{"id":"5692875f-e2bf-4c2d-be61-86c78cf91e03","arxiv_id":"2508.14679","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Multi-agent Q-learning routing fused with MERA and MST is claimed to improve node survival and battery balance in wireless sensor network simulations.","lead":"Sensor nodes in a wireless network act as learning agents that pick energy-efficient forwarding paths with Q-learning, and the learned paths are combined with two classical routing algorithms to balance battery drain. The abstract claims longer network life and more even battery use, but the body text is corrupted, so only the abstract-level claims could be checked.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Decentralized multi-agent Q-learning has no stated convergence guarantee; reported survival/SoC gains may be training-schedule artifacts rather than a stable learned policy.","rationale":"The reader's weakest assumption is decentralized convergence, and I agree that this is the most load-bearing gap. The central claim is empirical, and the downstream 'scalable and adaptive' assertion depends on the learned policy being meaningful. Since the body text is encoding-corrupted, no convergence proof, simulation setup, or Q-learning equations can be checked. The abstract explicitly allows decentralized learning, where multi-agent non-stationarity is a genuine technical issue, not merely a disagreement with consensus. I considered an alternative concern that the highest-SoC transmitter selection, rather than Q-learning, drives the gains; that is plausible but less load-bearing because the claim as written covers the whole hybrid system. The convergence issue is more fundamental: without a stable policy, the simulation outcome is not even well-defined. The paper provides no formal verification, code, or parameter-free derivation to independently support the result. The header-line inconsistency (arXiv:2508.14680v2 cond-mat.stat-mech vs. cs.NI metadata) is a provenance red flag but does not change the technical assessment. The verdict should remain UNVERDICTED, matching the reader's conclusion; no verdict adjustment is needed.","tokens_in":22660,"tokens_out":5199,"duration_ms":66926,"concrete_test":"Recover the original, non-corrupted PDF/source and inspect the Q-learning update (Section 4) and simulation protocol (Section 5). Then run the reported simulation with the same total episode budget but evaluate at 10%, 50%, and 100% of training, freezing policies at each point. If survival rate and SoC variance differ substantially across these checkpoints or oscillate, learning has not converged within the budget and the claimed gains are schedule artifacts. Additionally, run a decentralized-vs-cloud ablation with identical reward and exploration: one arm updates all Q-tables simultaneously; the other uses a single centralized controller. If decentralized performance is markedly worse or non-monotonic while centralized converges, the scalability/adaptivity claim fails specifically for the decentralized mode.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that simulations confirm the hybrid multi-agent Q-learning/MERA/MST router with highest-SoC transmitter selection improves survival and SoC variance. For this to be a property of the learned routing policy, the decentralized Q-learning must have settled on a stable policy by evaluation time. The abstract states each node is an autonomous agent that observes local state and selects actions, and learning occurs 'either in a decentralized manner or via a cloud-based controller.' In the decentralized mode, every node's Q-table is updated while all other nodes' policies are changing; this is a non-stationary stochastic game, and standard single-agent Q-learning convergence conditions (stationary MDP, all state-action pairs visited infinitely often, Robbins-Monro step sizes) do not apply. The available manuscript text provides no convergence argument, no epsilon/exploration schedule, no training-round budget, and no criterion for freezing policies before evaluation. If Q-values have not converged, the reported survival and SoC-variance numbers are functions of the training trajectory and random initialization, not of a well-defined routing policy. The cloud-based option might avoid non-stationarity, but the abstract does not state which mode the simulations used; if they used centralized learning, the 'each node as autonomous agent' narrative and the scalability/adaptivity claim are unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid routing method for WSNs that combines multi-agent Q-learning for cluster-head/route selection with two classical graph-based routing algorithms, MERA and MST, and selects transmitters by highest State of Charge (SoC). The reward is explicitly designed to incentivize balanced load distribution, hotspot avoidance, and energy-aware forwarding. The abstract claims that simulations show significant improvements in node survival rate, reduced