{"id":"7d231f34-7e55-4f9a-8731-a4cb24f5b432","arxiv_id":"2607.13880","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Learning when to communicate based on task value, not just link quality, improves simulated covert multi-AUV pursuit efficiency by roughly 20% over an IC3Net baseline.","lead":"This paper proposes SVR-MARL, a multi-agent reinforcement learning framework in which autonomous underwater vehicles decide when to transmit sensed information based on its estimated value to teammates, under realistic underwater acoustic and covert-detection constraints. A simulated four-AUV covert pursuit case study reports about 20% higher task efficiency than the IC3Net baseline.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 20% gain over IC3Net is not attributed to the value estimator: SVR-MARL also adds adaptive transmit power and constraint penalties, and no ablation isolates the novel mechanism.","rationale":"The reader's weakest assumption is that the simulator's communication and covert models are unfaithful to reality, making the 20% gain potentially an artifact of the simulator. That is a valid external-validity concern, but it applies to virtually any simulation result and is somewhat generic. I identify a more internal, directly testable gap: even taking the simulator at face value, the paper's central claim about value-driven scheduling is not supported because the experimental design confounds the value estimator with the newly introduced continuous power control. The comparison against IC3Net, which lacks power control, cannot attribute the improvement to the value mechanism. This is the single most load-bearing concern for the paper's stated contribution. The reader's verdict of CONDITIONAL remains appropriate—perhaps even more clearly so—since the missing ablation is a concrete additional condition for acceptance. I set agreement_with_reader to 'disagree' because we flag different weakest assumptions, though both point to insufficient empirical support.","tokens_in":10367,"tokens_out":4442,"duration_ms":51628,"concrete_test":"Train an SVR-MARL variant identical to the proposed method but with the SV-Est output replaced by a constant (e.g., a fixed 0.5) feeding the communication policy, while keeping the same gate-plus-power action space, reward penalties, and training procedure. Run both variants over multiple random seeds and report task efficiency with error bars. If the constant-value variant achieves statistically indistinguishable performance, the value estimator is not the source of the 20% gain and the central claim must be softened. If it drops significantly, the value mechanism is validated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that value-driven communication scheduling improves task efficiency by ~20% over IC3Net (Section IV.B). However, SVR-MARL differs from IC3Net in two major respects: (i) it adds the sensed-information value estimator (SV-Est, Section III.C.3) as an input to the communication policy, and (ii) it expands the communication action space from a binary gate to include transmit power drawn from a Beta distribution, with covert/reliability constraints folded into the reward via a Lagrangian method (Section III.C.5). IC3Net uses a fixed transmit power when its gate activates. Therefore the reported improvement could be entirely due to adaptive power control and explicit constraint penalization, not to the value estimator. The paper provides no ablation that isolates the value estimator—no variant with gate+power but without SV-Est, and no variant with SV-Est but fixed power. Without such an ablation, the central novelty claim ('value-driven communication scheduling') is not empirically supported; the result is consistent with a more mundane explanation: giving agents a continuous power knob helps them manage interference and covertness better than a fixed-power baseline. This is load-bearing because the paper's title and abstract foreground the value mechanism, and the comparison to IC3Net cannot distinguish it from the extra action dimension.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper addresses the gap between MARL with ideal communication and link-level communication optimization in covert multi-AUV collaboration. The authors propose SVR-MARL, a framework that decomposes each agent's policy into a communication phase and a task-execution phase, with a GRU-based encoder, a 'sensed information value estimator' (SV-Est), and a physics-grounded communication environment that models Thorp path loss, fading, propagation delay, overlap interference, decoding reliability, and KL-divergence-based covert detection. The framework is instantiated on a case study of four AUVs, one target, and one eavesdropper in a 3D pursuit task with bearing-only passive sonar, EKF state estimation, and PPO training. The paper reports that SVR-MARL reaches a task efficiency of about 1.2×10^-2, outperforming IC3Net by about 20%, and claims that value-driven scheduling improves localization and task completion efficiency while reducing unnecessary transmissions.","tokens_in":10705,"tokens_out":5131,"duration_ms":48394,"significance":"If the claimed result survives scrutiny, the paper makes a useful conceptual contribution by placing communication decision-making