{"id":"09dd4538-2823-4e42-8e6d-b00f9cc3d584","arxiv_id":"2607.05957","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":4,"one_line_summary":"A delay-aware multi-agent RL framework for active visual triangulation shows that Age-of-Information metadata, perception-consistent rewards, and multi-source covariance propagation each improve cooperative target localization under realistic communication delays.","lead":"This paper trains teams of camera-equipped drones to cooperatively locate a flying target by triangulation, while explicitly modeling the communication delays and sensor noise that real systems face. It shows that giving each drone a freshness timestamp on teammate data, using noise-aware reward signals, and modeling all uncertainty sources together each measurably improve tracking performance.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"Contribution 2's 'reward-observation alignment' interpretation is untested: the perception-consistent reward's advantage may stem from generic reward stochasticity rather than specific noise-observation correlation.","rationale":"The reader's concern about sim-to-real transfer and stationary delay models is valid but generic—it applies to any simulation-based RL paper and is explicitly acknowledged by the authors. My concern is more specific: Contribution 2's central interpretive claim ('reward-observation alignment as a design axis') is not causally isolated from a simpler alternative (reward stochasticity as exploration/regularization). The ablation is experimentally clean but mechanistically underdetermined. However, this concern affects the interpretation and practical recommendation, not the empirical validity of the measured outcomes. The ablation results in Table II stand regardless of which mechanism explains them. The reader's CONDITIONAL verdict (pending real-world validation) remains appropriate; this concern adds a caveat about Contribution 2's framing but does not change the verdict. The other two contributions (AoI and multi-source covariance) have cleaner causal stories: AoI provides directly measurable staleness metadata that recurrence cannot recover, and multi-source covariance captures more uncertainty terms by construction. These are less susceptible to alternative-explanation concerns.","tokens_in":10741,"tokens_out":6605,"duration_ms":444788,"concrete_test":"Run a third reward condition: clean delayed states (as in the privileged path) plus uncorrelated reward perturbations—e.g., multiplicative Gaussian noise on r_tri with standard deviation matched to the empirical reward variance of the perception-consistent path. Train under identical curriculum and evaluate on the same metrics. If this condition achieves RMSE and track loss within one standard error of the perception-consistent reward (0.547±0.217 m, 6.1 track losses), the alignment interpretation weakens substantially. If it performs closer to the clean reward (0.633±0.345 m, 8.3 track losses), alignment is supported as the operative mechanism.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper claims perception-consistent rewards outperform privileged rewards because 'the same configuration receives different reward values depending on its sensitivity to the noise actually present, so noise-fragile optima are penalized' (Section V-C), identifying 'reward-observation alignment as a design axis.' However, the noisy reward path introduces stochasticity into the reward signal that the clean path lacks. An alternative explanation is that this reward stochasticity acts as implicit exploration or regularization—preventing premature convergence to any fixed geometric equilibrium—regardless of whether the noise is correlated with the policy's observations. The paper does not test this alternative. The per-timestep analysis (Fig. 5) actually supports this alternative: the clean-reward policy converges to a 'stable geometric equilibrium' with 'monotonically decreasing RMSE,' while the noisy-reward policy 'lacks a fixed attractor' and shows 'gradual RMSE increase.' This pattern is consistent with reward noise disrupting convergence rather than specifically teaching noise-robustness. If the benefit is generic reward stochasticity, the practical recommendation changes from 'align reward with observation noise' (a specific architectural choice) to 'add reward noise or regularization' (a standard RL technique), weakening the novelty of Contribution 2. The ablation itself is clean—both conditions see identical observation noise and differ only in reward computation—but the causal mechanism attributed to the result is not isolated.