{"id":"ba939855-6a23-4cb4-82e6-b666c82cb153","arxiv_id":"2506.14992","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A GFlowNet learns to sample time-modulated IRS configurations with probability proportional to sum rate at intended user directions, enabling multi-user directional modulation after training on a tiny fraction of the configuration space.","lead":"This paper trains a generative flow network to choose the switching times and phases of a time-modulated intelligent reflecting surface, so that a single OFDM transmitter can send a clean signal toward multiple directions while scrambling it everywhere else. The idea is that a stochastic policy can produce many different good settings, which the system can rotate over time to frustrate eavesdroppers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Multi-user sum-rate in Eq. (10) is not achievable: the model has a single OFDM data stream d(k) broadcast to all directions, so C_total is not a valid multiuser rate.","rationale":"The reader's weakest assumption is the same as mine. The single-stream model in Eq. (1) and the absence of any multi-user transmit processing make the sum-rate objective in Eq. (10) infeasible as a measure of independent-user throughput. The GFlowNet machinery itself is not at fault; the single-user results and the released code give some evidence for the method. However, the multi-user claim as stated in the abstract and Section IV depends on the invalid sum-rate reward. A conditional verdict requiring the authors to reframe the contribution (or add real multi-stream support) is the right level of scrutiny.","tokens_in":9846,"tokens_out":9201,"duration_ms":86635,"concrete_test":"Retrain the GFlowNet with reward R = min_u C_u, the common-message broadcast rate, instead of the sum C_total in Eq. (10), under the same two-user setting, and regenerate Fig. 4. If both user directions still achieve low SER, the current claim is valid only as a broadcast/multicast scenario and the 'sum rate' terminology is incorrect. If one user's SER degrades, the published two-user result relied on an infeasible reward. Also report the ratio of ∑_{u=1}^U C_u to min_u C_u for the selected configurations.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The described transmitter feeds every ULA element with the same OFDM signal e(t) of Eq. (1), which contains a single data stream d(k). The IRS reflects that same signal toward all directions, so Eq. (7) gives the received symbol y_i(θ,φ) = (√N/K) ∑_k d(k) V(i−k, ...) + z_i. No precoding, scheduling, or per-user subcarrier allocation is described. Therefore every legitimate user receives the same message d(k) after equalization. Eq. (10) sums U per-user rates C_u as if each user had an independent data stream, but the achievable rate for this common-message broadcast channel is at most min_u C_u, not the sum. Optimizing ∑_u C_u may still shape the spatial response, but the central claim of supporting multiple legitimate users with a sum-rate objective is unsupported. The two-user SER plot in Fig. 4 merely shows that the same d(k) is decodable at two angles, which is a multi-direction broadcast rather than independent multi-user transmission. This is a load-bearing mismatch between the advertised contribution and the system model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a generative flow network (GFlowNet) based method for designing time-modulated intelligent reflecting surface (TM-IRS) parameters in an OFDM directional modulation system. The parameter selection is cast as a deterministic MDP, and a trajectory-balance GFlowNet is trained to sample TM-IRS configurations with probability proportional to a reward that combines the sum rate across authorized directions with a phase-constraint penalty. Simulations compare the angular SER profile with a rule-based design for a single user, show a two-user example, and illustrate a time-varying configuration strategy that randomizes low-SER leakage directions. The authors report convergence after 9e5 sampled trajectories out of a space of roughly 10^65 configurations and provide implementation code.","tokens_in":10115,"tokens_out":7150,"duration_ms":58710,"significance":"The paper introduces a novel application of GFlowNets to a physical-layer security problem, and the idea of using generative models to explore a huge discrete TM-IRS parameter space is interesting and potentially valuable. The single-user comparison to the rule-based method is a reasonable sanity check, and the diversity-based security enhancement via switching among sampled configurations is a thoughtful contribution. The availability of code and the self-contained simulation are strengths. However, the advertised multi-user support is not supported by the system model, as detailed in the major comments; the claimed sum-rate objective and the two-user SER demonstration only establish a common-message broadcast capability. If reframed or extended to independent user streams, the approach could be significant for secure multi-directional transmission.","major_comments":[{"comment":"The system model in Eq. (1) transmits a single OFDM data stream d(k), and Eq. (7) shows that the received symbol at every direction is a function of that same d(k). Consequently, all legitimate users receive the same common message, and the per-user rates C_u in Eq. (10) cannot be summed as independent user rates. For this common-message broadcast channel, the achievable rate is at most min_u C_u, not the