inside the loop of task value, physical-layer realism, and covert constraints—a combination rarely considered together in MARL. The framework is general and the case study is well motivated. However, the current evidence is too weak to establish the central claim: the evaluation is a single simulation with no error bars, no ablations, and no sensitivity analysis. The manuscript also leaves the notion of 'information value' under-specified. With a rigorous empirical evaluation (ablations, hyperparameters, seeds, sensitivity sweeps) and a formal definition of the value estimator's objective, the paper could become an interesting contribution to task-oriented communication and multi-agent learning.","major_comments":[{"comment":"The reported ~20% gain over IC3Net is the central quantitative claim, but SVR-MARL changes two things relative to IC3Net: it adds the SV-Est value estimator and replaces the binary gate/fixed-power communication with a continuous Beta-distributed power control plus Lagrangian constraint penalties. No ablation isolates the value estimator. Without variants such as (a) IC3Net-style gate+fixed power augmented with SV-Est and (b) SVR-MARL with the value-estimator input removed, the result is consistent with the explanation that adaptive power and explicit constraint shaping alone account for the improvement. The title/abstract foreground 'value-driven communication scheduling,' so this missing ablation is load-bearing.","section":"§IV.B and §III.C.3/5"},{"comment":"Task efficiency is reported as a single point estimate (≈1.2×10^-2; 'outperforming IC3Net by about 20%'). No standard deviation, number of seeds, training curves, or statistical test is provided, and no sensitivity analysis is given for the reward weights, Lagrangian penalty coefficients, EKF covariances, PPO hyperparameters, or the covertness KL threshold. The evaluation therefore does not establish that the improvement is robust, and the reader cannot reproduce or assess the significance of the claim. Provide at least mean±std over seeds and a table of hyperparameters.","section":"§IV.B"},{"comment":"The central concepts of the paper, 'value of sensed information (VSI)' and 'practical information value (PIV),' are described verbally but never given a formal definition or an equation. Specifically, the text says SV-Est supervision is 'based on changes in the receiver's state or task-related loss,' but the actual learning objective, the counterfactual benefit computation, and the communication reward function r_comm are unspecified. As presented, PIV appears to be the communication reward shaped from the same success signal (capture rate/task steps) used for evaluation, making it a reward-shaping term rather than an independently measured utility. This lack of formalization makes the framework's internal mechanism unfalsifiable and needs to be fixed.","section":"§III.C.3/5"},{"comment":"All performance results are generated inside the custom communication environment (Thorp path loss, fading, overlap interference, KL-divergence covert detection). No parameter sensitivity analysis or known-reference validation is provided for any of the acoustic or covertness parameters, including the detection threshold ε, transmit power bounds, noise levels, or interference model. Since the advantage of learned selective communication depends on the realism of Eve's detection statistics and the channel, the 20% improvement may be a simulator artifact. Add sensitivity sweeps over the key physical-layer parameters and compare at least one aggregate metric with an analytical benchmark.","section":"§III.C.4"}],"minor_comments":[{"comment":"The abstract uses 'practical information' while the body consistently uses 'practical information value'; standardize terminology throughout.","section":"Abstract"},{"comment":"The equations for f(γ_i) and D(Q_i^e||Q_i) appear only inside the figure; move them into the text and refer explicitly to them, since they are central to the covertness and reliability constraints.","section":"Fig. 3"},{"comment":"Reference [14] appears to have an author error ('Y. L. L. Brekhovskikh'); verify the correct author name and bibliographic details of the monograph.","section":"Reference [14]"},{"comment":"The passive sonar observation model is described qualitatively ('angular observation error increases with target distance'); state the exact sensor model and noise covariance used in the EKF.","section":"§IV.A"},{"comment":"It is unclear whether the 'True Obs' scheme uses the same movement policy as the other schemes or an oracle policy; clarify to interpret this upper bound correctly.","section":"§IV.B"}],"recommendation":"major_revision","confidential_remarks":"The missing ablation isolating the value estimator is the key technical issue. If the authors can provide that ablation, full hyperparameter tables, error bars, and a sensitivity analysis of the physical-layer model, the paper would be much stronger. The topic is suitable for the journal and the framework is promising, but the empirical evidence as it stands is insufficient to support the central claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper to know about: it extends MARL communication to realistic underwater acoustic channels and covert constraints by adding a learned value-of-sensed-information estimate that shapes both gating and transmit power. The problem is real, and the authors deserve credit for assembling the pieces—Thorp path loss, delay, packet overlap, decoding, and KL-divergence-based covertness—into a single framework that is clearly described. Table I is honest about prior work, and the case study is a sensible testbed for bearing-only pursuit.