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"This paper presents a delay-aware multi-agent reinforcement learning framework for active visual triangulation in Counter-UAS applications. The authors make three contributions: (1) a Dec-POMDP formulation with Age-of-Information (AoI)-augmented observations for handling stochastic communication delays, (2) a controlled comparison between perception-consistent (noisy) and privileged (clean) reward formulations, and (3) an extension of analytical triangulation covariance propagation to incorporate multi-source uncertainties (pixel, pose, gimbal, intrinsics). The framework is validated in Isaac Sim with 4096 parallel environments using MAPPO, with ablations isolating each contribution against external baselines (Gavin et al. angular-only covariance, no-AoI, MLP, clean-reward). The perception-consistent reward policy achieves 0.547 m RMSE with 78.1% triangulation validity.","tokens_in":10957,"tokens_out":1780,"duration_ms":280416,"significance":"The paper addresses a practically important problem: multi-agent visual triangulation under realistic communication delays. The ablation design is methodologically sound, with each contribution isolated by changing one variable while holding others constant, N>4000 episodes per condition, and 5th-95th percentile clipping. The covariance derivation (Eqs. 14-17) is a standard first-order linearization that is parameter-free given stated noise models. The reward comparison is a controlled experiment rather than a fitted prediction. The robustness envelopes (Figs. 6-7) test delay magnitude and target speed variation beyond training conditions. The honest reporting of the stability-robustness tradeoff in the reward comparison, including the gradual RMSE drift in the noisy-reward policy, is commendable.","major_comments":[{"comment":"Section V-C, Contribution 2 interpretation: The paper attributes the perception-consistent reward's advantage to 'reward-observation alignment' — specifically, that noisy rewards penalize noise-fragile optima because 'the same configuration receives different reward values depending on its sensitivity to the noise actually present.' However, an alternative explanation is that the stochasticity in the noisy reward path acts as implicit exploration or regularization, preventing premature convergence to any fixed geometric equilibrium — regardless of whether the noise is correlated with the policy's observations. The per-timestep analysis (Fig. 5) actually supports this alternative: the clean-reward policy converges to a 'stable geometric equilibrium' with 'monotonically decreasing RMSE,' while the noisy-reward policy 'lacks a fixed attractor' and shows 'gradual RMSE increase.' This pattern","section":null},{"comment":"is consistent with reward stochasticity disrupting convergence rather than specifically teaching noise-robustness. The paper does not test this alternative. A discriminating experiment would compare the perception-consistent reward against a clean-state reward with additive i.i.d. Gaussian noise matched in variance to the perception-consistent path's reward variance. If both yield similar aggregate metrics, the benefit is generic reward stochasticity rather than noise-observation correlation. Without this test, the 'reward-observation alignment as a design axis' claim (Section V-C, final paragraph) is not fully substantiated. This is load-bearing for Contribution 2's novelty framing.","section":null},{"comment":"Section V-C and Table II, clean-reward row: The clean-reward policy achieves higher triangulation validity (79.1% vs. 78.1%) and visibility (86.7% vs. 84.1%) than the noisy-reward policy, yet the paper frames the noisy-reward policy as superior in 'aggregate metrics.' The RMSE advantage (0.547 vs. 0.633 m) is the primary basis for this claim, but the clean-reward policy's higher validity and visibility suggest it maintains tracking more reliably. The paper should clarify which metrics constitute 'aggregate' superiority and acknowledge that the dominance is metric-dependent rather than uniform. The current framing in the abstract ('perception-consistent rewards outperform privileged clean-state rewards') overstates a mixed result.","section":null},{"comment":"Section VI, scaling beyond two agents: The paper notes that preliminary three-agent experiments show collision rates rising 9.5x. Given that multi-agent coordination is central to the paper's framing, this limitation is significant. The authors attribute it to insufficient collision-risk encoding in observations/rewards, but this suggests the reward formulation (Contribution 2) may not generalize beyond the minimal two-agent case. The paper should discuss whether the perception-consistent reward's advantage over the clean reward persists with more agents, or whether this is an artifact of the two-agent setting.","section":null}],"minor_comments":[{"comment":"Table II: The N values differ across conditions (4096, 4412, 4280, 4096, 6784). The MLP condition has substantially more episodes (6784), which the text explains is due to early termination, but the table caption or footnotes should clarify that N reflects completed/terminated episodes and that conditions with early termination accumulate more episodes within the same wall-clock budget.","section":null},{"comment":"Eq. (3): The AoI update rule sets Delta = d upon successful reception, but d is defined as end-to-end delay. If the delay includes both staleness and latency (as described in Section IV-A), then upon reception the AoI should equal the total delay experienced, which is correct. However, the notation could be clearer — d_k is used for delay in Eq. (2) and (3) but the three-stage pipeline (staleness, latency, dropout) is described only qualitatively. A brief equation showing how d_k composes from the three stages would improve reproducibility.","section":null},{"comment":"Section IV-B, Eqs. (14)-(17): The