{"id":"8bf20e06-fffc-4c43-85b8-1d70c912f19c","arxiv_id":"2607.07045","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":7,"one_line_summary":"Decomposing Wi-Fi MAC control into staged contention-then-aggregation decisions improves throughput in mixed legacy/MLO networks compared to flat one-shot policies.","lead":"EvoOMG is a multi-agent reinforcement learning framework that optimizes Wi-Fi throughput in networks where legacy single-link devices coexist with new multi-link (MLO) devices. It decomposes MAC-layer decisions into ordered stages—contention first, then aggregation—matching protocol causality and preserving distinct device behaviors.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The paper's own ablation (Fig. 10) shows removing the autoregressive staged decomposition costs only 0.5% throughput, undermining the central claim that staged decomposition is structurally superior.","rationale":"The reader correctly identified the lack of statistical significance and the unverified theoretical assumptions as concerns, but missed the more fundamental issue visible in the paper's own ablation: the central architectural innovation (autoregressive staged decomposition) contributes almost nothing to the headline throughput gains. The reader's focus on Theorem 3's Lipschitz assumption is somewhat misplaced because Theorem 3 is a standard supporting bound, not the core claim — and Theorem 2 (staged optimal factorization) is exact and uncontroversial. The real problem is empirical: the ablation that isolates the staged decomposition (w/o AR Gen) shows negligible throughput difference (0.5%), while the gains over MADDPG are driven by orthogonal components (centralized critic, device embedding, history encoding) that the baseline lacks. This means the comparison with MADDPG is confounded — MADDPG is missing multiple components, not just the staged decomposition. The paper would need either (a) a fair comparison where the baseline also has Transformer encoding, device embedding, and centralized critic but uses flat actions, or (b) a demonstration that the staged decomposition provides statistically significant gains in fairness or other metrics. The existing w/o AR Gen ablation already provides comparison (a) and shows no meaningful difference. This keeps the verdict at CONDITIONAL rather than ACCEPT: the engineering contribution (combining multiple known techniques for heterogeneous Wi-Fi) may be valuable, but the central theoretical and architectural claim about staged decomposition is not validated by the paper's own evidence.","tokens_in":19547,"tokens_out":1755,"duration_ms":72343,"concrete_test":"Run the 'w/o AR Gen' ablation and the full EvoOMG model across at least 5 random seeds with different mobility and traffic patterns. Report throughput and Jain fairness with 95% confidence intervals. If the confidence intervals overlap (which the single-run 1463.9 vs 1471.5 result suggests they will), the claim that autoregressive staged decomposition is structurally superior to flat action generation is not supported by the paper's own evidence.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central methodological contribution is the autoregressive staged decomposition: generating contention guidance first, then aggregation guidance conditioned on it (Eqs. 22-24). The claim is that this is 'structurally superior' to flat one-shot action mappings. However, the ablation in Fig. 10 directly tests this: the 'w/o AR Gen' variant — which replaces the staged autoregressive generation with a flat joint action while keeping all other components (Transformer encoder, device embedding, centralized critic, feasibility masking) — achieves 1463.9 Mbps vs. 1471.5 Mbps for the full model. This is a 0.5% difference, well within noise for a single-run NS-3 evaluation with no confidence intervals. The paper notes fairness is 'lower' for w/o AR Gen but does not quantify this meaningfully, and the Jain fairness bars in Fig. 10 appear nearly identical (~0.73 vs ~0.74). Meanwhile, the components that actually drive the ~29% gain over MADDPG are the centralized critic (w/o Cent-Critic: 1106.0 Mbps), device embedding (w/o Dev-Embed: 1301.7 Mbps), and history encoding (w/o History: 1344.4 Mbps). The baseline MADDPG lacks all of these, making the comparison confounded. The paper's title, abstract, and theoretical framework (Theorem 2) all center on the staged decomposition, but the empirical evidence shows it is not the load-bearing mechanism behind the performance gains.