{"id":"27d1db57-21ac-4479-9e1d-767e061252ce","arxiv_id":"2502.10027","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A DNN with a learned binary router jointly solves fourteen wireless power allocation tasks of varying dimensionality and objectives, nearing single-task performance.","lead":"The paper designs a single neural network with a learned router that can solve many different wireless resource allocation problems of different sizes and goals in one model. The router turns parts of the network on and off per task, which keeps one task's training from harming another.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The experiments do not isolate the learned routing: the two-phase retraining and sparse hard-parameter-sharing structure, rather than the rDNN's decisions, could explain the gains. A random-mask ablation is missing.","rationale":"The reader's weakest_assumption focused on the fixed subnetwork per task and the 0.5 thresholding. Those are plausible but not the most decisive issue. The paper's central contribution is the learned routing mechanism, yet the experiments only compare against baselines that differ in both mask presence and training schedule. The 'naive multi-task DNN' has no masks and no retraining; zero-padding also lacks masks and retraining. Thus, the reported improvement could come from (i) the two-phase retraining, which lets the bDNN specialize after masks are frozen, or (ii) the mere sparsity of task-specific masks, rather than the rDNN's learned decisions. A random-mask control would directly test the causal role of the routing. This is a concrete missing ablation, not a disagreement with the empirical trend; the paper may still be correct, but the current evidence does not isolate the mechanism it claims. The reader's parameter-count parity concern is secondary because the final inference model uses masked bDNN weights and is not larger than the single-task DNN; the rDNN is a training-time overhead. My verdict remains CONDITIONAL, consistent with the reader, but the condition should include the random-mask ablation to substantiate the central claim.","tokens_in":19292,"tokens_out":11465,"duration_ms":111951,"concrete_test":"Implement a random-mask control: for each task i, sample binary matrices H^{l,i}_rand with the same proportion of zeros (or same number of active weights) as the trained rDNN output after thresholding, freeze them, and run only the Phase-2 retraining of Algorithm 1 (lines 12-18) with the identical optimizer, epochs, and loss. Evaluate on the same 14 tasks and report the same metrics as Tables II-III. If the random-mask baseline is within one standard deviation (or statistically indistinguishable) of the proposed method over multiple seeds, the central claim that the learned routing is essential is not supported; if the proposed method clearly outperforms the random-mask control across tasks, the concern is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that the trained rDNN, by learning binary masks (Eq. 15) that are multiplied with bDNN weights (Eq. 14), is responsible for the near-single-task performance in Tables II-III. However, the comparison does not control for two confounds. First, Algorithm 1 includes a two-phase procedure: joint training with soft masks, hard thresholding, and then retraining of the bDNN with fixed masks (lines 9-18). The 'naive multi-task DNN' baseline is trained once without masks or retraining, so the retraining phase itself may account for part of the gap. Second, any fixed task-specific sparse mask, even a random one, would reduce inter-task interference relative to the all-ones naive mask; the paper never tests a random-mask or fixed-mask baseline with the same sparsity as the rDNN output. Without such an ablation, the numerical results support the MTL architecture as a whole but do not establish that the learned routing decisions, rather than the sparse subnetwork structure and retraining schedule, drive the improvement. This is load-bearing because the paper's novelty and contribution statements (Section I-B) rest specifically on the rDNN learning which weights to set to zero.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a multi-task learning (MTL) architecture for solving wireless resource allocation problems of varying dimensionality, objective, and constraints with a single DNN. The architecture consists of a base DNN (bDNN) shared across tasks and a routing DNN (rDNN) that outputs binary masks, which are multiplied element-wise with the bDNN weights to define a task-specific subnetwork (Eqs. (13)-(15)). Tasks are defined as different network sizes N in {5,...,20} for an FDMA delay-minimization problem (supervised) and an average sum capacity maximization problem with average power constraints (unsupervised), giving fourteen tasks. Algorithm 1 trains the two networks jointly with soft masks, thresholds the masks at 0.5, and then retrains the bDNN with fixed masks. Numerical results compare the proposed scheme with single-task DNNs, zero-padding, and a naive multi-task baseline without routing, and report that the proposed method performs close to single-task DNNs while the baselines degrade.","tokens_in":19582,"tokens_out":5453,"duration_ms":57126,"significance":"If the empirical findings are reliable, the framework is a useful step toward DNN-based optimization that can adapt to variable problem dimensions and conflicting objectives without retraining from scratch, and the complexity analysis suggests training cost grows only by a