Jacobian J_theta is described as capturing sensitivity to 'nuisance parameters theta_i (pose, gimbal, intrinsics)' but the specific parameterization of theta_i is not given. For reproducibility, a brief enumeration of the components of theta_i and their noise standard deviations (referenced as matching Section IV-B) would help.","section":null},{"comment":"Fig. 5 caption: The caption states 'averaged across N > 4,000 episodes' but does not specify which condition. It should clarify that this is for the noisy-reward and clean-reward comparison.","section":null},{"comment":"Section III-C, Eq. (11): The reward weights (s_ctr, s_sz, lambda_ctr, lambda_sz) are listed as free parameters in the axiom ledger but their values are not reported in the paper. Including these in a table or appendix would improve reproducibility.","section":null},{"comment":"Section V-A: The curriculum schedule boundaries (Phase 1: 0-60k, Phase 2: 20k-100k, Phase 3: 80k-200k) overlap. This is presumably intentional (linear ramps), but should be clarified to avoid confusion about whether phases are sequential or overlapping.","section":null},{"comment":"Abstract and Section I: '0.547m vs.0.633m' should have a space after 'vs.' for consistency.","section":null},{"comment":"Section V-C, 'Non-stationary delays' paragraph: The claim that 'AoI would still indicate stale teammate information' under non-stationary channels is correct in principle, but the policy was trained under stationary distributions. The paper should note that the policy's response to AoI values outside the training distribution is not validated.","section":null}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about Contribution 2's interpretation (reward stochasticity vs. reward-observation alignment) is well-founded and is the primary reason for the major_revision recommendation. The paper's ablation is clean in isolating the reward computation path, but the interpretive claim goes beyond what the ablation demonstrates. The discriminating experiment (clean reward + matched-variance additive noise) is straightforward and would resolve the ambiguity. If the authors can show that perception-consistent rewards outperform noise-injected clean rewards, the 'reward-observation alignment' framing is validated; if not, the contribution narrows to 'reward stochasticity helps' which is less novel but still practically useful. The other two contributions (AoI, multi-source covariance) are well-supported by the ablations."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive review. The referee correctly identifies that the ablation design, covariance derivation, and honest reporting of tradeoffs are strengths of the paper. We address each major comment below and commit to revisions where the referee's points reveal genuine gaps in our analysis or framing.","responses":[{"response":"The referee raises a legitimate and incisive alternative hypothesis. We agree that the current manuscript does not present a discriminating experiment to distinguish 'reward-observation alignment' from 'generic reward stochasticity as regularization.' The referee's proposed control — a clean-state reward with additive i.i.d. Gaussian noise matched in variance to the perception-consistent path's reward variance — is exactly the right experiment to run. We will conduct this experiment and report the results in the revised manuscript. If the matched-variance noisy clean reward performs comparably to the perception-consistent reward, we will revise Contribution 2's framing accordingly, downgrading the claim from 'reward-observation alignment as a design axis' to 'reward stochasticity as a regularization mechanism.' If the perception-consistent reward outperforms the matched-variance control, this would support the alignment hypothesis. Either way, the current claim is overstated without this test, and we will present the result honestly.","revision_made":"yes","referee_comment":"Section V-C, Contribution 2 interpretation: The paper attributes the perception-consistent reward's advantage to 'reward-observation alignment' — specifically, that noisy rewards penalize noise-fragile optima because 'the same configuration receives different reward values depending on its sensitivity to the noise actually present.' However, an alternative explanation is that the stochasticity in the noisy reward path acts as implicit exploration or regularization, preventing premature convergence to any fixed geometric equilibrium — regardless of whether the noise is correlated with the policy's observations. The per-timestep analysis (Fig. 5) actually supports this alternative: the clean-reward policy converges to a 'stable geometric equilibrium' with 'monotonically decreasing RMSE,' while the noisy-reward policy 'lacks a fixed attractor' and shows 'gradual RMSE increase.' This pattern"},{"response":"This is a continuation of the previous comment and we treat it as the same point. We agree that the discriminating experiment is necessary and will include it. We also note that there is a theoretical reason to expect the two hypotheses to yield different outcomes: under the perception-consistent reward, the noise injected into the reward is correlated with the noise in the policy's observations — the same measurement