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"The paper proposes EvoOMG, a multi-agent deep reinforcement learning framework for optimizing MAC-layer throughput in heterogeneous Wi-Fi networks where legacy (non-MLO) STAs and MLO-capable (Wi-Fi 7/8) STAs coexist. The core methodological contribution is an autoregressive, staged action generation process: agents first produce contention-window guidance, then produce aggregation-length guidance conditioned on the contention output, subject to standard-specific feasibility constraints (Eqs. 22-24). The system is evaluated in NS-3 against MADDPG, IDDPG, and static baselines across various legacy/MLO deployment ratios, traffic classes, and STR/NSTR PHY constraints, showing approximately 29% throughput gains. A theoretical analysis provides a performance gap bound (Theorem 3) under Lipschitz continuity assumptions.","tokens_in":19805,"tokens_out":1571,"duration_ms":171859,"significance":"The problem of heterogeneous legacy/MLO coexistence is timely and practically important for Wi-Fi 7/8 deployments. The NS-3 evaluation is commendably thorough, covering multiple deployment ratios, AC-level traffic differentiation, STR/NSTR constraints, latency CDFs, and an ablation study. The formulation of the MAC action space to respect protocol causality and device heterogeneity is a reasonable architectural choice. The inclusion of a theoretical performance bound, while relying on standard assumptions, adds analytical depth.","major_comments":[{"comment":"The central claim of the paper is that the autoregressive staged decomposition (contention-then-aggregation) is structurally superior to flat one-shot action mappings. However, the ablation study in Fig. 10 directly contradicts the magnitude of this claim. The 'w/o AR Gen' variant, which removes the staged decomposition and uses a flat joint action while retaining all other components, achieves 1463.9 Mbps compared to 1471.5 Mbps for the full model—a difference of only 0.5%. The paper states that fairness is 'lower' for w/o AR Gen but does not quantify this, and the Jain fairness bars in Fig. 10 appear nearly identical. Meanwhile, the components driving the ~29% gain over MADDPG are the centralized critic (w/o Cent-Critic: 1106.0 Mbps), device embedding (w/o Dev-Embed: 1301.7 Mbps), and history encoding (w/o History: 1344.4 Mbps). The baseline MADDPG lacks all of these, making the headto","section":null},{"comment":"head-to-head comparison confounded. The paper's title, abstract, and theoretical framework (Theorem 2) all center on the staged decomposition, but the empirical evidence shows it is not the load-bearing mechanism. The authors must either (i) provide a fair comparison where the baseline MADDPG is augmented with the Transformer encoder, device embedding, and centralized critic (leaving only the staged decomposition as the variable), or (ii) reframe the paper's central contribution to accurately reflect that the performance gains stem primarily from the combination of history encoding, device-aware heads, and centralized training, rather than the autoregressive decomposition. As it stands, the empirical evidence does not support the paper's central thesis.","section":null},{"comment":"The evaluation lacks confidence intervals or multiple-seed runs. All results in Figs. 4-10 appear to be from single-run NS-3 evaluations. Given that the difference between the full model and the 'w/o AR Gen' ablation is only 0.5% (7.6 Mbps), and that Wi-Fi MAC dynamics are inherently stochastic, it is impossible to determine whether this difference is statistically significant. The 29% gain over MADDPG is likely significant, but the relative contributions of individual components (especially the staged decomposition) cannot be assessed without variance estimates. Please provide results averaged over multiple seeds with confidence intervals or standard deviations.","section":null},{"comment":"The theoretical analysis (Theorem 3, Eq. 48) provides a performance gap bound under Assumption 1 (Lipschitz continuity of Q*) and Assumption 2 (bounded approximation errors). While these are standard in approximate RL theory, they are not verified for the specific NS-3 Wi-Fi MAC dynamics. The protocol-induced transition kernel P (Eq. 13) involves discrete CW sets, hard NSTR constraints, and threshold-based backoff, which can plausibly produce discontinuous value functions, violating the Lipschitz assumption. The bound is therefore not guaranteed to hold. The authors should either verify the Lipschitz property empirically (e.g., by measuring the sensitivity of Q* to perturbations in actions/states) or explicitly acknowledge this limitation and clarify that the theoretical analysis is motivational rather than guaranteeing the bound in practice.","section":null}],"minor_comments":[{"comment":"Fig. 5: The throughput numbers in the text do not match the bar chart values. For the MLO-dominant case, the text states EvoOMG reaches 1111 Mbps, but the figure shows 1117 Mbps. Please reconcile.","section":null},{"comment":"Fig. 6(a): The text states EvoOMG achieves 1482.7 Mbps, but the figure label shows 1482.8 Mbps. Minor inconsistency.","section":null},{"comment":"The paper mentions 'optional federated aggregation' (Section III-D) but does not evaluate it. If it is not evaluated, consider removing the detailed federated training equations (Eqs. 26-31) or moving them to an appendix to streamline the presentation.","section":null},{"comment":"Eq. (14): The notation for the previous MAC-action vector ψ_{i,a,k}(t-1) includes ρ, z, c, and l, but the text later says ρ and z are determined by feasibility projection rather