constant factor relative to independent single-task DNNs. The paper is honest about the main design choice: routing depends only on the one-hot task indicator, so each task is assigned one fixed subnetwork for all channel realizations. The strengths include a clear formulation of the multi-task mapping, a broad fourteen-task testbed spanning supervised and unsupervised settings, and an explicit two-phase training procedure. However, the contribution hinges on whether the measured gains are caused by the learned routing decisions rather than by the sparse subnetwork structure and the retraining schedule; the current experiments do not isolate that mechanism, and the tables report point estimates without variance or significance tests.","major_comments":[{"comment":"To support the claim that the rDNN's learned routing is responsible for the observed gains, the experimental comparison must include fixed random masks with matched sparsity and a naive baseline that undergoes the same two-phase retraining loop. As written, the naive multi-task DNN is trained once without the threshold-and-retrain procedure, so the retraining phase and the per-task sparse structure are confounded with the routing objective. Without such an ablation, the results support the overall MTL architecture but do not establish that the specific routing decisions learned by the rDNN drive the improvement.","section":"Section V, Tables II-III and Algorithm 1"},{"comment":"The tables report only point estimates, with no error bars, multiple training seeds, or statistical significance tests. The 'near-single-task' claim rests on small absolute differences, e.g., 0.0152 vs. 0.0171 for the SL task at N=20 and 3.883 vs. 3.863 for the UL task at N=20, which may be within run-to-run variation. Please report the mean and standard deviation over several independent training runs and a paired comparison across tasks, so the reader can assess whether the proposed method is statistically indistinguishable from the single-task DNN.","section":"Section V, Tables II-III"},{"comment":"The claimed parameter-fair comparison between the single-task DNN and the bDNN is not fully specified. For tasks with N_i < max{N_1,...,N_K}, the single-task DNN's input and output layers have dimension N_i, while the bDNN uses d1 = dL = max{N_1,...,N_K} (Eq. (12)) unless the hidden widths are adjusted separately. Thus the parameter counts will generally differ. Please state the exact architectures and parameter counts for every scheme, including zero-padding and the naive multi-task baseline, so the fairness of the comparison can be verified.","section":"Section V, parameter-fair comparison"},{"comment":"The hard thresholding step Sign(ReLU(yr,i - 0.5)) can in principle destroy the information in the jointly trained soft masks, and the paper provides no quantitative sensitivity analysis. The statement that retraining the bDNN 'slightly improves' performance is not accompanied by a comparison of final task performance before and after thresholding, nor by a sweep over the threshold value. Please add an ablation that reports the distribution of learned soft-mask values and the effect of the threshold on the final objective values, since the thresholding step is an essential part of the central mechanism.","section":"Section III-B, Eq. (15)"}],"minor_comments":[{"comment":"The text refers to 'Table 3' and 'Table 4' when discussing the results that appear in Tables II and III; please correct the cross-references.","section":"Section V, Fig. 5 and surrounding text"},{"comment":"The caption contains the typo 'Constrain Violation' instead of 'Constraint Violation'.","section":"Fig. 6 caption"},{"comment":"The Sign function used in yr,i <- Sign(ReLU(yr,i - 0.5)) is not defined at zero; for completeness, please state that Sign(0) = 0.","section":"Algorithm 1, line 10"},{"comment":"The abstract states that the rDNN's binary vector is multiplied with all bDNN weights, while Eq. (14) explicitly excludes the first and last layers; please adjust the wording for consistency.","section":"Abstract and Section III-B"}],"recommendation":"major_revision","confidential_remarks":"The manuscript builds directly on the authors' prior PIMRC paper [1], which is cited and used as the naive baseline. Given that the rDNN routing mechanism is the primary novelty, the requested ablations in the major comments are the decisive factor: without a random-mask or fixed-mask control and without a statistical assessment, the central contribution is not yet established. The editor may also wish to check that the description of the 'naive multi-task DNN [1]' baseline in Section V is consistent with the methods reported in [1]."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: a useful application paper, not a methods breakthrough. The genuinely new bit is applying learned binary routing - conditional computation - to multi-task wireless resource allocation where tasks vary in dimensionality, objective, and constraints. The previous conference paper only handled dynamic dimensionality with a naive shared network; here the rDNN decides which bDNN weights to zero per task, and the fourteen-task evaluation across supervised and unsupervised problems is a real step up. The architecture is clearly described, Algorithm 1 is complete, and the qualitative ordering (single-task, proposed, naive, zero-padding) is consistent across all seven sizes for both objectives. Credit is earned there. The soft spots are real but not fatal. Most important: the experiments