{"id":"71c4182e-5246-4e9d-b2c8-ecbd16901b62","arxiv_id":"2505.05224","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A GFlowNet-based active learner discovers radio resource allocations for integrated sensing, communication, and computing networks, reporting 20% utility gains over PPO and MCMC baselines with fewer acquisition rounds.","lead":"The paper uses a generative flow network inside an active learning loop to allocate radio resources to devices that need communication, sensing, and computing at once. It reports finding high-utility allocation patterns with about 20% gains over two baselines while using half as many evaluation rounds.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section IV's feasibility guarantee is false as stated: unique resource blocks plus exactly U modified entries does not prevent the same device from occupying several blocks while another device gets none, which violates Section III's one-block-per-device rule and can distort the utilities behind…","rationale":"The reader's conditional verdict rests on a real gap: the only formal guarantee in the generator section is logically insufficient. I checked whether any other text closes the gap—Section IV's 'previously selected blocks cannot be re-used by other users' addresses entry reuse, not device uniqueness, and the Fig. 2 illustration does not specify a per-device mask. This concern is load-bearing because the entire empirical comparison (Figs. 3 and 4) measures f(X) under constraint (13), and infeasible terminal states can have undefined or distorted utilities. The attack is not a disagreement with the community's consensus about GFlowNets; it is an internal inconsistency between the system model's constraints and the generator's stated constraints. I do not see grounds to reject outright, because the implementation could include the missing device mask, but the ambiguity blocks verification and reproducibility. Hence I keep the reader's CONDITIONAL verdict; the concrete test above would settle whether the concern actually changes the reported numbers.","tokens_in":8043,"tokens_out":7415,"duration_ms":83709,"concrete_test":"Rebuild the Section IV generator with the exact stated action mask: an action is valid iff X_{m,w}=0 and the trajectory length is < U. Enumerate or sample terminal matrices for U=20, M=5, W=10 and count how many contain a device with two or more assignments and a device with zero assignments. Then re-run the Fig. 3b experiment with an additional mask that forbids selecting a device already present in X. If the utilities or selected allocations change materially, the reported 20% gain is not robust to the missing constraint; if no change occurs, the authors should still state the device-uniqueness mask explicitly, because the current feasibility proof in Section IV is incomplete.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III defines a feasible allocation by the rule that each device is allocated to one MEC server over one resource block, and each resource block is used by one device. Section IV's generator, however, only forbids re-selecting an already occupied block (m,w) and requires the number of modified entries to be U. These two conditions are necessary but not sufficient: a terminal matrix with U nonzero entries in distinct blocks can still place device d in multiple blocks and leave device d' with no block. Such a matrix violates the system model. Moreover, the utility fd(X) in (12) is not even well-defined for an omitted device: the computing latency in (11) has no assigned term, and duplicated blocks can inflate the communication and sensing terms through (5) and (9). The paper asserts 'we guarantee that terminal objects sampled by the GFlowNet are feasible resource allocation matrices X' without adding a device-uniqueness constraint to the action mask. If the implementation silently enforces that constraint, the text is incomplete; if it does not, the abstract's 20%-gain and half-round claims may be computed over infeasible allocations. Either way, the central experimental claim is not verifiable from the manuscript as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an active-learning framework for radio resource allocation in an integrated sensing, communication, and computing network. A Gaussian process surrogate is iteratively updated with utility evaluations, and a GFlowNet is trained against this surrogate to propose diverse high-return resource allocation matrices. The authors report that the method achieves about 20% utility gains over PPO and MCMC baselines in a scalability experiment, and reaches the baseline performance in about half the acquisition rounds in a convergence experiment. The central novelty is the use of a GFlowNet as a discrete compositional sampler inside an active-learning loop.","tokens_in":8262,"tokens_out":3604,"duration_ms":39495,"significance":"If