SoC variance, and enhanced network resilience. However, the submitted full text after the abstract is almost entirely unreadable due to character-encoding corruption; equations, algorithm pseudocode, simulation tables, figures, and section content are not decipherable. Only the abstract and a few fragmentary captions are usable, so the central simulation-based claims cannot be technically verified from the manuscript.","tokens_in":22969,"tokens_out":5264,"duration_ms":71427,"significance":"If the claimed results were fully substantiated, a MARL-based routing policy that fuses learned decisions with classical graph methods could be a useful contribution to energy-efficient WSN routing. The intended formulation is coherent: local state features, a shaped reward, hybrid RL/MERA/MST fusion, and highest-SoC transmitter selection are all sensible components. The paper does not provide machine-checked proofs, reproducible code, or raw experimental data in the readable portion, and the abstract itself contains no concrete simulation numbers. The significance is therefore currently prospective: sound architecture, but no verifiable evidence that the proposed method outperforms baselines in practice.","major_comments":[{"comment":"The body of the manuscript is unreadable mojibake; all equations, algorithm boxes, tables, and figures after the abstract are unusable. I could not evaluate the Q-learning update, the exact reward function, the MERA/MST fusion rule, or any simulation configuration. Because the abstract's central claim is that 'simulations confirm' the method's benefits, this is a load-bearing presentation defect: the claim cannot be checked. The authors must resubmit a correctly encoded PDF and ensure every equation and table is readable.","section":"Main text (after abstract)"},{"comment":"No convergence or stability argument is provided for the decentralized multi-agent Q-learning setting. In the decentralized mode, each node updates its Q-table while all other nodes' policies are simultaneously changing, making the environment non-stationary; standard single-agent Q-learning convergence guarantees do not apply. The abstract does not state which learning mode the simulations used, what exploration schedule was followed, or whether policies were frozen before evaluation. Without this information, the reported survival and SoC-variance gains could be artifacts of the training schedule rather than properties of a stable learned policy. Provide convergence diagnostics or clearly use a centralized/cloud controller for the experimental results.","section":"Abstract (learning mode)"},{"comment":"The reward is described as deliberately incentivizing balanced load distribution, hotspot avoidance, and energy-aware forwarding, while the headline evaluation metrics are node survival rate and reduced SoC variance. Consequently, part of the reported improvement over classical baselines is by construction: a policy that optimizes these incentives should improve these same metrics. To demonstrate a genuine algorithmic advantage, report ablations with different reward weights, and compare against a parameter-matched non-RL baseline (e.g., greedy energy-aware routing) under the same energy and radio model.","section":"Abstract (reward design)"},{"comment":"The abstract's simulation claim is not accompanied by any quantitative details: no node counts, energy model, channel/traffic model, simulation duration, baseline algorithms, confidence intervals, or absolute lifetime/energy figures. Even if the body were fully readable, the abstract alone cannot support 'scalable and adaptive' conclusions. The revision must include a complete experimental section with network sizes spanning at least one order of magnitude, multiple random seeds, and error bars on the survival and SoC-variance metrics.","section":"Abstract (simulation claim)"}],"minor_comments":[{"comment":"State of Charge (SoC) is used without a formal definition. Specify whether it is normalized residual battery energy, a voltage-based estimate, or something else, and give units.","section":"Throughout"},{"comment":"The phrase 'learning occurs either in a decentralized manner or via a cloud-based controller' is ambiguous. Clarify whether these are two interchangeable modes, whether the simulations use both, and whether the reported results correspond to one mode or an average.","section":"Abstract"},{"comment":"The arXiv header lists cond-mat.stat-mech while the title and abstract are in cs.NI; the metadata should be corrected to avoid category mismatch.","section":"Metadata"}],"recommendation":"major_revision","confidential_remarks":"Editor: I could not perform a substantive technical review because the supplied full text is unreadable. If this is a conversion artifact, please ask the authors to resubmit a clean PDF and then re-review. My recommendation is based on the need to provide a readable, complete manuscript and to address the convergence, baseline, and experimental-detail concerns above; it is not a statement that the proposed ideas are necessarily wrong."