a transmit power, and receives a reward computed either selfishly or as a shared function of all agents' normalized throughputs. The shared-reward functions, namely average, max-min, and proportional fairness, are what turn independent learners into a coordinated system, because each agent's update uses the others' rewards over the MAPC link. The action-selection strategies ε-greedy and Thompson sampling determine how exploration is conducted, and the paper's claimed results follow from pairing these strategies with the right sharing rule.","core_discovery":"On its own terms, the paper establishes that coordinating bandit agents through shared rewards changes the equilibrium that spatial-reuse learning converges to. In a two-BSS toy deployment, uncoordinated selfish bandits get stuck selecting the most conservative transmit-power and packet-detect setting (A4) because aggressive settings only pay off when both agents choose them; coordinated agents with an average reward instead settle on the jointly optimal configuration (A1) most of the time. In random nine-BSS deployments, the same coordination improves the minimum throughput by up to 210% and caps maximum access delay at under 3 ms, where the OBSS/PD baseline reaches 58 ms. The paper concludes that explicit exploration (ε-greedy) with average or proportionally fair rewards is the most reliable combination, while Thompson sampling adapts better under the max-min reward.","pith_inferences":["If the lossless-reward assumption is relaxed, algorithms that tolerate missing or delayed reward updates would be the natural next step, and max-min sharing may degrade fastest because a single stale low reward punishes the whole group.","The same reward-sharing machinery could be applied to other Wi-Fi learning tasks, like channel selection or TXOP sharing, wherever MAPC already provides a communication channel.","A testable extension is to compare coordinated bandits against explicit coordinated spatial reuse under identical channel variability, since the paper compares against uncoordinated baselines only."],"forward_implications":["Coordinated bandits with shared rewards give a low-complexity alternative to explicit coordinated spatial reuse, requiring only reward exchange rather than joint parameter optimization.","The proportionally fair reward is a robust choice for improving the worst-off access points, raising minimum throughput by up to 210% in the simulated dense scenarios.","Keeping maximum access delay under 3 ms, versus 58 ms for OBSS/PD, suggests learning-based spatial reuse can support latency-sensitive applications in dense Wi-Fi.","Because the approach builds on the existing OBSS/PD mechanism and only adds MAPC-based reward sharing, it can be layered onto current hardware generations rather than waiting for Wi-Fi 8.","The choice of reward function changes the equilibrium reached: average reward maximizes mean throughput, while max-min helps fairness but can trap ε-greedy in a conservative weak equilibrium."],"supporting_citations":[{"why":"Supplies the OBSS/PD spatial-reuse baseline mechanism and the detailed simulation parameters that the coordinated MAB results are measured against.","marker":"[4]"},{"why":"Provides the multi-armed bandit formulation for decentralized spatial reuse that this paper extends with MAPC-based coordination.","marker":"[12]"},{"why":"Demonstrates the weak-equilibrium pitfall of selfish bandits in spatial reuse, which motivates the shared-reward coordination studied here.","marker":"[13]"},{"why":"Supplies the Wi-Fi simulator used for all performance evaluations in the paper.","marker":"[17]"}],"fun_headline_variants":["Wi-Fi bandits learn to share boosting worst-case throughput 210%","Coordinated MABs lift Wi-Fi spatial reuse: +15% throughput, +210% fairness","AI-native Wi-Fi: shared bandit rewards triple worst-case throughput","Bandit agents coordinate to cut Wi-Fi max delay from 58ms to 3ms"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that agents can exchange reward information instantly and without loss at the end of every learning step, so coordinated agents always act on perfect knowledge of other access points' performance.","fun_headline_variants_meta":{"raw":{"variants":["Wi-Fi bandits learn to share boosting worst-case throughput 210%","Coordinated MABs lift Wi-Fi spatial reuse: +15% throughput, +210% fairness","AI-native Wi-Fi: shared bandit rewards triple worst-case throughput","Bandit agents coordinate to cut Wi-Fi max delay from 58ms to 3ms"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000757,"raw_usage":{"total_tokens":3349,"prompt_tokens":915,"completion_tokens":2434,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":2346}},"tokens_in":531,"tokens_out":2434,"duration_ms":16531,"temperature":1.0,"reasoning_tokens":2346,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:48:03.195099+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the same 9-BSS deployments while injecting realistic MAPC message loss, delay, or a nonzero cost for reward exchange; if the minimum-throughput and delay gains over OBSS/PD spatial reuse shrink or disappear as loss or delay grows, the central claim fails for real deployments.","supporting_citations":[{"cited_title":"Spatial reuse in IEEE 802.11 ax WLANs,","cited_arxiv_id":null,"evidence_quote":"Supplies the OBSS/PD spatial-reuse baseline mechanism and the detailed simulation parameters that the coordinated MAB results are measured against."},{"cited_title":"Potential and pitfalls of multi- armed bandits for decentralized spatial reuse in WLANs,","cited_arxiv_id":null,"evidence_quote":"Provides the multi-armed bandit formulation for decentralized spatial reuse that this paper extends with MAPC-based coordination."},{"cited_title":"Collaborative spatial reuse in wireless networks via selfish multi-armed bandits,","cited_arxiv_id":null,"evidence_quote":"Demonstrates the weak-equilibrium pitfall of selfish bandits in spatial reuse, which motivates the shared-reward coordination studied here."},{"cited_title":"Komondor: A wireless network simulator for next-generation high-density WLANs,","cited_arxiv_id":null,"evidence_quote":"Supplies the Wi-Fi simulator used for all performance evaluations in the paper."}],"review_version":1}