the reported results hold, the paper makes a useful contribution by showing that GFlowNet-based generative active learning can be a viable heuristic for high-dimensional discrete resource allocation, where classical RL and MCMC suffer from mode collapse or slow mixing. The evaluation is not circular: the GFlowNet is trained on the GP surrogate while the reported utilities are computed from the true system model, so the numbers are not an artifact of the surrogate. The main strengths are the clear problem formulation, the use of trajectory balance with a stated convergence guarantee, and a sensible comparison setup. However, the current manuscript leaves the central empirical claims unverifiable because of an incorrect feasibility argument, an ill-defined sensing SINR expression, and missing experimental details such as seeds, error bars, and hyperparameter values.","major_comments":[{"comment":"The feasibility guarantee stated in Section IV does not follow from the constraints described. The text says that previously selected resource blocks cannot be re-used and that the number of modified entries equals U, then concludes that terminal objects are feasible allocation matrices. Section III defines feasibility as each device being allocated to one MEC server over one resource block. The two stated constraints only ensure that U distinct entries are non-zero; they do not prevent the same device from occupying several blocks while another device receives none. Such a matrix violates the system model, and the computing latency in (11) is undefined for an omitted device because the numerator has no allocated block. The fix is to add an explicit constraint in the action mask forbidding a device that has already been assigned, and to state that constraint in the text. If the implementation silently enforces this, the description is incomplete; if it does not, the reported utilities may be computed over infeasible allocations.","section":"Section IV, Generator Model"},{"comment":"The abstract combines two separate gains into one claim. Figure 3a shows that GFlowNet reaches a solution about 8% better than PPO in half the acquisition rounds (5 versus 10), while Figure 3b shows roughly 20% improvement after 10 rounds when the number of devices is varied. As written, the abstract's statement '20% performance gains against benchmarks, while requiring less than half of the number of acquisition rounds' implies the 20% gain is achieved at the half-round point, which is not supported by the displayed curves. Please report these two results separately, or show a joint trade-off curve so that the claim matches the evidence.","section":"Abstract and Section V, Figs. 3a and 3b"},{"comment":"The sensing SINR in Eq. (10) is not well defined. The numerator sums over the MEC index m, but the denominator contains the term m' ≠ m with m appearing as a free variable outside the numerator's summation convention. As typeset, the expression has no unambiguous meaning. Moreover, if the generator can produce a matrix in which device d occupies multiple blocks, the numerator would sum over multiple terms, which is inconsistent with the one-block-per-device rule of Section III. Please rewrite Eq. (10) with a fixed MEC index or a clear convention, and ensure it is consistent with the feasibility constraints.","section":"Section V, Eq. (10)"},{"comment":"No random seeds, number of independent runs, or error bars are reported for any figure. GFlowNet, PPO, and MCMC are all stochastic algorithms, and the text makes strong statements such as 'consistently outperforming benchmarks by around 20%' based on what appears to be a single run. Please add the number of seeds, report mean and standard deviation (or confidence intervals), and describe the variance across runs. This is necessary to verify the robustness of the central claim.","section":"Section V, Simulation Results"},{"comment":"Several parameters that are essential for reproducibility are missing: the utility weights ω_d^ρ, ω_d^ζ, and ω_d^ξ in Eq. (12) are not specified; the GP kernel hyperparameters ν and l, and the architecture and training details of the deep kernel network, are not reported; and the number of GFlowNet training steps per active-learning round and the backward policy parameterization are not given. Without these details, the reported gains could depend on unstated tuning choices. Please provide the full experimental configuration.","section":"Section V, Simulation Setup"}],"minor_comments":[{"comment":"In the definition of terminal states, the phrase 'states with no outgoing edges are terminal states are terminal states' contains a duplicated fragment. Please revise.","section":"Section II-A"},{"comment":"The feasibility rule 'each device can be allocated to communicate with one MEC server over one resource