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: I can't review this paper. The full text is encoding-corrupted — equations, simulation setup, results, and references are all mojibake. All I have is the abstract, one header line, and a table-of-contents skeleton. So any verdict has to be about the abstract alone, and the abstract doesn't give enough numbers to verify anything.\n\nWhat the paper does well, on paper: it's a sensible fusion of known pieces — multi-agent Q-learning for route selection, classical MERA/MST graph paths as a hybrid fallback, and a transmitter-selection rule based on highest remaining state of charge. That is a legitimate engineering combination in the WSN-lifetime literature, and if the simulations actually support it, it would be a useful incremental result. I'm not going to pretend it's a breakthrough; RL-based routing and cluster-head selection are decades old.\n\nThe soft spots are mostly consequences of not being able to read the body. The abstract says \"simulations confirm... significantly improves\" without node counts, baselines, energy model, or error bars. The reward function explicitly incentivizes balanced load, hotspot avoidance, and energy-aware forwarding, and the headline wins are reduced SoC variance and improved survival — so part of the benefit holds by construction. That's not necessarily fatal; it just means the honest claim should be \"this routing policy trades in a reasonable way,\" not \"nature gives us these gains.\"\n\nThe stress-test point about decentralized Q-learning is real. The abstract says learning happens either decentralized or via a cloud controller. If decentralized, every node updates its Q-table while others update theirs — a non-stationary game, and standard Q-learning convergence guarantees don't apply. The abstract offers no exploration schedule, convergence argument, or training-round budget. If the reported gains come from an unsettled training trajectory, they're artifacts. But again — I can't check because the body is unreadable.\n\nOne mechanical red flag: the body's header line shows \"arXiv:2508.14680v2 [cond-mat.stat-mech]\" — a different paper ID, and cond-mat, not cs.NI. That suggests a flawed PDF build or copy-paste error. The authors need to fix that before anyone can take the submission seriously.\n\nBottom line: this is not ready for peer review in this form. Send it back for a clean, readable resubmission. If you want a reading-group case study on how decentralized multi-agent RL claims can outrun their convergence guarantees, the abstract alone is a good start — but the paper itself, as submitted, doesn't deserve referee time.","headline":"The abstract describes a plausible RL-routing mashup, but the body is unreadable mojibake, so the claimed simulation wins are unverifiable.","tokens_in":23457,"tokens_out":2626,"would_cite":false,"duration_ms":28610,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a multi-agent Q-learning router with highest-SoC transmitter selection extends WSN lifetime and balances energy better than classical graph routing.","keywords":["wireless sensor networks","energy-efficient routing","multi-agent reinforcement learning","Q-learning","cluster-head selection","network lifetime","load balancing","IoT"],"falsifier":"Run the decentralized learning mode and log Q-values and selected actions per node across many random seeds; if policies keep oscillating, or if the survival-rate gain over MERA/MST disappears when all nodes update simultaneously (rather than using pre-trained or centrally computed policies), then the central claim fails.","tokens_in":22551,"feed_emoji":"📡","tokens_out":2589,"duration_ms":36947,"temperature":0.7,"pith_summary":"The paper is trying to establish that wireless sensor networks last longer and drain batteries more evenly when every sensor node learns its own routing decisions with Q-learning, instead of relying only on fixed graph-based routes. Each node acts as an autonomous agent, observing its residual energy, distance to the sink, hop count, and proximity to hotspots, then choosing forwarding actions that maximize a reward tied to energy efficiency. The learned decisions are fused with two classical routing schemes, MERA and MST, and the node with the highest remaining State of Charge is picked as transmitter to avoid draining any single node. Simulations are offered as evidence that this hybrid approach improves node survival rate, reduces SoC variance, and makes the network more resilient. If true, the method would give energy-constrained WSN deployments a self-adapting routing strategy that scales to dynamic IoT settings.","feed_headline":"Q-learning router extends sensor network lifetime","feed_subtitle":"Each node learns to route around hotspots, and simulations report better survival plus balanced battery drain.","key_machinery":"The