\n\nThe soft spot is exactly what the stress-test note says. SVR-MARL differs from IC3Net in two major ways: it adds the value estimator, but also swaps binary gating with fixed power for a continuous Beta-distributed power action plus Lagrangian penalty terms for covertness and reliability. The reported 20% improvement over IC3Net could come entirely from the extra control knob and the constraint penalties, not from the \"value-driven\" mechanism the title and abstract foreground. There is no ablation isolating SV-Est, and the paper does not report hyperparameters, reward weights, or statistical variance. That is load-bearing, since the central claim is that value-aware scheduling is what helps.\n\nI also share the reader's circularity concern: \"practical information value\" is defined through the same task-efficiency reward used for training and evaluation, so it is closer to reward shaping than an independently predicted benefit. That is not disqualifying for an engineering paper, but it should be said plainly, and the authors should at least discuss what would distinguish a genuine value estimator from a shaped reward.\n\nThe math and modeling are internally consistent, and I see no contradiction that would sink it. The citation pattern looks fair; related work is represented accurately. The paper just does not yet provide enough evidence for the headline claim.\n\nWho gets value from this? Researchers working on underwater multi-robot systems or MARL with physically grounded communication will appreciate the framework and the clearly identified gap, even if the empirical support is thin. It deserves a serious referee but needs major revision: add ablations that separate the value estimator from adaptive power/constraint penalties, release code and seed data, report error bars, and provide a full specification of the SV-Est objective and reward terms.\n\nMy recommendation: send it to peer review, but with a strong request for those ablations and reproducibility details before acceptance.","headline":"A well-motivated integration of task-oriented information value with realistic acoustic and covert constraints, but the 20% gain over IC3Net is not pinned on the value estimator—missing ablations and reproducibility details keep this conditional.","tokens_in":11202,"tokens_out":1327,"would_cite":false,"duration_ms":18133,"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":"In covert underwater pursuit, AUVs should transmit only when the sensed information's task value survives acoustic and covert constraints, not merely when a link exists.","keywords":["autonomous underwater vehicles","multi-agent reinforcement learning","covert communication","task-oriented communication","bearing-only localization","underwater acoustic channels","value of information","pursuit-evasion"],"falsifier":"A field trial or high-fidelity re-simulation using measured underwater acoustic propagation statistics and a realistic passive eavesdropper detector, comparing the value-driven policy against always-transmit and learned-gating baselines, would settle whether the reported 20% task-efficiency gain persists. Specifically, if the real eavesdropper's detection probability as a function of received power deviates strongly from the KL-divergence model, or if packet-overlap interference behaves differently in practice, the policy's transmission schedule could become either too aggressive or too conser","tokens_in":10259,"feed_emoji":"🌊","tokens_out":2824,"duration_ms":30857,"temperature":0.7,"pith_summary":"This paper argues that in covert multi-AUV cooperation, communication decisions should be driven by the task-level value of the sensed information after it has been degraded by real underwater acoustic channels and covert-detection constraints. It proposes SVR-MARL, a multi-agent reinforcement learning framework that decomposes each AUV's policy into a communication phase and a movement phase, and defines a 'value of sensed information' that is reshaped by propagation delay, interference, decoding reliability, and eavesdropper detection risk into a 'practical information value.' In a simulated bearing-only pursuit scenario, the paper claims this approach achieves a task efficiency of about 1.2 × 10^-2, roughly 20% higher than a learned-gating baseline and better than always-transmit or no-communication schemes. If correct, the work demonstrates that task-oriented covert communication is an information-selection problem, not just a link-optimization problem.","feed_headline":"Covert AUV teams gain 20% by transmitting only when it matters","feed_subtitle":"Task-aware acoustic messaging beats always-on and learned-gating baselines in simulated covert multi-AUV pursuit.","key_machinery":"The key machinery is the conversion of 'value of sensed information' (the potential collaboration benefit of a local message under