than being learned. Clarify whether ψ includes the projected or raw values.","section":null},{"comment":"The paper uses 'EvoOMG' and 'Proposed' interchangeably in figures. Consistency would improve readability.","section":null},{"comment":"Reference [25] is cited for the path loss model, but the parameters (η, X_σ, d_0) are not specified in Table I or the text. Please provide these values for reproducibility.","section":null},{"comment":"The paper claims EvoOMG 'improves convergence stability' (abstract), but Fig. 4(a) shows the reward oscillating between 2.9 and 3.1, which is not particularly stable. Please clarify what 'stability' means here relative to the baselines.","section":null}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about the ablation study is valid and is the most serious issue. The paper's framing centers on the autoregressive staged decomposition, but the authors' own ablation shows it contributes negligibly to throughput. The real drivers are the centralized critic and device embedding. This is a framing problem: the paper may be 'overselling' the staged decomposition. If the authors reframe the contribution as a holistic framework (history + device-aware + centralized + staged) rather than centering the staged decomposition, the paper could be acceptable. The theoretical analysis is correct but somewhat disconnected from the empirical results—it does not explain why the staged decomposition helps (or doesn't). The lack of multiple-seed runs is also a concern for a journal submission."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"The headline you need: this paper's own ablation undercuts its central claim. The autoregressive staged decomposition — contention first, then aggregation — is pitched as the core contribution, but removing it costs only 0.5% throughput (1463.9 vs 1471.5 Mbps in Fig. 10). The real work is done by the centralized critic, device embedding, and history encoder, which are standard MADRL components. That's the thing to know before reading further. What is genuinely good: the problem formulation is well-motivated. Mixed legacy/MLO coexistence in Wi-Fi 7/8 is real and underexplored. The system model captures STR/NSTR constraints, TID-to-link mapping, and per-AC scheduling at the right granularity. The NS-3 evaluation covers multiple deployment ratios, AC-level differentiation, and PHY constraint modes. The standard-aware feasibility masking for heterogeneous action spaces is a reasonable engineering contribution. The 29% gain over MADDPG in the main comparison is real in direction, though reported without error bars or multi-seed runs, so I can't tell if it's robust. The soft spots, in order of severity: First, the stress-test concern lands. The paper's title, abstract, and Theorem 2 all center on staged decomposition being structurally superior. But the ablation shows it contributes almost nothing to throughput. The paper hand-waves that w/o AR Gen has 'lower fairness' but the Jain index bars look nearly identical (~0.73 vs ~0.74) and are not quantified. This is a load-bearing gap between claim and evidence. Second, the theoretical bound (Theorem 3) relies on Lipschitz continuity of Q* and bounded approximation errors — standard assumptions in approximate RL theory, but unverified for discrete Wi-Fi MAC dynamics where CW sets, NSTR constraints, and threshold-based backoff can produce discontinuities. The bound is not wrong, just not informative for this specific setting. Third, no code or data shipped, and no confidence intervals anywhere. For a paper whose key ablation difference is 0.5%, that matters. Who gets value: researchers working on learning-based Wi-Fi MAC optimization, especially MLO scheduling. The system model and NS-3 setup are useful references even if the staged decomposition thesis doesn't hold up. Recommendation: deserves a serious referee because the problem is important and the evaluation is substantive, but the referee needs to push hard on the ablation discrepancy. The authors should either reframe the contribution around the components that actually drive gains, or provide stronger evidence (multi-seed, fairness quantification) that staged decomposition matters. I'd lean conditional accept pending that revision.","headline":"EvoOMG: staged Wi-Fi MAC decomposition that its own ablation shows barely matters","tokens_in":20311,"tokens_out":1327,"would_cite":false,"duration_ms":67781,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Splitting Wi-Fi MAC decisions into stages yields 29% throughput gain","keywords":["Wi-Fi 7","multi-link operation","heterogeneous networks","multi-agent reinforcement learning","autoregressive policy","MAC layer optimization","contention window","frame aggregation"],"falsifier":"If a flat one-shot policy with sufficient capacity and standard-aware action masking matched or exceeded the staged autoregressive policy in NS-3 throughput under identical mixed-standard deployments, the central claim that staged decomposition is structurally superior would be