block; conversely, each subcarrier can be retained for one device only' is stated informally. It would be clearer to define the set X of feasible allocations explicitly, including whether a device must receive exactly one block or at most one block.","section":"Section III"},{"comment":"The claim that problem (13) is 'a variant of the knapsack problem which belongs to the class of NP-complete problems' is stated without a reduction or reference. A brief justification or citation would be appropriate.","section":"Section III-D"},{"comment":"The claim that the GFlowNet solution is 'much better' than the baselines is not quantified. Please add numerical values or a table, since the 3D scatter plot alone does not show the utility differences clearly.","section":"Section V, Fig. 4"}],"recommendation":"major_revision","confidential_remarks":"The paper reads more like a workshop or short conference contribution than a full journal paper in its current form. The active-learning framing and the GFlowNet sampler are reasonable, but the feasibility and reporting issues are substantial enough that I would not recommend acceptance without major revision. If the authors can fix Eq. (10), make the feasibility constraint explicit, and add full experimental details with multiple seeds, the contribution could be publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a legitimate new application—GFlowNet-based active learning for joint communication, sensing, and computing resource allocation—and the problem formulation is sensible. But the paper as written does not support its headline claims. The generator model in Section IV says that forbidding reuse of a block (m,w) and setting the number of modified entries to U \"guarantees\" feasible allocations. That guarantee is false. Nothing stops the same device from occupying two blocks while another device gets none, which violates Section III's rule that each device is allocated to one MEC server over one block. That is not a cosmetic detail: if the implementation does not mask already-assigned devices, the reported utilities can be computed over infeasible matrices, artificially inflating throughput and sensing for duplicated devices and zeroing the utility for omitted ones. If the implementation does enforce device uniqueness, the text is simply incomplete—but either way, the abstract's 20% gain and half-round claim are not verifiable from the manuscript.\n\nWhat the paper does well: the active learning loop (GP surrogate plus GFlowNet sampler) is a coherent way to handle an expensive, discrete, high-dimensional utility, and the comparison against PPO and MCMC is a reasonable sanity check. The multi-objective utility product in (12) is a fair way to capture communication, sensing, and computing trade-offs. The simulation setup is clearly described, and the figures suggest the method finds higher-utility solutions faster than the baselines, at least in the scenarios shown.\n\nSoft spots, in proportion: the feasibility gap is load-bearing and needs to be fixed. There are no error bars, no seeds, and no code, so the empirical results are hard to trust even after the feasibility issue is resolved. The abstract combines two separate results—Fig. 3a's 8% gain in convergence speed and Fig. 3b's ~20% gain in scalability—into one \"20% gains while requiring less than half the rounds\" headline, which is misleading. The sensing SINR expression in (10) also looks inconsistent: the numerator sums over MEC m while the denominator excludes m' = m, so the interference model is not clearly specified. Minor: the novelty is mostly the application; the GFlowNet and active learning machinery is off-the-shelf, and the authors should say that more plainly.\n\nWho this is for: researchers working on data-driven radio resource management, especially those interested in active learning or generative samplers for combinatorial allocation. A serious referee should engage with this, but the paper needs major revision before the empirical claims can be accepted. If the authors add the missing device-uniqueness constraint, report multiple seeds with standard deviations, and separate the convergence and scalability claims, the core idea could be a useful contribution.