{"id":"248a6d1d-ae8d-4620-b793-ac38696dfde8","arxiv_id":"2608.13032","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"The authors propose a two-stage action-space compression for RL-based RIS resource allocation, claiming lossless Pareto mapping plus autoencoder compression yields faster training and lower cost, but the lossless claim is flawed.","lead":"This paper proposes PAAERL, a reinforcement learning framework that compresses high-dimensional RIS antenna settings into a small latent code via a Pareto mapping and an autoencoder, to speed up online resource allocation in IoT networks. The simulations show faster training and lower cost in a RIS-assisted mobile edge computing setup, but the core theoretical guarantee that the compression loses no Pareto information is not valid for non-convex rate regions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Directional max-min mapping is not bijective and not even well-defined: P3's inequality constraints admit multiple rate vectors per direction, so the lossless Pareto-representation guarantee in Section III.C is unproven.","rationale":"The reader correctly identifies a false geometric premise: the ray through a Pareto point is unique, but this does not imply a one-to-one correspondence between all positive-orthant directions and Pareto-optimal points. However, the specific L-shape counterexample is slightly imprecise because P3 uses inequality constraints, so for omega = (1,1) the optimum may also return Pareto points such as (2,1) or (1,2), depending on the solver. The precise failure is non-uniqueness and the presence of dominated points in the optimal set, which still invalidates the paper's stated bijection. The deeper issue is that the theoretical guarantee of no information loss is used to justify replacing {f, beta, theta} with omega; the paper does not prove the weaker but sufficient surjectivity claim, and it also does not prove that the AO/SCA feasibility solver can certify feasibility with zero slack. Because the central theoretical contribution is a claimed exact, lossless mapping, and that claim is false as stated, the reader's reject verdict remains appropriate. The approach may be repairable, but the manuscript as written does not contain the required corrected proof or solver guarantees.","tokens_in":16755,"tokens_out":19013,"duration_ms":223545,"concrete_test":"Enumerate all solutions of P3 for a two-user non-convex comprehensive rate region, e.g., R = ([0,2]x[0,1]) union ([0,1]x[0,2]), at omega = (1,1). Check whether the optimal rate vector set is a singleton and consists only of Pareto-optimal points; if the set contains dominated points or depends on solver tie-breaking, the bijective mapping claim fails. Separately, maximize a monotone objective f(R) over all omega and compare with the true Pareto-optimal objective value; if equality holds, the weaker no-information-loss claim survives and the paper needs only a corrected theorem and a feasibility-certification proof, but the current manuscript does not provide them.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing claim is the bijective directional mapping M in Section III.C. The proof is invalid: uniqueness of the ray through a Pareto point does not imply that every positive-orthant direction has a Pareto point, and P3's constraints R_k >= omega_k^PT R_s are inequalities, not equalities. For a comprehensive non-convex region (e.g., R = ([0,2]x[0,1]) union ([0,1]x[0,2])), with omega = (1,1), the maximum R_s is 1, and the optimal feasible rate set contains dominated points such as (1,1) as well as Pareto points such as (2,1) and (1,2); the returned rate vector is therefore solver-dependent and M is not uniquely defined. The paper's assertion that searching omega attains the global optimum of any nondecreasing objective can be salvaged by a weaker surjectivity argument (every Pareto point p is obtained by taking omega proportional to p), but that corrected proof is not what the paper supplies. Moreover, the inner AO/SCA feasibility solver needed to realize even the weaker claim is not validated: P5's slack can be positive at a stationary point even when a zero-slack feasible configuration exists, so the binary search in Algorithm 1 may misclassify feasible R_s as infeasible. The central theoretical guarantee is therefore not established as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes PAAERL, a hierarchical reinforcement-learning framework for online resource allocation in RIS-assisted multi-user IoT/MEC systems. The first compression stage replaces the high-dimensional physical variables (transmit beamforming, RIS phase shifts and amplitudes) with a positive weight vector ω_PT by solving a directional max-min scaling problem P3, and claims that this establishes a