{"id":"3ea8234f-9812-490e-9988-558c8a7391bd","arxiv_id":"2607.18604","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A cloud/edge LLM hierarchy that re-weights a fast UAV controller's reward function is claimed to reduce collisions and improve throughput, but the simulations do not directly plot those two headline metrics.","lead":"Large language models on a ground station and on each drone team up with a fast reinforcement-learning controller to guide delivery drones through 3D sky highways. The paper simulates this 'cloud-edge agentic' design and reports safer, better-connected flights, but the headline claims of lower collision rates and higher throughput are not directly plotted.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract claims collision-rate and throughput gains are never directly measured; Fig. 3 shows only proxy rewards with no error bars or significance tests, leaving the strongest claim unsupported.","rationale":"The reader's weakest_assumption focused on non-stationary reward weights causing potential DDQN instability. That is a legitimate theoretical risk, but the more immediate and decisive gap is that the abstract's quantified claims (collision rate, aggregate throughput) are never directly measured in the reported results. The paper only presents proxy rewards and survival steps, without error bars or statistical tests. If the direct metrics were supplied and showed significant improvements, the non-stationary reward concern might be empirically mitigated; conversely, even a perfectly stable training process would not justify the headline claim if the wrong metrics are reported. Thus the missing metrics are more load-bearing. The reader's verdict of CONDITIONAL remains appropriate: the architecture is promising but requires a strengthened evaluation. We do not escalate to REJECT because the observed proxy improvements are consistent with the claim and the problem is fixable with additional experiments.","tokens_in":8643,"tokens_out":3082,"duration_ms":32796,"concrete_test":"Re-run the three methods (DDQN, Envelope MORL, Proposed Framework) in the same simulator for at least 10 random seeds, and report direct metrics: (i) collision rate (e.g., collisions per 1,000 flight-seconds) and (ii) aggregate system throughput (sum of per-UAV achieved rates in Mbps) for M=10, 20, 30. Compute means and 95% confidence intervals. If the proposed framework's collision rate is not significantly lower and aggregate throughput not significantly higher than both baselines (with non-overlapping CIs), the abstract's central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest claim is that the proposed architecture 'significantly reduces collision rates and improves aggregate system throughput' (Abstract). However, the numerical results in Section V-B and Fig. 3 do not report collision rates or aggregate throughput at all. Instead, Fig. 3a-b show 'transportation reward' and 'survival step count,' and Fig. 3c-d show 'communication reward' and 'handover probability.' Survival step count is only a proxy for collision avoidance and can be inflated by conservative flight behavior that reduces throughput; communication reward is a weighted sum that includes the handover penalty term, not the actual aggregate throughput in Mbps. Additionally, no error bars, confidence intervals, or multi-seed statistics are presented, so 'significantly reduces' is not statistically supported. Consequently, even if the non-stationary reward-weight tuning is stable, the central empirical claim is unverified because the decisive metrics are missing. The reader's concern about DDQN stability is plausible, but the measurement gap is more load-bearing: without direct collision-rate and throughput measurements, there is no way to confirm the headline result.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hierarchical LLM-driven control framework for multi-UAV navigation in ITNTNs. A cloud-based LLM on a HAPS performs slow-timescale global load balancing, while lightweight edge-LLMs on UAVs periodically adjust the reward weights of a fast-timescale DDQN controller. The problem is formulated as a hierarchical multi-objective POMDP. Simulations in a gym-pybullet-drones environment compare the proposed framework against DDQN and Envelope MORL baselines. The abstract claims 'significantly reduces collision rates and improves aggregate system throughput,' but the reported results are proxy metrics only.","tokens_in":8990,"tokens_out":3365,"duration_ms":40527,"significance":"If the claimed gains are real, the architecture would demonstrate a practical way to inject strategic, semantic reasoning into real-time multi-UAV control without violating latency constraints. The timescale decomposition (HAPS at 5 s, edge-LLM at 1 s, DDQN at 0.05 s) is well motivated, and the use of a high-fidelity physics simulator is a strength. However, the current evidence does not support the headline empirical claims, and the non-stationary reward tuning raises unresolved stability concerns.","major_comments":[{"comment":"The abstract claims that the architecture 'significantly reduces collision rates and improves aggregate system throughput,' but neither metric is directly measured or reported. Fig. 3 shows total reward, survival step count, communication reward, and handover probability. Survival step count is only a proxy for collision avoidance and can increase under conservative flying that sacrifices throughput; communication reward is a weighted combination that includes the handover penalty, not throughput in Mbps. No error bars, confidence intervals, or multi-seed statistics are provided, so 'significantly' is not supported. The paper must report direct collision rates (e.g., per episode or per flight hour) and aggregate