undermined","tokens_in":19588,"feed_emoji":"📶","tokens_out":2186,"duration_ms":77885,"temperature":0.7,"pith_summary":"The paper argues that Wi-Fi throughput optimization in networks where legacy single-link stations coexist with MLO-capable multi-link stations should be reformulated as a staged decision process rather than a flat one-shot action mapping. The central mechanism is an autoregressive policy that first generates contention-window guidance and then produces aggregation-length guidance conditioned on the contention output, following the natural protocol order of contention before transmission. This factorization, combined with standard-aware policy heads that distinguish legacy and MLO action spaces, allows a unified multi-agent framework to respect both device heterogeneity and protocol causality. The authors prove that the staged decomposition is lossless relative to joint optimization under the exact action-value function, derive a performance gap bound showing each component reduces a specific approximation error, and demonstrate approximately 29% throughput gains over flat MADDPG baselines in NS-3 simulations.","feed_headline":"Splitting Wi-Fi MAC decisions into stages yields 29% throughput gain","feed_subtitle":"Decomposing MAC decisions by protocol order outperforms flat one-shot policies in heterogeneous legacy/MLO networks.","key_machinery":"The autoregressive staged policy factorization pi_i(a_i | O_i, d_i) = pi_i^cw(c_i | O_i, d_i) * pi_i^tr(u_i | c_i, O_i, d_i), which splits the MAC action into a contention-window stage and a transmission stage, conditioned on a device descriptor d_i that encodes whether the station is legacy single-link or MLO-capable. A Transformer-based history encoder processes recent channel, queue, contention, and transmission states, and standard-aware feasibility projection masks invalid actions for each device type.","core_discovery":"The performance gap in mixed legacy-and-MLO Wi-Fi networks stems not from function approximation difficulty alone but from a structural mismatch between flat one-shot MAC policies and the sequential, standard-constrained nature of Wi-Fi MAC operations. Decomposing the joint MAC action into a contention stage followed by a transmission stage—conditioned on the preceding access outcome and device-specific feasibility constraints—is provably lossless under the exact value function and empirically yields substantial throughput gains over flat baselines.","pith_inferences":[],"forward_implications":["Mixed-standard Wi-Fi deployments can be optimized within a single multi-agent framework without sacrificing protocol compliance for either legacy or MLO devices","The staged decomposition principle could extend to other protocol-layer optimization problems where decisions have natural temporal ordering and device-specific feasibility constraints","The performance gap bound suggests that improving history encoding, staged generation, and standard-aware masking each independently contribute to closing the heterogeneity gap","The federated aggregation extension indicates the framework can scale across multiple Wi-Fi domains with different legacy/MLO ratios without cross-standard drift"],"fun_headline_variants":["Staged MAC decisions boost throughput in mixed legacy/MLO Wi-Fi","Multi-agent staged MAC guidance outperforms flat policies in Wi-Fi","Decomposing Wi-Fi MAC into contention then transmission lifts throughput","Evolution-oriented multi-agent framework lifts mixed Wi-Fi goodput","Autoregressive MAC guidance improves mixed legacy/MLO Wi-Fi throughput"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The theoretical performance bound relies on the optimal action-value function being Lipschitz continuous, which is a standard assumption in approximate reinforcement learning but is not verified for the discrete, threshold-based dynamics of the NS-3 Wi-Fi MAC stack. If the protocol's transition dynamics produce discontinuous value functions, which is plausible given discrete contention windows and hard NSTR constraints, the bound does not hold, though the empirical results do","fun_headline_variants_meta":{"raw":{"variants":["Staged MAC decisions boost throughput in mixed legacy/MLO Wi-Fi","Multi-agent staged MAC guidance outperforms flat policies in Wi-Fi","Decomposing Wi-Fi MAC into contention then transmission lifts throughput","Evolution-oriented multi-agent framework lifts mixed Wi-Fi goodput","Autoregressive MAC guidance improves mixed legacy/MLO Wi-Fi throughput"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":659,"prompt_tokens":587,"completion_tokens":72,"prompt_tokens_details":null},"tokens_in":587,"tokens_out":72,"duration_ms":34613,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T21:01:42.607598+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If a flat one-shot policy with sufficient capacity and standard-aware action masking matched or exceeded the staged autoregressive policy in NS-3 throughput under identical mixed-standard deployments, the central claim that staged decomposition is structurally superior would be undermined","supporting_citations":[],"review_version":1}