perturbation affects both what the policy sees and what it is rewarded for. Under i.i.d. additive reward noise, the reward perturbation is independent of the observation noise. This structural difference means that perception-consistent rewards provide gradient information about which configurations are robust to the specific noise realization the policy is currently experiencing, whereas i.i.d. noise provides no such configuration-specific signal. However, we acknowledge this is a theoretical argument, not empirical evidence, and the referee is correct that without the experiment the claim is not substantiated. We will run the experiment and revise the framing based on the results.","revision_made":"yes","referee_comment":"is consistent with reward stochasticity disrupting convergence rather than specifically teaching noise-robustness. The paper does not test this alternative. A discriminating experiment would compare the perception-consistent reward against a clean-state reward with additive i.i.d. Gaussian noise matched in variance to the perception-consistent path's reward variance. If both yield similar aggregate metrics, the benefit is generic reward stochasticity rather than noise-observation correlation. Without this test, the 'reward-observation alignment as a design axis' claim (Section V-C, final paragraph) is not fully substantiated. This is load-bearing for Contribution 2's novelty framing."},{"response":"The referee is correct. The clean-reward policy does achieve higher triangulation validity (79.1% vs. 78.1%) and visibility (86.7% vs. 84.1%), and our use of 'aggregate metrics' and 'outperform' in the abstract overstates what is a mixed result. The perception-consistent reward's advantages are specifically in RMSE (0.547 vs. 0.633 m), track losses (6.1 vs. 8.3 per episode), and collisions (0.155% vs. 0.251%), while the clean-reward policy maintains better tracking continuity. We will revise the abstract and Section V-C to state this explicitly: the perception-consistent reward achieves lower RMSE and fewer track losses/collisions, while the privileged reward achieves higher validity and visibility. We will replace 'outperform' with a precise enumeration of which metrics favor each formulation and acknowledge that neither dominates. The phrase 'aggregate metrics' will be removed or replaced with the specific metrics referenced.","revision_made":"yes","referee_comment":"Section V-C and Table II, clean-reward row: The clean-reward policy achieves higher triangulation validity (79.1% vs. 78.1%) and visibility (86.7% vs. 84.1%) than the noisy-reward policy, yet the paper frames the noisy-reward policy as superior in 'aggregate metrics.' The RMSE advantage (0.547 vs. 0.633 m) is the primary basis for this claim, but the clean-reward policy's higher validity and visibility suggest it maintains tracking more reliably. The paper should clarify which metrics constitute 'aggregate' superiority and acknowledge that the dominance is metric-dependent rather than uniform. The current framing in the abstract ('perception-consistent rewards outperform privileged clean-state rewards') overstates a mixed result."},{"response":"We agree that the three-agent collision result is a significant limitation that the paper should discuss more thoroughly. To be transparent: we have not yet conducted the perception-consistent vs. clean-reward comparison with three agents, so we cannot state whether the advantage persists. The 9.5x collision rate increase was observed with the perception-consistent reward configuration only. The referee's concern that the reward formulation may not generalize is legitimate — the pairwise collision penalty (Eq. 12) does not scale gracefully to larger teams because it does not encode joint collision risk (multiple agents approaching the same region simultaneously). We will expand Section VI to explicitly state that the perception-consistent vs. clean-reward comparison has only been validated for two agents, that we cannot currently claim the advantage generalizes, and that the collision-encoding deficiency may interact with the reward formulation in ways we have not tested. We will also note this as a boundary on Contribution 2's scope rather than deferring it entirely to future work.","revision_made":"yes","referee_comment":"Section VI, scaling beyond two agents: The paper notes that preliminary three-agent experiments show collision rates rising 9.5x. Given that multi-agent coordination is central to the paper's framing, this limitation is significant. The authors attribute it to insufficient collision-risk encoding in observations/rewards, but this suggests the reward formulation (Contribution 2) may not generalize beyond the minimal two-agent case. The paper should discuss whether the perception-consistent reward's advantage over the clean reward persists with more agents, or whether this is an artifact of the two-agent setting."