throughput (e.g., bps/Hz or Mbps), with multiple random seeds and statistical significance tests.","section":"Abstract and Section V-B, Fig. 3"},{"comment":"The edge-LLM dynamically adjusts the reward weights α_{1,2,3} every T_LLM = 1 s while the DDQN continues training on the resulting reward. This makes the reward non-stationary, and standard DDQN convergence guarantees no longer apply. The paper offers no convergence analysis and no ablation against fixed weights or a simple rule-based weight scheduler. Without this comparison, the gains in Fig. 3 could be attributed to reward shaping rather than to LLM reasoning. Please provide learning curves under fixed weights, a hand-crafted scheduling rule, and the proposed LLM-based tuning, with multiple seeds.","section":"Section IV-B, Eq. (7)"}],"minor_comments":[{"comment":"The number of UAVs is reported inconsistently: the text says M ∈ {10,20,30}, while Table I lists M ∈ {5,10,15,20,25,30}. Clarify which set is used for the experiments in Fig. 3.","section":"Section V-A and Table I"},{"comment":"The DDQN baseline is cited to [4] and [15], but [15] appears to be a paper on programmable data-planes, unrelated to deep Q-networks. This citation seems erroneous.","section":"Section V-A, references"},{"comment":"There is a typo: 'rewrards' should be 'rewards'. Also, the notation R_m^t is overloaded with the communication rate R^{m,c}_t; consider renaming one to avoid confusion.","section":"Eq. (7) and surrounding text"},{"comment":"The caption says 'Training convergence and scalability evaluation,' but panels (c) and (d) are not training-convergence curves. Adjust the caption to reflect that (a)-(b) are training curves and (c)-(d) are parameter sweeps.","section":"Fig. 3 caption"},{"comment":"The abstract and introduction emphasize 'zero-shot' reasoning, but the edge-LLM uses few-shot examples retrieved from a memory buffer. The term 'zero-shot' is therefore inaccurate; please qualify it (e.g., 'zero-shot strategic directives at the HAPS' vs. 'few-shot in-context tuning at the edge').","section":"Section IV-D and abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the venue, and the architectural idea is interesting. The main blocker is the mismatch between the headline claims and the reported metrics; direct collision-rate and throughput measurements are essential. The non-stationary reward issue is also significant. If the authors can supply the missing measurements and a stability ablation, the work could become publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The architecture is a sensible engineering idea, and the paper does a good job motivating the timescale separation: edge-LLM latency at 0.6 s, cloud LLM at 2.55 s, and a 20 Hz control loop. Putting the big LLM on the HAPS for slow global load balancing and a small LLM on each UAV for tactical reward weighting is a concrete, non-obvious way to inject semantic reasoning into real-time control. The prompt template and the few-shot retrieval scheme are also useful details. This clearly extends the authors' prior hybrid LLM-DDQN work to the multi-UAV case with a HAPS meta-controller.\n\nThe problem is the evaluation. The abstract claims the architecture 'significantly reduces collision rates and improves aggregate system throughput,' but Figure 3 shows total reward, survival step count, communication reward, and handover probability. None of these is a direct collision-rate or throughput curve. Survival step count is a proxy that can be inflated by conservative flying, and the communication reward is a scalarized weighted sum, not Mbps. There are no error bars, confidence intervals, or multi-seed statistics, so 'significantly' is not supported in any statistical sense. This is the load-bearing gap: without direct metrics, the paper's strongest claim is unverified.\n\nThe reader's concern about non-stationary rewards is also real. The edge-LLM retunes alpha_1, alpha_2, alpha_3 every second, which makes the DDQN's reward a moving target. Standard convergence guarantees do not apply, and there is no ablation against fixed weights or a rule-based scheduler. Without that ablation, you cannot tell whether the gains come from LLM reasoning or just from dynamic reward shaping. The HAPS meta-controller's directives (Offload/Recall/Idle) are also described only abstractly; it is not clear how they are enforced locally.\n\nThat said, the core comparison is not circular. The baselines are external, and the architecture itself is coherent. The missing pieces are measurement and attribution, not conceptual integrity. I think the paper deserves a serious referee because the architecture is novel and the latency analysis is thoughtful, but it needs major revision: direct collision/throughput results, error bars, and an ablation against a simpler weight scheduler. Without those, the claims should be toned down. I would not cite the results yet, but I would point colleagues to the design ideas.","headline":"Plausible hierarchical LLM-DDQN architecture for multi-UAV ITNTN control, but the headline metrics (collision rate, throughput) are never measured; the evaluation is the weak link.","tokens_in":9441,"tokens_out":2038,"would_cite":false,"duration_ms":100065,"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":"A two-tier LLM hierarchy cuts drone collisions and improves network throughput in 3D aerial highways.","keywords":["multi-UAV navigation","aerial highways","large language models","deep reinforcement learning","handover management","integrated terrestrial and non-terrestrial networks","hierarchical control","reward shaping"],"falsifier":"A controlled comparison where the edge LLM's weight adjustments are replaced either by fixed weights or by a simple rule-based scheduler that maps the same semantic