bijective, information-lossless mapping between weight vectors and the Pareto-optimal frontier of the achievable rate region, regardless of convexity. A second, data-driven autoencoder compresses the priority profile into a low-dimensional latent action space for a PPO agent. The paper provides Algorithms 1 and 2, a complexity analysis, and simulations in an RIS-assisted MEC scenario comparing six strategies on convergence speed, total cost, and training time.","tokens_in":16901,"tokens_out":12500,"duration_ms":141907,"significance":"If the bijective Pareto-representation claim were valid, the proposed model-driven plus data-driven compression would be a conceptually attractive way to decouple physical-layer optimization from high-level control, and the simulation results suggest practical gains in training time and cost. The paper is also transparent that the autoencoder stage is lossy, which is a reasonable design choice. However, the central theoretical guarantee—the claimed lossless substitution of the physical variables by the weight vector—is false as stated. The directional max-min problem P3 is not a single-valued map onto the Pareto frontier for non-convex comprehensive rate regions, and the inner feasibility solver is not shown to be an exact oracle. Because the hierarchical RL design and the paper's central 'Pareto-aware' claim rest on this guarantee, the main contribution is not established in the submitted form.","major_comments":[{"comment":"The claimed one-to-one correspondence between positive-orthant direction vectors and Pareto-optimal points is false for non-convex comprehensive rate regions. For R = ([0,2]×[0,1]) ∪ ([0,1]×[0,2]) and ω=(1,1), P3 gives max R_s = 1, but the feasible rate set at R_s=1 contains the dominated point (1,1) as well as the Pareto points (2,1) and (1,2); hence the ray in direction ω does not terminate at a Pareto point, and P3 does not define a unique rate vector. The observation that the ray through a given Pareto point is unique does not imply that every ray contains a Pareto point. This invalidates the deterministic mapping M:ω_PT→R* and the 'no theoretical information loss' claim that follows it.","section":"Section III.C, P3 (Eq. (11))"},{"comment":"Even apart from the counterexample, the constraints in P3 are R_k ≥ ω_k R_s, not equalities. Consequently, for a fixed optimal R_s there are generally many feasible rate vectors satisfying the constraints, so the output of P3 depends on the arbitrary choice of the inner feasibility solver; the paper never specifies a tie-breaking rule. A directional representation of the Pareto frontier requires either equality constraints or an additional projection step that selects a definite rate vector. The existence result that can actually be proved is weaker: for every Pareto point p, choosing ω=p and solving P3 returns p (if the feasibility oracle is exact). The paper should replace the bijection theorem with this statement and re-derive the RL justification from it.","section":"Section III.C, P3 constraints"},{"comment":"Algorithm 1's binary search uses P4/P5 as an exact feasibility oracle, but P5 is not validated as one. P5 minimizes the sum of slack variables d_k without imposing d_k ≥ 0, so a zero-slack feasible point need not be the minimizer. More importantly, P5a/P5b replace the nonconvex signal-power term by an SCA lower bound; this relaxes the constraint, and the alternating SCA procedure is only claimed to reach a stationary point of the relaxed problem. A stationary point can have positive slack even when P4 is feasible, causing the bisection to misclassify a feasible R_s as infeasible. The paper provides no convergence or exactness proof showing that the binary search returns the true maximum R_s; the statement that the 'SCA lower bounds guarantee that the objective value is non-decreasing and bounded' is not by itself sufficient for a minimization problem.","section":"Section III.D, P5 and Algorithm 1"}],"minor_comments":[{"comment":"Please clean up the notation typos: the reference to Table I appears as 'Table.', and Eq. (6) has bracket/brace errors around the logarithm expression.","section":"Section II and Eq. (6)"},{"comment":"The vector ω_PT is introduced before its normalization or domain is defined; specify whether it is normalized to sum to one, lies on the unit sphere, or is otherwise made scale-invariant, and explain how scaling redundancy is handled.","section":"Section III.C"},{"comment":"The instruction to 'find R_max when P4 has no feasible solution' is ambiguous; the bisection initialization and the convention for checking feasibility at R_s=0 should be stated precisely.","section":"Algorithm 1, line 11"},{"comment":"The statement that imperfect AE reconstruction 'will only cause a small amount of performance