sum. The optimization problem in Eq. (11) and the reward in Eq. (17) therefore do not maximize a valid multi-user sum rate. The two-user result in Fig. 4 merely shows that one common message is decodable at two angles. Please either extend the system model with per-user precoding, scheduling, or subcarrier allocation that creates independent data streams, or reframe the contribution as multi-directional common-message broadcast and replace the sum-rate objective with an appropriate common-rate metric.","section":"Section II, Eq. (10) and Section IV, Fig. 4"},{"comment":"The state vector is described as s in R^{M x Q}, partitioned into M blocks with Q = Q1+Q2+Q3 entries, and later the text states that after 3M actions every block contains a single 1. This is inconsistent: each IRS element has three TM-IRS parameters (cmn, tau_o_mn, Delta_tau_mn), so there are 3M parameters. A one-hot representation would require either 3M blocks or a different encoding (e.g., a block size of Q1*Q2*Q3 per element). The stated network output dimension 2M x Q similarly reflects only M blocks. Please clarify the exact state, action, and network output dimensions; the current description is not reproducible.","section":"Section III-B, state representation"},{"comment":"The SINR expression has a malformed denominator: \"eta * sum_{i j=i-(K-1)} |V_j|^2 - |V0|^2 + sigma^2\" leaves the summation range ambiguous and does not define V_j for indices outside [0, K-1]. Since the achievable rate C in Eq. (9) and hence the reward in Eq. (17) depend directly on this SINR, please rewrite the summation explicitly (e.g., as sum_{j=0, j != i}^{K-1} |V_{i-j}|^2) and specify the definition and range of V_l.","section":"Section II, Eq. (8)"}],"minor_comments":[{"comment":"The claim that 9e5 trajectories out of roughly 10^65 configurations demonstrates \"remarkable efficiency\" would be more persuasive with a comparison against random sampling or a simple baseline; as presented, the decreasing TB loss does not by itself establish that the sampled configurations are near-optimal or that the search is more efficient than random search.","section":"Section IV"},{"comment":"The threshold xi_u = pi/5 is used for QPSK modulation, but Section II states that for M-PSK the threshold must be smaller than pi/M (pi/4 for QPSK). Please justify the specific choice and state whether it satisfies the required margin.","section":"Section IV, two-user experiment"},{"comment":"The procedure for sampling the four diverse TM-IRS configurations used in the time-varying security strategy is not described; please specify whether they are sampled from the trained forward policy and whether any temperature or filtering is applied.","section":"Section III-B and Section IV"},{"comment":"The paper contains several typographical and spacing errors (e.g., \"configured\" in the abstract, inconsistent hyphenation of \"sum rate\" and \"sum-rate\"); a careful proofreading pass is needed.","section":"Abstract and throughout"},{"comment":"The notation V(l, Omega_mn, theta, phi) is later abbreviated to Vi-k without an explicit definition of the indexing convention; please define Vi-k consistently, including the range of l for a finite number of harmonics used in the simulations.","section":"Section II, Eq. (5)-(7)"}],"recommendation":"major_revision","confidential_remarks":"The central multi-user claim needs substantial reworking; the state-space description also suggests the implementation may not match the paper, so the authors should verify their code against the described MDP. If the authors can either add a genuine multi-user transmission model or honestly reframe the results as common-message broadcast, the paper could become acceptable. The paper is within the journal's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent application of GFlowNets to a hard discrete parameter search for time-modulated IRS, with code and a fair single-user comparison. The advertised multi-user contribution is undercut by a real mismatch: the system model broadcasts one OFDM stream d(k) to all directions, so Eq. (10) sums per-user rates as if each user had an independent message when no independent messages exist. The two-user SER plot shows the same data decodable at two angles, which is a multi-direction broadcast, not multi-user transmission.\n\nWhat is genuinely new: prior TM-IRS design [18] is rule-based and single-direction, and this is the first GFlowNet-based parameter selection for TM-IRS that I know of. The MDP and trajectory-balance machinery is standard, but the application and the diversity-based security enhancement are legitimately new. The paper does the right thing empirically for the single-user case: Fig. 2 shows comparable SER at the intended direction and scrambling elsewhere, and the dynamic switching between diverse configurations (Fig. 3) is a sensible use of the sampler's mode coverage. Code is on GitHub, which makes the simulation reproducible.\n\nSoft spots, in order of severity. First, the multi-user sum-rate issue is load-bearing. Eq. (1) sends the same d(k) on every antenna; Eq. (7) is that same data convolved with V(i−k). There is no precoding, scheduling, or per-user subcarrier allocation. Every legitimate user receives the same message, so the total achievable rate is at most min_u C_u, not ∑ C_u. The stated objective of maximizing sum rate across authorized directions is not an information-theoretically valid multi-user metric. This does not affect the single-user results, but it invalidates the claim of supporting multiple independent users. It is fixable by adding transmit precoding or subcarrier assignment, or by reframing the contribution as multi-direction secure broadcast.