}],"tokens_in":10653,"tokens_out":1495,"duration_ms":190405,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"Two things to know: (1) this paper extends Gavin et al.'s multi-agent triangulation framework with three concrete, ablation-validated additions — AoI-augmented observations, multi-source covariance propagation, and a dual-path reward comparison — and the ablations are clean with large sample sizes. (2) The most interesting finding, that perception-consistent rewards outperform privileged clean-state rewards, is empirically solid but the causal explanation is not fully isolated, which matters because the practical recommendation changes depending on the mechanism at work. The stress-test note raises a legitimate concern here that I think partially lands. The paper attributes the perception-consistent reward advantage to noise-observation alignment — the idea that noise-fragile geometric optima get penalized because the reward sees the same noise the policy does. The alternative explanation is that the noisy reward path simply adds reward stochasticity that acts as implicit regularization, preventing premature convergence to any fixed equilibrium. The per-timestep analysis in Fig. 5 actually supports this alternative somewhat: the clean-reward policy converges to a stable equilibrium with monotonically decreasing RMSE, while the noisy-reward policy shows gradual drift and lacks a fixed attractor. That pattern is consistent with reward noise disrupting convergence rather than specifically teaching noise-robustness. The ablation itself is well-designed — both conditions see identical observation noise and differ only in reward computation — but the paper does not run the control that would distinguish these mechanisms (e.g., adding uncorrelated reward noise to the clean path). That said, the empirical result stands regardless of mechanism: perception-consistent rewards produce better aggregate metrics. The practical recommendation just shifts from a specific architectural principle to a more generic regularization insight. The other two contributions are straightforward and well-executed. The AoI ablation is convincing — 10.6 percentage point validity improvement, and the MLP baseline collapsing to 0.7% validity confirms recurrence is necessary. The multi-source covariance extension is a standard first-order linearization but a sensible one, and the 2.8-fold RMSE degradation from angular-only modeling is a meaningful quantitative result. The covariance derivation is parameter-free given stated noise models, which is good. The main limitation is that everything is simulation-only with stationary delay distributions. The paper acknowledges this honestly. The two-agent restriction with 9.5x collision rate at three agents is a real scalability concern but not a load-bearing flaw for what the paper claims. Reward weights and noise parameters are not fully specified, which limits reproducibility without contacting the authors. This paper is for researchers working on multi-robot perception, delay-aware RL, or C-UAS coordination. It deserves a serious referee who can push on the reward mechanism question and check whether the perception-consistent finding generalizes beyond this specific setup.","headline":"Solid empirical paper on delay-aware multi-agent triangulation; the perception-consistent reward finding is real but the causal mechanism is under-tested.","tokens_in":11501,"tokens_out":631,"would_cite":false,"duration_ms":125562,"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":"Stale Data Beats Clean Data for Training Drone Teams to Track Targets","keywords":[],"falsifier":"If, in real-flight experiments with onboard sensing and wireless networking, the perception-consistent reward policy fails to outperform the privileged-reward policy, or if AoI-augmented observations provide no measurable benefit over AoI-free recurrent policies, the paper's central design recommendations would not hold outside simulation. The paper itself acknowledges this: 'real-flight validation with onboard sensing and wireless networking remains necessary.'","tokens_in":10968,"feed_emoji":"🛰️","tokens_out":1004,"duration_ms":186340,"temperature":0.7,"pith_summary":"This paper tackles a practical bottleneck in using teams of camera-equipped drones to localize aerial targets: the observations each drone receives about its teammates are delayed, noisy, and sometimes dropped. The authors argue that three specific design choices, tested in controlled ablations, each measurably improve cooperative triangulation under these realistic conditions. First, tagging each inter-agent message with its Age-of-Information (AoI) -- a timestamp-derived measure of how stale the data is -- gives the policy explicit freshness metadata, improving the fraction of time steps with valid triangulation by 10.6 percentage points. Second, computing training rewards from the same noisy, delayed observations the policy actually sees (perception-consistent rewards) outperforms rewards computed from clean ground-truth states, yielding 14% lower localization error and 27% fewer track losses. This is counterintuitive: training against the true state should be better, but the authors show that clean-state rewards can steer policies toward geometric configurations that look optimal but are fragile to the noise the policy will actually face at deployment. Third, propagating uncertainty from all sources -- pixel detection, pose estimation, gimbal calibration, and camera intrinsics -- through the triangulation covariance, rather than modeling only angular noise, reduces localization error 2.8-fold. The