never isolate the learned routing. The proposed scheme differs from the naive baseline in two ways - it has task-specific sparse masks, and it undergoes a two-phase retrain after hard thresholding. A random-mask baseline with the same sparsity, plus a no-threshold control, would tell you whether the rDNN's specific decisions matter or whether any sparse subnetwork plus retraining does the work. Without that ablation, the main contribution claim about the rDNN learning which weights to cut is not fully supported, even though the multi-task architecture as a whole clearly beats the baselines. No error bars or significance tests appear, so the 'near single-task' language is descriptive; the claim that the single-task DNN is 'provably optimal' overstates the cited result; and the parameter-count parity is ambiguous because the rDNN's own parameters are apparently excluded. The fixed masks after thresholding are a design choice, but they also mean the router only issues one path per task, ignoring within-task variation. Who this is for: researchers in wireless MTL and learning-to-optimize who want a working recipe for heterogeneous tasks. It deserves a serious referee; the missing ablation and statistics should be requested. The core empirical comparison is probably reproducible, and the paper is honest about being an extension of the authors' own work.","headline":"A useful application paper that extends prior work with learned binary routing for multi-task wireless optimization, but the routing component is not isolated from sparse-mask and retraining effects.","tokens_in":646,"tokens_out":1500,"would_cite":false,"duration_ms":51354,"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 single deep network, guided by a small router that learns a binary mask per task, can jointly solve wireless resource-allocation problems that differ in size, objective, and constraints, nearly matching separately trained single-task…","keywords":["multi-task learning","conditional computation","routing DNN","wireless resource allocation","deep learning for optimization","supervised and unsupervised learning","variable dimensionality","power allocation"],"falsifier":"Re-run the fourteen-task experiment with the router allowed to see the input realization as well as the task index, and compare against the frozen per-task masks; if per-input routing yields a material gain in objective value or feasibility, the paper's fixed-path-per-task premise is falsified.","tokens_in":19120,"feed_emoji":"📡","tokens_out":7450,"duration_ms":71398,"temperature":0.7,"pith_summary":"Wireless networks face many resource-allocation problems that differ in the number of users or channels, in what is being optimized, and in the constraints. The paper proposes a single deep network that can handle all of these variants without retraining a new network for each one. The mechanism is conditional computation with routing: a small router network picks a subnetwork of a shared base network for each task by multiplying the base weights with binary masks. In experiments covering fourteen tasks, from convex delay minimization with labels to stochastic sum-capacity maximization without labels, the routed multi-task network nearly matches separately trained single-task networks, while naive parameter sharing and zero-padding fall clearly behind.","feed_headline":"A router layer lets one DNN solve 14 conflicting wireless tasks","feed_subtitle":"Learned per-task masks stop interference between shared weights, giving near-single-task accuracy across 14 allocation problems.","key_machinery":"The carrying object is the pair (bDNN, rDNN) with hard parameter sharing. The rDNN maps a one-hot task index $z_i$ to binary matrices $H^{l,i}$ for each hidden layer $l$; these are multiplied element-wise with the base weights, giving task-specific parameters $\\Theta_i = \\{ W^1[1:N_i,:], b^1, \\{ H^{l,i} \\odot W^l, b^l \\}_{l=2}^L, W^{L+1}[1:N_i,:], b^L \\}$ (Eq. 14), so each task flows through a distinct subnetwork of the same total parameter set. The hard threshold $y^{r,i} = \\mathrm{Sign}(\\mathrm{ReLU}(y^{r,i}(\\Phi,z_i)-0.5))$ (Eq. 15) converts soft outputs into binary masks. Training first updates $\\Omega=\\{\\Theta,\\Phi\\}$ jointly, including primal-dual steps for unsupervised constraints, then freezes the router and retrains the base network.","core_discovery":"The paper claims that hard parameter sharing with learned routing is enough to make multi-task learning work for heterogeneous optimization problems. During joint training, the router outputs soft masks; after training, these are thresholded at 0.5 and frozen, and each task runs through its own masked subnetwork of the shared base network, after which the base network is retrained. On the fourteen tested tasks, the proposed scheme stays close to the optimal single-task DNN, while naive multi-task and zero-padding baselines are substantially worse. The same framework supports both supervised tasks, where optimal labels are available, and unsupervised tasks, where a primal-dual Lagrangian loss enforces constraints.","pith_inferences":["A natural extension the paper leaves implicit is per-instance routing: giving the router the channel realization as well as the task index would test whether the optimal computation path ever varies within a single task, which the current frozen-mask design assumes it does not.","The framework's behavior on unseen tasks is untested; a concrete next experiment would be to take a trained bDNN and either evaluate a new