{"id":"50f9a8d4-fdac-410a-90b3-bc158ab28d75","arxiv_id":"1908.04685","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Vehicles learn to compress their radio observations into three values, and a deep Q-network uses those values to allocate V2X spectrum at about 97% of the optimal sum rate in a simulated 4x4 network.","lead":"A hybrid deep reinforcement learning system for vehicular networks learns to compress channel information into a few numbers, which a base station uses to assign spectrum, achieving about 97% of the brute-force optimal rate in simulations. The same idea is adapted to a fully distributed version where each vehicle decides locally from compressed global information.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'learn to compress' claim is unsupported: Section V never compares against fixed/non-learned feedback with the same bit budget, so near-optimal ARP may reflect the small 4x4 scenario rather than learned compression.","rationale":"The reader identifies exponential action-space scalability as the weakest assumption. I agree that scalability is a real scope limitation, but the more load-bearing concern for the paper's central novelty (learning to compress) is the absence of a fixed-compression control with the same feedback budget. The paper's headline numbers show that 3 real values or 36 bits suffice, but they do not show that the learned compression is responsible, because no non-learned baseline is tested. A cheap and decisive control experiment would settle whether the learned representation adds value or whether the small scenario is simply easy. This does not move the verdict: the paper should remain CONDITIONAL on providing such a baseline and on clearly restricting the near-optimal claim to the tested network size. The reader's rationale mentions missing comparisons but not this specific control, hence partial agreement.","tokens_in":17473,"tokens_out":12907,"duration_ms":138225,"concrete_test":"Run a fixed-compression baseline in the same N=K=4 simulator: each V2V feeds back its top-3 preferred channel indices (or, to match the 36-bit budget, quantized local channel-gain vectors to 12 bits per value); the BS reconstructs an approximate state and picks the allocation with the same exhaustive-search rule used for 'Optimal'. Evaluate over the same 2000 testing episodes and compare ARP to the C-Decision curves. If the fixed-compression ARP is within about 3% of C-Decision, the learned-compression contribution is not established; if it is substantially lower, the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V-C reports ARP approximately 97-98% with Nk=3 real-valued feedback (Fig. 4) and Section V-D reports 94% with 36 bits (Fig. 5), leading to the conclusion that each V2V link needs only three learned real values for near-optimal centralized allocation. The paper's central claim, reflected in the title, is that the DNN learns what to compress. Yet the only comparators are the full-CSI exhaustive-search 'Optimal' scheme and 'Random Action' (55%); there is no control where each V2V link sends a fixed, non-learned summary of comparable size (e.g., its top-3 channel indices, or 36 bits of quantized local CSI) and the BS then applies the same exhaustive-search allocation. Without such a control, the evidence cannot distinguish 'the learned feedback encodes the right information' from 'the 4-link, 4-channel problem is easy enough that any three scalars per link suffice.' The selection of Nk=3 and 36 bits from the same testing curves used to report final performance (Figs. 4(b), 5) adds a further selection-bias risk. This is a missing-control issue, not an internal inconsistency, but it directly targets the 'learn to compress' contribution.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two deep-reinforcement-learning-based architectures for spectrum sharing in a vehicular (V2X) network. In the C-Decision scheme, each V2V link compresses its local observation with a DNN, feeds a few real-valued numbers (or their quantized bits) back to a base station, and a DQN at the base station performs centralized channel allocation. In the D-Decision scheme, each V2V link also uses a local DQN to make decisions, while the base station aggregates compressed feedback into a small global message (AGI). A binary quantization layer based on the straight-through estimator is added for signaling reduction. Simulations with N=4 V2I links and K=4 V2V links show that the C-Decision scheme achieves about 97% of the brute-force optimal average return with 3 real-valued feedback values and about 94% with 36 feedback bits, while the D-Decision scheme reaches about 96% with 16 real-valued AGI and 90% with 80 AGI bits. The paper also reports robustness to feedback interval, input noise, and feedback noise.","tokens_in":17748,"tokens_out":3169,"duration_ms":34094,"significance":"If the central claim is established, the paper would demonstrate an attractive design: each V2V link needs only a handful of learned real values (or tens of bits) to support near-optimal centralized scheduling in a V2X network, and a similarly compact aggregated message for distributed scheduling. The authors deserve credit for evaluating against a brute-force exhaustive-search optimum computed on the same realizations, which avoids circularity of the near-optimality claim, and for showing convergence curves from multiple seeds in the training phase. The quantization-layer approach is a reasonable and clearly described extension. However, the core 'learn to compress' claim is not yet convincingly supported because no fixed/non-learned feedback baseline is compared, the headline hyperparameters are selected from the same testing curves used for performance reporting, and the