ideal sharing) into 'practical information value' (the benefit after passing through a physics-grounded communication environment). This environment computes received power from the Thorp path-loss model and fading, models packet arrival windows and overlap interference, checks decoding via SINR, and enforces covertness through a KL-divergence detection constraint. The policy uses separate communication and task heads, a GRU-based temporal context encoder, a learned sensed-information value estimator, and PPO updates with Lagrangian penalty terms.","core_discovery":"The central claim is that a multi-AUV system can improve covert collaborative localization and pursuit by learning when to transmit and at what power, based on an estimate of how much the local sensing message will actually help teammates after passing through realistic acoustic links and covert constraints. The framework introduces a two-stage policy: a communication actor decides whether and how to transmit, while a task actor decides movement actions. The environment simulates Thorp-model path loss, large- and small-scale fading, propagation delays, packet-overlap interference, decoding outcomes, and a KL-divergence-based covert detection constraint. The communication reward captures coun","pith_inferences":["The same value-realization logic could generalize to other resource-constrained cooperative sensing domains, such as radio-frequency or optical networks, where transmission carries detection or energy costs.","A practical extension would be to test sensitivity to the eavesdropper's detection threshold: if Eve is weaker or stronger than modeled, the learned selective-transmission advantage may shrink or grow, suggesting an operating envelope for the policy.","The paper's framing implies that optimal communication in such settings is event-triggered rather than periodic, driven by spikes in estimation uncertainty or task-critical states.","The counterfactual reward design points toward a general principle for task-oriented communication: schedule transmissions by expected reduction in teammate uncertainty per unit exposure risk, rather than by information content or link quality alone."],"forward_implications":["Task-aware, physics-aware communication scheduling can outperform both always-transmit and learned-gating baselines in covert multi-AUV pursuit, with roughly a 20% task-efficiency gain.","Covert communication and task movement are coupled: task geometry affects exposure risk, so communication and movement policies should be learned jointly rather than designed separately.","Under realistic acoustic links, indiscriminate full communication is suboptimal because channel contention and decoding failures degrade efficiency.","Adaptive transmit power allocation uses the covert budget more effectively, increasing transmission investment when target uncertainty is high and reducing it when localization is confident.","The framework produces fully distributed policies, where each AUV decides based on local observations and received messages, avoiding reliance on a central scheduler."],"fun_headline_variants":["Learn when to ping: covert AUVs share less, track better","Sparsity wins: AUVs cut transmissions, boost pursuit","Covert AUV teamwork: transmit only if it helps teammates","SVR-MARL: AUVs decide what to share under acoustic limits","Task-aware acoustic chatter improves covert AUV pursuit"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The entire evaluation rests on the assumption that the custom simulator's underwater acoustic channel model (Thorp path loss, fading, packet-overlap interference) and the KL-divergence covert detection model faithfully represent real covert underwater acoustic links; if they do not, the learned policy's selective-transmission advantage is an artifact of the simulation.","fun_headline_variants_meta":{"raw":{"variants":["Learn when to ping: covert AUVs share less, track better","Sparsity wins: AUVs cut transmissions, boost pursuit","Covert AUV teamwork: transmit only if it helps teammates","SVR-MARL: AUVs decide what to share under acoustic limits","Task-aware acoustic chatter improves covert AUV pursuit"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000187,"raw_usage":{"total_tokens":1154,"prompt_tokens":721,"completion_tokens":433,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":465,"completion_tokens_details":{"reasoning_tokens":342}},"tokens_in":465,"tokens_out":433,"duration_ms":4468,"temperature":1.0,"reasoning_tokens":342,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T03:25:08.249917+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A field trial or high-fidelity re-simulation using measured underwater acoustic propagation statistics and a realistic passive eavesdropper detector, comparing the value-driven policy against always-transmit and learned-gating baselines, would settle whether the reported 20% task-efficiency gain persists. Specifically, if the real eavesdropper's detection probability as a function of received power deviates strongly from the KL-divergence model, or if packet-overlap interference behaves differently in practice, the policy's transmission schedule could become either too aggressive or too conser","supporting_citations":[],"review_version":1}