\n\nSecond, the efficiency claim is under-evidenced. Converging in 9e5 trajectories out of ~10^65 configurations tells us little without comparison to random search, greedy search, or another generative baseline. The comparison to exhaustive search is trivially true; what matters is comparison to plausible heuristics.\n\nThird, no error bars or seed variance are reported. For a stochastic sampler, single-run SER curves make it hard to know how robust the policy is. Minor: the two-user experiment reports SER, not the sum rate that the objective nominally optimizes.\n\nThe circularity worry raised in the stress-test is not a real issue: the reward and SER come from the same simulation model, but no fitted constants are hidden, which is normal for this line of work.\n\nOverall, the paper is a solid, honest single-user study that overclaims on multi-user. I'd send it to review, but the multi-user framing needs a serious rewrite before acceptance.","headline":"GFlowNet for TM-IRS is a fresh, reproducible application, but the multi-user sum-rate claim is unsupported by the single-stream system model.","tokens_in":10608,"tokens_out":3853,"would_cite":true,"duration_ms":52317,"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 generative flow network can choose time-modulated reflecting-surface settings that serve multiple users while scrambling all other directions.","keywords":["intelligent reflecting surface","time modulation","directional modulation","physical layer security","GFlowNets","OFDM","Markov decision process"],"falsifier":"Transmit two distinct data streams to the two users at $(40^\\circ, 30^\\circ)$ and $(-40^\\circ, 30^\\circ)$ using the same trained GFlowNet configuration and check whether each user decodes its own stream; because Eq. (7) contains a single data symbol $d(k)$, the received signals at both directions will carry the same broadcast data, so the multi-user claim would collapse unless separate streams are actually delivered.","tokens_in":9665,"feed_emoji":"📡","tokens_out":5883,"duration_ms":53728,"temperature":0.7,"pith_summary":"This paper claims that the configuration of a time-modulated intelligent reflecting surface (TM-IRS) — which switches individual reflecting elements on and off in time to shape the transmitted OFDM waveform — can be chosen by a generative flow network rather than by hand-crafted rules. The goal is directional modulation: signals stay readable at the directions of legitimate users and are scrambled everywhere else. The authors cast configuration selection as a Markov decision process whose terminal states are full TM-IRS parameter settings and train a GFlowNet to sample settings with probability proportional to the achievable sum rate. Experiments show low symbol error rates at two user directions with scrambling elsewhere, and convergence after training on fewer than 0.000001% of the roughly $10^{65}$ configuration space. A sympathetic reader would take the paper as evidence that generative sampling can replace rule-based beamforming for physical-layer security in multi-user OFDM links.","feed_headline":"AI sampling picks secure reflecting-surface settings for multi-user links","feed_subtitle":"A GFlowNet trained on under a million configurations finds TM-IRS patterns that keep two users readable and scramble the rest.","key_machinery":"The load-bearing object is the GFlowNet with trajectory-balance loss: a generative model that treats each complete TM-IRS parameter set as a terminal state of a deterministic, acyclic Markov decision process, with actions assigning one discretized value to one parameter at a time. Flow-matching training makes the forward policy sample terminal states with probability proportional to the reward function $R = C_{\\mathrm{total}} \\prod_{u=1}^U H(\\xi_u - |\\arg(V_0)_u|)$, so high sum-rate configurations that respect constellation phase constraints are visited frequently without exhaustive search. The discrete parameter space is roughly $10^{65}$ configurations, and the trained network generalizes beyond the $9 \\times 10^5$ sampled trajectories.","core_discovery":"The central discovery is that a GFlowNet trained with the trajectory-balance objective can learn a stochastic policy over TM-IRS parameters (phase shifts, turn-on instants, and on-durations for each element) such that the sampling distribution is proportional to a sum-rate reward. The reward is the total achievable rate across authorized directions, multiplied by a Heaviside factor that enforces a per-user phase constraint on the main-tap coefficient $V_0$. In simulation with a $6 \\times 6$ IRS and 16 subcarriers, the policy yields near-zero symbol error rate at the intended directions while scrambling other directions, matches the single-user rule-based baseline, and extends to a second user at $(-40^\\circ, 30^\\circ)$. The authors further use the policy's diversity to switch among four configurations over time, averaging out accidental low-SER sidelobes.","pith_inferences":["One implication the paper leaves implicit: because the sampling policy is stochastic, an eavesdropper who knows the channel but not the sampling seed cannot predict which configuration is active, so the configuration space itself becomes a source of secret randomness; quantifying that secrecy gain would be