evaluation is confined to a 4-by-4 network whose joint action space (256) is enumerated in the DQN output layer. These issues are fixable but require additional experiments and careful qualification.","major_comments":[{"comment":"The central claim that each V2V link 'learns' what to compress is not tested against a non-learned fixed feedback baseline with the same feedback budget. For example, a control scheme could send the indices of the strongest 3 channels (or 36 bits of quantized local CSI) from each link and then let the BS apply the same exhaustive-search allocation. Without such a control, the near-optimal ARP at Nk=3 and 36 bits could reflect the small 4-link, 4-channel problem being easy, rather than the learned feedback encoding the right information. This missing control directly targets the main contribution advertised in the title and abstract.","section":"V-C and V-D (Figs. 4 and 5)"},{"comment":"The choices Nk=3 and 36 feedback bits appear to be made by inspecting the testing ARP curves and are then reported as the operating points in the abstract and conclusion. Selecting hyperparameters on the same data used to report final performance introduces optimistic bias. The authors should either use a separate validation set for tuning, use cross-validation, or clearly state that the reported numbers are post-selection and provide a correction or hold-out evaluation.","section":"V-C and V-D (Figs. 4(b) and 5)"},{"comment":"The centralized DQN enumerates all K^N channel allocations in its output layer (256 outputs for N=K=4). The paper claims 'near-optimal performance' without qualification, but the architecture is only feasible at this toy scale; for realistic numbers of V2V links and channels the action space grows exponentially. The claims of near-optimality and the title's general promise should be explicitly restricted to the evaluated small network, or the authors should provide evidence that the method scales, for instance via a factorized action representation or a larger-scale experiment.","section":"III-B and Table III"},{"comment":"All headline quantitative results (97%, 94%, 96%, 90%, etc.) are reported as single averages with no confidence intervals, standard deviations, or number of seeds for the testing phase. Given that training is stochastic, the reader cannot assess whether the differences between schemes or between feedback sizes are statistically significant. Reporting multiple seeds with error bars or confidence intervals is necessary to support the quantitative comparisons.","section":"V-C through V-G"}],"minor_comments":[{"comment":"The vertical axis label reads 'Noramlized Average Return' and should be corrected to 'Normalized Average Return'.","section":"V-F, Fig. 7"},{"comment":"The notation for the number of real-valued AGI and binary AGI is inconsistent between the text (N_r^g, N_b^g' around Fig. 10) and the figure captions. Please define these symbols where they are first introduced.","section":"V-G and Fig. 10"},{"comment":"In several sentences the text says 'the APR reaches 90%' and 'the APR nearly achieves its maximal value', but the metric is elsewhere called ARP (Average Return Percentage). Please use one consistent acronym.","section":"V-G"},{"comment":"In Table III, the BS DQN input dimension is written as 'K × Nk'; it would be clearer to write 'K·Nk' or 'K N_k' to indicate product rather than a 2D structure.","section":"III-A and Table III"},{"comment":"Fig. 3(b) is said to be computed under 10 different random seeds with epsilon=0, but the paper does not state whether the final reported ARP values in later figures are also averaged over seeds; please clarify the test-phase averaging procedure.","section":"V-B"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the topic is timely, but the central 'learn to compress' contribution is under-supported by the current evaluation design. The missing fixed-feedback control is the most important issue and should be addressed before publication. The small network size and test-set hyperparameter selection should also be dealt with, either by additional experiments or by substantially tempering the generality of the claims. I see no indication of misconduct; the concerns are purely about experimental design and reporting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe useful way to read this paper is as a modest, honest extension of the authors' own DRL resource allocation work: each V2V link compresses its local CSI with a DNN, optionally quantizes it, and a DQN at the BS makes the allocation. The simulation results in the 4x4 scenario are qualitatively solid: the training curves converge, and the 97/94% ARP against a brute-force optimum is credible. The comparison against the true optimum is the right baseline, so the near-optimality claim is not circular.\n\nWhat is genuinely new is the joint training of the per-link compression DNNs with the BS DQN (and, in the D-Decision variant, the BS aggregation DNN), including a straight-through binary layer. That is a real combination step beyond [26] and [27]. The robustness checks to feedback interval and noise are worth having. The paper is clearly written and the method is reproducible in principle.\n\nThe soft spots are in proportion to the central claim. The biggest one is the missing control: nothing compares learned feedback against a fixed, non-learned summary of the same size (say, three raw channel-gain indices, or 36 quantized bits of local CSI) with the same exhaustive-search allocation. Without that, the evidence does not distinguish 'the DNN learned what to compress well' from 'the 4x4 problem is easy enough that any few scalars per link suffice.' That gap directly targets the title. Second, Nk=3 and 36 bits are picked from the same testing curves used to report the final ARP numbers, so there is a selection-bias risk that is not addressed. Third, there are no confidence intervals on the headline percentages — Fig. 3(b) averages seeds, but the 97% and 96% claims appear to be single-run averages. Fourth, the C-Decision DQN action space is K^N (256 outputs at 4x4); the paper does not discuss how this scales to realistic V2V link counts. No code or data is released, which compounds all of these.