loss' is not quantified; the simulations do not report reconstruction error or its effect on the achieved cost, so this claim is currently unsupported.","section":"Section III.E"},{"comment":"The convergence and scaling plots do not include error bars or multiple-seed statistics, so the reported improvements over WSRRL and PARL cannot be separated from run-to-run variance.","section":"Section IV"}],"recommendation":"reject","confidential_remarks":"The false bijection claim is not a presentation issue; it is the central theoretical foundation of the paper. A corrected version would need to replace the one-to-one claim with a weaker existence theorem and would need to repair or re-validate the feasibility oracle used in Algorithm 1. Those changes go beyond local revision, and the empirical study does not compensate for the invalid guarantee as submitted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead this one. The headline is that the theoretical centerpiece doesn't hold: the claim in Section III.C that P3 sets up a one-to-one correspondence between direction vectors and the Pareto frontier is false. In a non-convex but comprehensive rate region, a ray can leave the feasible set at a dominated point, and P3's inequality constraints don't force the solver to pick a Pareto point even when the max scaling is achieved. The L-shaped region example (union of [0,2]×[0,1] and [0,1]×[0,2], direction (1,1)) is exactly right: the optimal R_s is 1, and (1,1) is feasible but dominated by (2,1) and (1,2). So the \"no theoretical information loss\" guarantee collapses, and the RL agent's action space is contaminated with directions that map to dominated rate vectors.\n\nThat said, the paper is not a mess. The basic idea—compress the RIS beamforming variables to a low-dimensional priority vector using a geometric mapping, then use an autoencoder to compress that vector further for RL—is a reasonable engineering direction. The authors include an ablation (PARL without autoencoder) and compare against a decent set of baselines. They also honestly report that when the number of RIS elements grows large, PARL and WSRRL beat their full PAAERL method (Fig. 5), which quietly contradicts the abstract's sweeping claim.\n\nThe soft spots beyond the bijectivity issue: the inner feasibility solver uses SCA with slack variables, and there is no proof that a zero-slack solution is found when one exists, so the binary search in Algorithm 1 may misclassify feasible points. The experiments have no error bars, no code, and the conclusion overstates the results. On the theoretical side, a weaker surjectivity result is probably true—every Pareto point can be reached by taking the direction through that point—but that is not what the paper proves, and the uniqueness claims are not salvageable as written.\n\nWho should read this? Someone working on RL for RIS or hybrid model/data-driven action compression might get ideas for what to fix. It is not a publishable paper in its current form. But it deserves a real referee, not a desk reject: the flaw is specific and a competent reviewer could identify it and require a rewrite that either proves the weaker claim or abandons the lossless-language entirely.\n\nRecommendation: send to peer review, expect a major revision or reject after review. Don't take the abstract's performance claims at face value.","headline":"The directional Pareto mapping that carries the paper is not a bijection, so the lossless-compression claim breaks; the engineering is a repairable idea but the paper's central theorem is false as stated.","tokens_in":17540,"tokens_out":3873,"would_cite":false,"duration_ms":39674,"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":"This paper claims that every efficient rate allocation in a RIS-assisted IoT network can be encoded by a direction vector, allowing reinforcement learning to act on a tiny space instead of thousands of beamforming and phase variables.","keywords":["Reconfigurable intelligent surface","Hierarchical reinforcement learning","Pareto optimality","Directional max-min scaling","Autoencoder compression","Online resource allocation","Mobile edge computing","Internet of Things"],"falsifier":"Take a two-user rate region that is the union of the downsets of $(1,10)$ and $(10,1)$. This region is non-convex and contains $(1,1)$, but the point $(1,10)$ dominates $(1,1)$. Solving P3 with direction $\\omega=(1,1)$ returns $R_s=1$ and the rate vector $(1,1)$, a dominated point, even though a Pareto point exists with the same first coordinate. This directly contradicts the claimed one-to-one correspondence and can be checked with the paper's own feasibility solver on two users.","tokens_in":16437,"feed_emoji":"📡","tokens_out":10319,"duration_ms":102970,"temperature":0.7,"pith_summary":"This