task index directly or fine-tune the router on a new objective, to see whether routing transfers or must be relearned.","Because the final binary masks define sparse task-specific subnetworks, the method connects naturally to pruning and architecture search: the masks could be analyzed as a learned allocation of parameters across tasks, possibly yielding interpretable per-task model sizes.","In an Open RAN setting, the authors' consolidation motivation could be tested end-to-end by deploying one multi-task model in place of several single-task xApps and measuring lifecycle overhead against the small performance gap reported here."],"forward_implications":["A single trained allocator can serve network configurations of different sizes, objectives, and constraints without changing architecture or retraining from scratch.","Because routing uses the same parameter set for all tasks, the model size stays comparable to one single-task DNN rather than growing with the number of tasks.","Switching between tasks at inference time reduces to changing the one-hot task index, which is cheap and compatible with real-time operation.","The framework applies to both supervised problems with known optimal solutions and unsupervised non-convex or stochastic problems where optimal labels cannot be generated.","Training complexity is roughly a constant factor larger than training K separate DNNs, since the router overhead scales with the number of tasks but is independent of the base network's per-task cost."],"supporting_citations":[{"why":"Supplies the single-task DNN baseline and the primal-dual unsupervised learning procedure that the multi-task UL tasks build on.","marker":"[14]"},{"why":"Provides the multi-task learning survey and the modular-sharing and conditional-computation framing that motivates the bDNN-rDNN split.","marker":"[7]"},{"why":"Introduces routing networks, the idea the rDNN adapts from soft to hard parameter sharing.","marker":"[8]"},{"why":"Provides slimmable networks, the basis for selecting the first $N_i$ input and output neurons per task to handle variable dimensionality.","marker":"[12]"},{"why":"Defines the naive multi-task benchmark, an untrained router with output one everywhere, that the experiments must beat.","marker":"[1]"},{"why":"A zero-padding approach for variable-dimensionality DNNs used as the ZP benchmark in the numerical results.","marker":"[22]"},{"why":"Supports the training choices of sampling tasks uniformly and the observation that inter-task interference can speed convergence.","marker":"[10]"},{"why":"Supports the claim that unsupervised deep learning outperforms numerical methods in non-convex settings, motivating the UL formulation.","marker":"[13]"}],"fun_headline_variants":["One DNN, 14 jobs: router masks carve per-task paths","Routing masks let a shared DNN handle 14 tasks at once","Per-task masks in one net solve 14 wireless optimization problems","A router network directs a single DNN across 14 tasks","Learned routing creates unique subnets inside one DNN for wireless"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework assumes a single frozen subnetwork, found during joint training and fixed by a 0.5 threshold, remains the right computation path for every instance of a task, even though the router never sees the task's actual data.","fun_headline_variants_meta":{"raw":{"variants":["One DNN, 14 jobs: router masks carve per-task paths","Routing masks let a shared DNN handle 14 tasks at once","Per-task masks in one net solve 14 wireless optimization problems","A router network directs a single DNN across 14 tasks","Learned routing creates unique subnets inside one DNN for wireless"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001291,"raw_usage":{"total_tokens":5298,"prompt_tokens":996,"completion_tokens":4302,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":4212}},"tokens_in":612,"tokens_out":4302,"duration_ms":25964,"temperature":1.0,"reasoning_tokens":4212,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T19:39:23.651055+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the fourteen-task experiment with the router allowed to see the input realization as well as the task index, and compare against the frozen per-task masks; if per-input routing yields a material gain in objective value or feasibility, the paper's fixed-path-per-task premise is falsified.","supporting_citations":[{"cited_title":"Deep Learning for Distributed Optimization: Applications to Wireless Resource Management,","cited_arxiv_id":null,"evidence_quote":"Supplies the single-task DNN baseline and the primal-dual unsupervised learning procedure that the multi-task UL tasks build on."},{"cited_title":"Multi-Task Learning for Resource Allocation in Wireless Networks of Dynamic Dimensionality,","cited_arxiv_id":null,"evidence_quote":"Defines the naive multi-task benchmark, an untrained router with output one everywhere, that the experiments must beat."},{"cited_title":"A Deep Learning based Resource Allocation Algorithm for Variable Dimensions in D2D-Enabled Cellular Networks,","cited_arxiv_id":null,"evidence_quote":"A zero-padding approach for variable-dimensionality DNNs used as the ZP benchmark in the numerical results."},{"cited_title":"Learning to optimize: Training deep neural networks for wireless resource management,","cited_arxiv_id":null,"evidence_quote":"Supports the claim that unsupervised deep learning outperforms numerical methods in non-convex settings, motivating the UL formulation."}],"review_version":1}