\n\nThe citation pattern is fine; the paper builds on the authors' prior works transparently and does not overclaim relative to the literature in the text.\n\nWho is this for: people working on V2X spectrum sharing with DRL, and anyone interested in learned feedback compression. It is a useful method paper for a niche, not a landmark. It deserves a serious referee, but a revision needs the fixed-feedback control at minimum, plus confidence intervals and a scaling discussion. As is, the 'learn to compress' wording overstates what is shown.","headline":"Solid small-scenario DRL compression paper, but the 'learn to compress' claim needs a fixed-feedback control to be believed.","tokens_in":18291,"tokens_out":2159,"would_cite":false,"duration_ms":21890,"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":"Learned 3-value feedback reaches 97% of optimal V2X spectrum sharing","keywords":["vehicular networks","V2X communications","deep reinforcement learning","spectrum sharing","channel state information compression","binary feedback","resource allocation","deep Q-network"],"falsifier":"Compute, for the trained C-Decision scheme, the fraction of testing steps in which the action chosen coincides with the brute-force optimal action, alongside the 97% return ratio. If the exact-match rate is far below the return ratio, the near-optimal return is carried by many near-tied actions with similar rates, and the claim would not transfer to scenarios where suboptimal allocations are sharply worse.","tokens_in":17251,"feed_emoji":"📡","tokens_out":7140,"duration_ms":66445,"temperature":0.7,"pith_summary":"This paper tries to show that a vehicular network can offload nearly all spectrum-allocation intelligence to a base station while each vehicle sends only a few learned numbers instead of full channel state. The proposed C-Decision architecture trains each vehicle's neural network to compress its local observations into three real-valued feedback values (or 36 quantized bits), and trains a deep Q-network at the base station to map those compressed values to channel assignments for all vehicle-to-vehicle links. In simulations with four V2I channels and four V2V links, the scheme reaches about 97% of the brute-force optimal sum rate with real-valued feedback and about 94% with binary feedback. A distributed variant, D-Decision, lets each link decide locally using a small broadcast summary from the base station and reaches about 96% of optimal with 16 aggregated values. The paper also reports robustness to feedback interval, input noise, and feedback noise.","feed_headline":"Learned 3-value feedback reaches 97% of optimal V2X spectrum sharing","feed_subtitle":"The scheme hits 97% of optimal sum rate while random choice gets 55%, simulations show.","key_machinery":"The load-bearing object is the jointly trained compression-and-allocation chain: a per-vehicle compression DNN with a quantized feedback layer, feeding a base-station deep Q-network (C-Decision), or a base-station aggregation DNN feeding per-link decision DQNs (D-Decision). The mechanism is that the compression DNN and the Q-network are trained end-to-end by Q-learning updates, so the learned feedback is optimized for the downstream allocation task rather than for reconstructing CSI. The quantization layer uses tanh to map values to $[-1,1]$, the sign function to binarize, and the straight-through estimator to supply gradients through the nondifferentiable sign. The action space is the set of all channel assignments, encoded by the indicator vector $\\rho_k[n]$, and the reward is the weighted sum of V2I and V2V rates.","core_discovery":"The central claim is that the information bottleneck is not the bottleneck: a handful of learned feedback values per link is enough for near-optimal centralized resource allocation in V2X spectrum sharing, provided the feedback is trained jointly with the allocation policy rather than hand-designed. Each V2V link compresses its local observation $o_k = \\{h_k, I_k, P_k^d, h_{k,B}\\}$ through a DNN into $b_k$, and the base station's DQN treats the concatenated compressed vector as the state and outputs one of all $K^N$ joint channel allocations. The reward couples V2I and V2V throughput through a weighted sum rate; adjusting the weight shifts V2I quality of service. Binary feedback is obtained by a learned pre-binary tanh layer followed by a sign function trained with the straight-through estimator, so quantization is learned rather than fixed. The same training mechanism, with a base-station aggregation DNN producing a compact global summary $\\varphi$, yields the distributed D-Decision scheme with only about 2-4% performance loss relative to C-Decision.","pith_inferences":["Because the feedback vector is trained jointly with the Q-network, the same architecture should extend to other resource-allocation tasks, such as joint power control and channel selection, by changing the action space and reward; the paper names joint power control as future work, but the compression mechanism itself is task-agnostic.","The enumeration of all $K^N$ allocations in the DQN output layer limits the claim to very small networks; a testable extension is replacing the flat action layer with a factored or sequential action decomposition