\n\nRecommendation: send to peer review, but with the expectation of substantial revision.","headline":"A plausible new application of GFlowNet active learning to ISCC resource allocation, but the headline gains are not verifiable as written because the generator's stated constraints do not enforce the system model's one-block-per-device rule.","tokens_in":8842,"tokens_out":3226,"would_cite":false,"duration_ms":34571,"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":"An active learning loop pairing a GFlowNet sampler with a Gaussian-process surrogate allocates radio resources with 20% higher utility in less than half the acquisition rounds of PPO or MCMC.","keywords":["Radio resource management","Integrated sensing communication and computing","Active learning","Generative flow network","Gaussian process surrogate","Trajectory balance","Resource allocation","Wireless networks"],"falsifier":"Sample a few thousand terminal matrices from the trained GFlowNet and check the system model's per-device constraint: if the same device appears in multiple resource blocks while some other device appears in none, the claimed feasibility guarantee fails and the reported utilities are computed over invalid allocations. A simpler version: re-run the active loop with an explicit one-block-per-device constraint; if the 20% gain disappears, the original advantage depended on those invalid configurations.","tokens_in":7778,"feed_emoji":"📡","tokens_out":10653,"duration_ms":98427,"temperature":0.7,"pith_summary":"This paper targets radio resource allocation in wireless networks that must simultaneously serve communication, sensing, and computing traffic. The authors formulate the problem as maximizing a sum of per-device utilities and observe that the resulting search over allocation matrices is NP-hard and high-dimensional. They propose a generative active learning loop: a Gaussian process models the expensive true utility from past allocations, and a GFlowNet—trained so that sampled matrices are proportional to the surrogate reward—proposes the next set of allocations to evaluate. The paper reports that this loop reaches higher-utility allocations in about half the acquisition rounds of PPO or MCMC and keeps a roughly 20% advantage as the number of devices grows. A sympathetic reader would care because it suggests a practical, general recipe for quickly configuring multi-functional wireless networks at the start of each scheduling slot.","feed_headline":"GFlowNets find wireless allocations 20% better in half the rounds","feed_subtitle":"Pairing a GFlowNet sampler with a cheap surrogate beats PPO and MCMC at radio resource allocation","key_machinery":"The load-bearing mechanism is the combination of a Gaussian-process surrogate with a GFlowNet generator inside a sequential active learning loop. A GFlowNet is a generative model over compositional objects: it builds an M×W allocation matrix one entry at a time from an all-zero start, with each action setting X_{m,w}=d, and it is trained with the trajectory balance loss so that terminal objects are sampled with probability proportional to the reward—here the GP's predicted utility. The GP uses a Matérn kernel applied to neural-network embeddings of the matrices, keeping the surrogate tractable. The paper argues that proportional-to-reward sampling yields better mode coverage than PPO's expected-return maximization and faster mixing than MCMC, which is what allows fewer acquisition rounds.","core_discovery":"The paper's central claim is that a generative flow network trained with the trajectory balance objective can serve as the sampler in an active learning loop for combinatorial radio resource allocation, outperforming both a deep RL policy (PPO) and a classical MCMC sampler. The utility optimized is the sum over devices of weighted products of communication bitrate, sensing mutual information, and inverse computing latency, which couples MEC-server selection and subcarrier assignment. In simulations with 20 devices, the GFlowNet agent reaches its best solution in roughly five acquisition rounds—about half the ten rounds PPO needs—with a solution about 8% higher, while MCMC takes about twenty rounds for a solution 4% above PPO. Scaled over device counts up to 40, the GFlowNet consistently outperforms both baselines by about 20% after ten rounds. The authors directly attribute the gain to GFlowNet's training objective, which makes the probability of sampling an allocation proportional to its surrogate reward and therefore keeps exploring diverse high-return modes.","pith_inferences":["A clean control that keeps the GP surrogate but replaces the GFlowNet with random sampling would isolate whether the reported gains come from the sampler's proportional-to-reward coverage or simply from the GP learning; the paper does not report such an ablation.","Since the generator's stated constraints do not explicitly stop one device from taking several resource blocks while a neighbor gets none, the natural robustness check is to add a one-block-per-device constraint and re-measure the 20% gain; the paper leaves that check unstated.","The conclusion names multi-fidelity active learning as future work; a concrete version would let the sampler choose the evaluation fidelity, spending cheap surrogate evaluations early and expensive oracle evaluations only around promising modes."],"forward_implications":["If the reported gains are correct, a GFlowNet-based scheduler can configure a multi-functional