central mechanism is a multi-agent Q-learning controller for routing, in which each sensor node is an independent learning agent. The action space is the choice of next-hop neighbor; the state includes residual energy, distance to sink, hop count, and hotspot proximity; and a designed reward function incentivizes balanced load distribution, hotspot avoidance, and energy-aware forwarding while preserving signal quality. The learned policy is fused with two classical graph-based routing algorithms, MERA and MST, and the final forwarding step selects the neighboring transmitter node with the highest remaining State of Charge. That combination — learned next-hop selection plus highest-SoC tr","core_discovery":"The central claim is that a multi-agent reinforcement learning approach for cluster-head selection and multi-hop routing can outperform classical energy-aware routing in wireless sensor networks. Every sensor is modeled as a Q-learning agent that uses local state information — residual energy, distance to sink, hop count, and hotspot proximity — to pick next-hop actions that maximize a long-term energy-efficiency reward. After paths are computed, data is forwarded through intermediate nodes to the transmitter with the highest remaining State of Charge, which prevents premature node depletion. The RL decisions are blended with Minimum Energy Routing Algorithm (MERA) and Minimum Spanning Tree","pith_inferences":["Editorial extension: if decentralized convergence can be formally shown or empirically stabilized, the same reward-and-selection design could transfer to mobile ad hoc networks and general IoT mesh networks with minimal changes.","Editorial extension: the highest-SoC transmitter rule is a simple greedy heuristic that depends only on neighbor state; combining it with learned next-hop selection might generalize to energy-harvesting nodes where remaining charge is stochastic.","Editorial extension: a direct testable comparison between decentralized and cloud-based training — tracking convergence speed and final policy quality across random seeds — would clarify when offloading computation actually pays off.","Editorial extension: the reward structure could be ported to other network-control tasks, such as mobile sink routing or UAV relay placement, wherever load balancing and hotspot avoidance are the chief objectives."],"forward_implications":["If the simulation results hold, WSN deployments can keep more nodes alive for the same traffic load, delaying network partition and extending useful sensing lifetime.","Lower SoC variance implies battery drain is spread more evenly, so early deaths from overused relay nodes become less common.","The adaptive Q-learning component lets routing respond to changing energy distributions, which matters when deployments are dynamic rather than static.","The optional cloud-based controller offers a path to offload learning computation in large-scale networks without changing the local reward logic.","The hybrid fusion with MERA and MST means the method can inherit classical routing guarantees while adding learned adaptability."],"supporting_citations":[],"fun_headline_variants":["Q-learning routes around hotspots to save sensor energy","Multi-agent RL balances battery drain and extends sensor life","Sensor nodes learn to dodge energy hotspots for longer survival","Reinforcement learning improves energy efficiency in sensor networks","Adaptive routing uses local energy data to keep sensors alive"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The decentralized training mode assumes the multi-agent Q-learning process settles on a stable policy within the simulation round budget, even though each node's environment changes as every other node learns, so textbook Q-learning convergence guarantees do not apply.","fun_headline_variants_meta":{"raw":{"variants":["Q-learning routes around hotspots to save sensor energy","Multi-agent RL balances battery drain and extends sensor life","Sensor nodes learn to dodge energy hotspots for longer survival","Reinforcement learning improves energy efficiency in sensor networks","Adaptive routing uses local energy data to keep sensors alive"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000635,"raw_usage":{"total_tokens":2775,"prompt_tokens":764,"completion_tokens":2011,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":1934}},"tokens_in":508,"tokens_out":2011,"duration_ms":18866,"temperature":1.0,"reasoning_tokens":1934,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T18:22:22.265746+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the decentralized learning mode and log Q-values and selected actions per node across many random seeds; if policies keep oscillating, or if the survival-rate gain over MERA/MST disappears when all nodes update simultaneously (rather than using pre-trained or centrally computed policies), then the central claim fails.","supporting_citations":[],"review_version":1}