a natural next step.","The efficiency result suggests the same flow-matching setup could handle other combinatorial parameter searches in physical-layer security, such as joint IRS phase and position selection or waveform design with discrete phase shifters, whenever an analytical reward is available.","The single-stream system model means the multi-user demonstration is, strictly speaking, a broadcast to several directions; to make the sum-rate claim a genuine multi-user throughput result, the transmitter would need per-user precoding or subcarrier allocation, which the paper does not describe."],"forward_implications":["Training on $9 \\times 10^5$ trajectories out of a roughly $10^{65}$-state space suffices to approximate the reward distribution, so exhaustive search is not needed for this class of configuration problems.","The stochastic sampling policy yields many distinct high-reward configurations, which lets the system switch TM patterns over time and average away directions that accidentally have low symbol error rate.","The same reward formulation extends from one legitimate user to multiple users by enforcing a phase constraint per user, as demonstrated for two users.","The approach matches the rule-based single-user baseline in desired-direction symbol error rate while also scrambling all other directions, making it a viable replacement for the existing closed-form design."],"supporting_citations":[{"why":"Provides the rule-based single-user TM-IRS design that the proposed method is compared against and replaces.","marker":"[18]"},{"why":"Introduces GFlowNets and the flow-matching principle that makes sampling proportional to reward possible.","marker":"[20]"},{"why":"Supplies the trajectory-balance loss used to train the forward and backward policies.","marker":"[24]"},{"why":"Establishes time-modulated OFDM directional modulation transmitters as the physical-layer security setting the paper builds on.","marker":"[15]"},{"why":"Gives the security analysis of time-modulated-array OFDM directional modulation that motivates scrambling and diversity-based defense.","marker":"[16]"}],"fun_headline_variants":["GFlowNets tune intelligent reflecting surface for secure multi-user links","AI picks reflecting surface settings that scramble signals for eavesdroppers","Learning to twiddle IRS knobs boosts security for many users at once","Generative AI designs secure time-modulated IRS for multi-user OFDM","Sampling smart IRS configurations with GFlowNets beats exhaustive search"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's multi-user claim depends on each legitimate user being served an independent data stream, but the system model transmits a single OFDM data symbol $d(k)$ identically in all directions with no described precoding or scheduling, so if users are not actually served independent messages, the sum-rate objective measures broadcasting the same data to several directions rather than multi-user communication.","fun_headline_variants_meta":{"raw":{"variants":["GFlowNets tune intelligent reflecting surface for secure multi-user links","AI picks reflecting surface settings that scramble signals for eavesdroppers","Learning to twiddle IRS knobs boosts security for many users at once","Generative AI designs secure time-modulated IRS for multi-user OFDM","Sampling smart IRS configurations with GFlowNets beats exhaustive search"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00081,"raw_usage":{"total_tokens":3570,"prompt_tokens":980,"completion_tokens":2590,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":2498}},"tokens_in":596,"tokens_out":2590,"duration_ms":17751,"temperature":1.0,"reasoning_tokens":2498,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:10:20.412794+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Transmit two distinct data streams to the two users at $(40^\\circ, 30^\\circ)$ and $(-40^\\circ, 30^\\circ)$ using the same trained GFlowNet configuration and check whether each user decodes its own stream; because Eq. (7) contains a single data symbol $d(k)$, the received signals at both directions will carry the same broadcast data, so the multi-user claim would collapse unless separate streams are actually delivered.","supporting_citations":[{"cited_title":"Time-modulated intelligent re flecting surface for waveform security,","cited_arxiv_id":null,"evidence_quote":"Provides the rule-based single-user TM-IRS design that the proposed method is compared against and replaces."},{"cited_title":"Flow network based generative models for non-iterative diverse candidate gen- eration,","cited_arxiv_id":null,"evidence_quote":"Introduces GFlowNets and the flow-matching principle that makes sampling proportional to reward possible."},{"cited_title":"Tra jectory balance: Improved credit assignment in gflownets,","cited_arxiv_id":null,"evidence_quote":"Supplies the trajectory-balance loss used to train the forward and backward policies."},{"cited_title":"Time-modulate d ofdm directional modulation transmitters,","cited_arxiv_id":null,"evidence_quote":"Establishes time-modulated OFDM directional modulation transmitters as the physical-layer security setting the paper builds on."},{"cited_title":"On the Security of Directional Modulation via Time Modulated Arrays Using OFDM Waveforms","cited_arxiv_id":"2408.10522","evidence_quote":"Gives the security analysis of time-modulated-array OFDM directional modulation that motivates scrambling and diversity-based defense."}],"review_version":1}