wireless network at the start of each slot with less than half the system evaluations required by PPO or MCMC.","Because the trajectory balance objective makes sampling proportional to the surrogate reward, the active loop keeps proposing diverse high-utility allocations rather than collapsing onto one optimum, which is the failure mode the paper attributes to PPO.","Applied to other combinatorial radio resource problems with expensive objectives—beamforming pattern selection, user association, subcarrier assignment—the same GFlowNet-plus-GP loop would need only a redefinition of the DAG's actions.","The reduction in acquisition rounds translates directly into shorter scheduling delay in the time-slotted scenario the paper assumes, since the allocation is fixed for the duration of the slot."],"supporting_citations":[{"why":"Frames radio resource management as a Bayesian/active learning problem that the paper adapts with a GFlowNet sampler.","marker":"[13]"},{"why":"Introduces GFlowNets, the generative model family used to sample allocation matrices.","marker":"[14]"},{"why":"Provides the theoretical grounding for GFlowNet sampling proportionally to reward, which motivates the method's diversity.","marker":"[17]"},{"why":"Gives the trajectory balance loss used for training and the proposition ensuring proportional sampling at zero loss.","marker":"[18]"},{"why":"Defines the Gaussian process and Matérn kernel that form the surrogate model of the true utility.","marker":"[20]"},{"why":"Supplies deep kernel learning, used to embed high-dimensional allocation matrices before applying the GP kernel.","marker":"[22]"},{"why":"Defines PPO, one of the two baselines against which the GFlowNet method's gains are measured.","marker":"[23]"}],"fun_headline_variants":["GFlowNet active learning cuts wireless rounds in half, lifts gains 20%","GFlowNet sampler finds 20% better wireless allocations in half the rounds","GFlowNet beats PPO and MCMC in wireless resource allocation","GFlowNet active learning: 20% gain, half the acquisition rounds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim assumes the GFlowNet's stated constraints—never reusing a resource block and editing exactly as many entries as there are devices—guarantee feasible allocations, but those rules still allow one device to take several blocks while another device gets none, and the paper does not say whether device-uniqueness is enforced.","fun_headline_variants_meta":{"raw":{"variants":["GFlowNet active learning cuts wireless rounds in half, lifts gains 20%","GFlowNet sampler finds 20% better wireless allocations in half the rounds","GFlowNet beats PPO and MCMC in wireless resource allocation","GFlowNet active learning: 20% gain, half the acquisition rounds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000491,"raw_usage":{"total_tokens":2402,"prompt_tokens":918,"completion_tokens":1484,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":1402}},"tokens_in":534,"tokens_out":1484,"duration_ms":9723,"temperature":1.0,"reasoning_tokens":1402,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:09:24.394592+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Sample a few thousand terminal matrices from the trained GFlowNet and check the system model's per-device constraint: if the same device appears in multiple resource blocks while some other device appears in none, the claimed feasibility guarantee fails and the reported utilities are computed over invalid allocations. A simpler version: re-run the active loop with an explicit one-block-per-device constraint; if the 20% gain disappears, the original advantage depended on those invalid configurations.","supporting_citations":[{"cited_title":"Bayesian optimization for radio resource management: Open loop power control,","cited_arxiv_id":null,"evidence_quote":"Frames radio resource management as a Bayesian/active learning problem that the paper adapts with a GFlowNet sampler."},{"cited_title":"Flow network based generative models for non-iterative diverse candidate generation,","cited_arxiv_id":null,"evidence_quote":"Introduces GFlowNets, the generative model family used to sample allocation matrices."},{"cited_title":"Gflownet foundations,","cited_arxiv_id":null,"evidence_quote":"Provides the theoretical grounding for GFlowNet sampling proportionally to reward, which motivates the method's diversity."},{"cited_title":"Trajectory balance: Improved credit assign- ment in gflownets,","cited_arxiv_id":null,"evidence_quote":"Gives the trajectory balance loss used for training and the proposition ensuring proportional sampling at zero loss."},{"cited_title":"Stochastic variational deep kernel learning,","cited_arxiv_id":null,"evidence_quote":"Supplies deep kernel learning, used to embed high-dimensional allocation matrices before applying the GP kernel."}],"review_version":1}