state (distance zones, link stability) to weights. If the rule-based version matches or beats the LLM version in collision rate and throughput, the paper's claim that LLM zero-shot reasoning drives the improvement would be falsified; if instead the LLM version clearly wins, the claim gains support.","tokens_in":8546,"feed_emoji":"🚁","tokens_out":6258,"duration_ms":71654,"temperature":0.7,"pith_summary":"This paper is trying to show that the complementary strengths of large language models and deep reinforcement learning can be combined in a hierarchy for controlling swarms of high-speed drones in 3D aerial highways. The key idea is to let a cloud-based language model handle slow, global decisions like balancing load across base stations, while smaller onboard language models convert each drone's local situation into a short-term tactical goal by adjusting the weights of a multi-objective reward. A fast reinforcement-learning controller then executes those goals at 20 Hz. If the simulations are right, this arrangement produces fewer collisions, longer survival times, and better overall network throughput than reinforcement learning alone, and it offers a template for injecting semantic reasoning into real-time control loops.","feed_headline":"Reduce UAV collisions and raise throughput with cloud-edge LLM hierarchy","feed_subtitle":"Slow cloud reasoning plus fast onboard AI keeps drone swarms safe and networks unclogged.","key_machinery":"The load-bearing mechanism is the cognitive reflection loop: an edge LLM converts a discretized semantic state (distance zones, link-stability labels) into a new scalarization of the multi-objective reward every 1 s, thereby steering the fast Deep Double Q-Network without injecting action commands directly. This turns the LLM into a reward shaper, and the HAPS cloud LLM adds a slower loop that issues load-balancing meta-actions to prevent base-station saturation.","core_discovery":"The paper sets out to show that the latency of large language models need not disqualify them from real-time drone control. It proposes a hierarchy: a cloud LLM on a high-altitude platform reasons every five seconds about global load balancing, while each drone runs a smaller edge LLM once per second that reads a semantic description of its local state and outputs a new weighting of the reward components (transit progress, link quality, collision penalty). A Deep Double Q-Network then acts on that weighted reward at 20 Hz, picking rotor speeds and network handovers. In simulation, this stack converges to higher total reward, survives longer without collisions, and sustains lower handover pro","pith_inferences":["Inference: The reward-reshaping pattern could be reused in other multi-agent settings where a slow semantic layer must guide a fast control layer, e.g., connected ground vehicles or warehouse robots, without coupling the LLM to the control loop.","Inference: The paper does not ablate the edge-LLM's weight tuning against a fixed or rule-based scheduler; if a simple heuristic that maps semantic zones to weights performed equally, the observed gains would land on the reward adaptation, not on LLM reasoning. A direct comparison would settle this.","Inference: Because the edge LLM retrieves few-shot examples from local memory, the framework may exhibit self-improvement over time; a long-horizon deployment could make the weight adjustments more consistent, a property the current simulations do not measure."],"forward_implications":["The LLM-guided framework overtakes the DRL baselines around episode 1,500 and reaches a higher steady-state reward.","Collisions that terminate episodes early in the baseline are largely avoided, so survival step counts approach the maximum episode length.","Handover probability stays lowest and flattest as the number of drones grows from 5 to 30.","The communication reward is highest once multiple base stations create interference, because the HAPS meta-controller coordinates associations."],"fun_headline_variants":["LLM hierarchy cuts drone crashes and boosts network flow","Cloud LLM plans, edge LLM reacts, DRL flies drones","Split-second drone control with layered LLM reasoning","Slow cloud brain, quick onboard LLM: fewer collisions"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The framework assumes that the reward weights tuned by an onboard language model every second will not destabilize the deep reinforcement learning controller that must keep learning from them; no convergence or ablation evidence directly supports that.","fun_headline_variants_meta":{"raw":{"variants":["LLM hierarchy cuts drone crashes and boosts network flow","Cloud LLM plans, edge LLM reacts, DRL flies drones","Split-second drone control with layered LLM reasoning","Slow cloud brain, quick onboard LLM: fewer collisions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000154,"raw_usage":{"total_tokens":1031,"prompt_tokens":710,"completion_tokens":321,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":454,"completion_tokens_details":{"reasoning_tokens":254}},"tokens_in":454,"tokens_out":321,"duration_ms":4732,"temperature":1.0,"reasoning_tokens":254,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T14:51:33.960919+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled comparison where the edge LLM's weight adjustments are replaced either by fixed weights or by a simple rule-based scheduler that maps the same semantic state (distance zones, link stability) to weights. If the rule-based version matches or beats the LLM version in collision rate and throughput, the paper's claim that LLM zero-shot reasoning drives the improvement would be falsified; if instead the LLM version clearly wins, the claim gains support.","supporting_citations":[],"review_version":1}