paper also shows that recurrent policies (GRUs) are essential: feedforward MLPs with frame stacking achieve near-zero valid triangulation (0.7%) because they cannot adapt to the variable, non-uniformly spaced delays in the observation stream. The overall system, trained with MAPPO across 4096 parallel environments, achieves 0.547 m localization RMSE with 78.1% triangulation validity under 50-1000 ms communication delays.","feed_headline":"","feed_subtitle":"","key_machinery":"The paper's argument is carried by three mechanisms: (1) AoI-augmented observations, where each inter-agent message carries a directly measured staleness tag computed as current time minus message timestamp, enabling the recurrent policy to discount stale teammate information; (2) a dual-path delay architecture that separates the observation path (always noisy and delayed) from the reward path, allowing controlled comparison of perception-consistent vs. privileged rewards under identical observation conditions; and (3) multi-source analytical covariance propagation that linearizes the triangulation around pixel, pose, gimbal, and intrinsics uncertainties, yielding a closed-form 3x3 position","core_discovery":"The central finding is that when training multi-agent policies for tasks requiring geometric coordination under communication delays, aligning the reward signal with the policy's actual noisy perception yields more robust behavior than rewarding against ground-truth states. Clean-state rewards create a stability-robustness tradeoff: they produce temporally stable coordination but can guide policies toward noise-fragile geometric optima (e.g., near-degenerate ray intersection angles). Perception-consistent rewards implicitly penalize such fragility because the same configuration receives different reward values depending on its sensitivity to the noise actually present. This identifies reward","pith_inferences":[],"forward_implications":["If perception-consistent rewards generalize beyond triangulation, RL practitioners training multi-agent systems under sensor noise should reconsider the common practice of using privileged clean-state rewards during training, particularly when the deployment environment involves significant observation noise.","The finding that AoI tags provide information that recurrent memory alone cannot recover suggests that explicit freshness metadata should be a standard observation component in any multi-agent system with asynchronous, variable-latency communication, not just visual triangulation.","The 2.8-fold degradation from angular-only covariance modeling indicates that prior single-source uncertainty models may significantly underestimate the true error budget in multi-agent perception tasks, potentially invalidating coordination strategies that were optimized against incomplete noise models.","The stability-robustness tradeoff between perception-consistent and privileged rewards identifies a deployment-dependent design axis: missions requiring sustained long-horizon tracking may benefit from privileged rewards, while missions requiring fault tolerance benefit from perception-consistent rewards."],"fun_headline_variants":["Perception-consistent rewards beat clean-state rewards for multi-agent triangulation","Rewarding against noisy perception yields more robust multi-agent coordination","Delay-aware MARL for counter-UAS: perception-consistent rewards reduce track losses 27%","Clean-state rewards guide policies toward noise-fragile geometric optima","Recurrent memory essential for delay compensation in multi-agent triangulation"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The simulation's stochastic delay model uses stationary distributions for latency, staleness, and dropout. Real-world communication interference can be bursty and non-stationary, and if the actual delay distribution differs substantially from the simulated one, the learned policy may not transfer and the relative advantage of each design choice could change.","fun_headline_variants_meta":{"raw":{"variants":["Perception-consistent rewards beat clean-state rewards for multi-agent triangulation","Rewarding against noisy perception yields more robust multi-agent coordination","Delay-aware MARL for counter-UAS: perception-consistent rewards reduce track losses 27%","Clean-state rewards guide policies toward noise-fragile geometric optima","Recurrent memory essential for delay compensation in multi-agent triangulation"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":649,"prompt_tokens":556,"completion_tokens":93,"prompt_tokens_details":null},"tokens_in":556,"tokens_out":93,"duration_ms":24358,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T20:00:52.487799+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If, in real-flight experiments with onboard sensing and wireless networking, the perception-consistent reward policy fails to outperform the privileged-reward policy, or if AoI-augmented observations provide no measurable benefit over AoI-free recurrent policies, the paper's central design recommendations would not hold outside simulation. The paper itself acknowledges this: 'real-flight validation with onboard sensing and wireless networking remains necessary.'","supporting_citations":[],"review_version":1}