to keep the learned-feedback advantage at realistic scales.","The robustness to input and feedback noise suggests the learned feedback encodes allocation-relevant structure rather than raw channel values; one could test this by measuring how much of the optimal action's variance is explained by the three feedback values under noisy observations.","Comparing the 97% C-Decision figure with random action's 55% suggests most of the gain comes from the base station's global view rather than from per-link rate-maximizing choices, which is consistent with the paper's motivation but is not separately ablated in the paper."],"forward_implications":["With only 3 real-valued feedback numbers per V2V link, the C-Decision scheme attains about 97% of the brute-force optimal average return; adding more feedback values does not improve it.","Quantizing the 3 values to 36 bits keeps performance at about 94% of optimal, and binary feedback is more robust to feedback noise than real-valued feedback.","The D-Decision scheme reaches about 96% of optimal with 16 real-valued aggregated global information values, and about 90% with 80 aggregated bits, at the cost of fully distributed local decisions.","The learned allocation is robust: performance degrades only slowly with feedback interval up to large intervals, and the scheme retains about 60% of optimal even under very large input noise, still above random action.","The V2I/V2V weight $\\lambda_c$ gives a practical knob: raising it shifts the V2I sum-rate CDF rightward while V2V sum rate stays close to optimal."],"supporting_citations":[{"why":"Supplies the urban simulation scenario, vehicle mobility model, and system parameters used in all performance evaluations.","marker":"[5]"},{"why":"Supplies the premise that cross-channel gains $h_{k,B}$ can be estimated at the BS and broadcast to V2V links, and the multi-agent RL spectrum-sharing baseline the reward design draws on.","marker":"[27]"},{"why":"Supplies the delay-free feedback assumption that lets the V2V transmitter know its local channel information.","marker":"[29]"},{"why":"Provides the Q-learning update rule on which the DQN training is based.","marker":"[30]"},{"why":"Provides the DQN architecture, experience replay, target network, and the approximate target value used in the training loss.","marker":"[32]"},{"why":"Supplies the two-step binary quantization procedure (continuous values followed by binarization) used for the learned feedback.","marker":"[34]"},{"why":"Supplies the straight-through estimator that makes the sign function trainable by backpropagation.","marker":"[35]"},{"why":"Supplies the V2V path-loss model (LOS WINNER+B1 Manhattan) used in the channel simulations.","marker":"[36]"}],"fun_headline_variants":["Learned compression yields near-optimal V2X spectrum sharing","Tiny learned feedback hits 97% of optimal in V2X networks","V2X resource sharing with compressed learned CSI feedback","Compressed CSI feedback enables near-optimal V2X allocation","Learned 3-bit feedback drives 97% optimal V2X spectrum use"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The scheme assumes the base station's deep Q-network can enumerate every joint channel allocation of all V2V links in its output layer, which is only feasible for tiny networks; in the simulations this is 4 channels and 4 links, giving 256 actions.","fun_headline_variants_meta":{"raw":{"variants":["Learned compression yields near-optimal V2X spectrum sharing","Tiny learned feedback hits 97% of optimal in V2X networks","V2X resource sharing with compressed learned CSI feedback","Compressed CSI feedback enables near-optimal V2X allocation","Learned 3-bit feedback drives 97% optimal V2X spectrum use"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000177,"raw_usage":{"total_tokens":1310,"prompt_tokens":981,"completion_tokens":329,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":238}},"tokens_in":597,"tokens_out":329,"duration_ms":4634,"temperature":1.0,"reasoning_tokens":238,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:55:47.956998+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute, for the trained C-Decision scheme, the fraction of testing steps in which the action chosen coincides with the brute-force optimal action, alongside the 97% return ratio. If the exact-match rate is far below the return ratio, the near-optimal return is carried by many near-tied actions with similar rates, and the claim would not transfer to scenarios where suboptimal allocations are sharply worse.","supporting_citations":[{"cited_title":"Technical spefica tion group radio access network: Study on LTE-based V2X serv ices,","cited_arxiv_id":null,"evidence_quote":"Supplies the urban simulation scenario, vehicle mobility model, and system parameters used in all performance evaluations."},{"cited_title":"Spectrum sharing in vehicu lar networks based on multi-agent reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the premise that cross-channel gains $h_{k,B}$ can be estimated at the BS and broadcast to V2V links, and the multi-agent RL spectrum-sharing baseline the reward design draws on."},{"cited_title":"Q-learning,","cited_arxiv_id":null,"evidence_quote":"Provides the Q-learning update rule on which the DQN training is based."},{"cited_title":"Human-level control through deep reinforcement learnin g,","cited_arxiv_id":null,"evidence_quote":"Provides the DQN architecture, experience replay, target network, and the approximate target value used in the training loss."},{"cited_title":"IST-4-027756 WINNER I I d1. 1.2 v1. 2 WINNER II channel models","cited_arxiv_id":null,"evidence_quote":"Supplies the V2V path-loss model (LOS WINNER+B1 Manhattan) used in the channel simulations."}],"review_version":1}