paper proposes a way to make reinforcement learning practical for real-time resource allocation in RIS-assisted IoT networks by cutting the action space in two stages. The first stage replaces the high-dimensional physical variables—transmit beamformers, RIS phase shifts, and reflection amplitudes—with a low-dimensional weight vector $\\omega$ using a directional max-min scaling problem. The paper argues this mapping is lossless: every Pareto-optimal rate vector of the achievable rate region, convex or not, is reachable by some direction $\\omega$, so no efficient operating point is lost. The second stage compresses $\\omega$ further with an offline-trained autoencoder, so the RL agent acts on a tiny latent vector while a frozen decoder and the model-driven projection restore physical configurations. Simulations on a multi-user mobile edge computing network indicate faster convergence, shorter training time, and lower delay-plus-energy cost than weighted-sum-rate or uncompressed RL baselines.","feed_headline":"A direction vector reaches every efficient rate point","feed_subtitle":"A max-min scaling map plus an autoencoder lets one RL agent control dense RIS IoT networks in real time.","key_machinery":"The load-bearing object is the directional max-min scaling problem P3. It replaces linear scalarization (weighted sum-rate) with a search along a ray: for a direction vector $\\omega$, maximize a scalar $R_s$ such that each user's rate is at least $\\omega_k R_s$; the outer loop is a binary search on $R_s$, and the inner loop is a feasibility problem P4 solved by alternating optimization with successive convex approximation, producing convex subproblems for beamforming and for RIS coefficients. The second mechanism is a symmetric autoencoder whose frozen decoder $D_\\psi$ is embedded in the online loop, translating the RL agent's low-dimensional latent action $Z$ into the weight-and-application vector $(\\omega, X)$ that P3 then maps to a Pareto-optimal physical configuration.","core_discovery":"The central claim is that the set of Pareto-optimal operating points of an RIS-assisted multi-user network can be indexed bijectively by normalized direction vectors in the positive orthant, and that the directional max-min scaling problem P3—maximize a common scaling factor $R_s$ subject to $R_k \\ge \\omega_k R_s$ for every user—recovers the unique Pareto-optimal rate vector for a chosen direction $\\omega$. Because any sensible objective is non-decreasing in the user rates, its global optimum lies on this frontier, so optimizing over $\\omega$ and application-level variables $X$ is equivalent to optimizing over the original $\\{f,\\beta,\\theta,X\\}$. The paper then chains an autoencoder onto this mapping, so the RL agent outputs only a latent code $Z$; the decoder reconstructs the weight vector, and P3 projects it onto the Pareto boundary. The claimed payoff is that PAAERL inherits the geometric exactness of P3 while avoiding both action-space explosion and the information loss the paper attributes to weighted sum-rate scalarization in non-convex rate regions.","pith_inferences":["If the one-to-one claim fails, PAAERL would degrade gracefully into an approximate compressor whose error is the depth of the dent in the rate region; the paper's own caveat about autoencoder reconstruction error already admits approximate behavior in the second stage.","A minimal two-user test of the geometric claim is to run P3 for direction $(1,1)$ on a rate region formed by the union of the downsets of $(1,10)$ and $(10,1)$ and check whether the returned point is dominated.","If the geometric mapping is only approximately bijective in practice, the autoencoder could be trained to avoid directions that point into dents, effectively learning which rays to distrust; this is a testable extension the paper does not explore."],"forward_implications":["The number of RIS elements $M$ stops being part of the RL action space: once the P3 projection is in place, the policy sees only $K$ weight coordinates plus application variables, regardless of $M$.","Any network objective that is non-decreasing in user rates—latency, energy, sum rate—can be optimized by searching the same low-dimensional weight space, so the architecture transfers to STAR-RIS, UAV-relayed, and other coupled RIS systems.","Because the autoencoder is trained offline on Pareto-optimal profiles, online execution is a single decoder pass plus one P3 projection per slot, making decisions fast enough for quasi-static channels even in dense deployments.","The paper's convergence plots show the Pareto-aware variants reach lower steady-state cost than weighted-sum-rate variants, which the paper reads as evidence that the non-convex 'dented' parts of the rate region are exactly where P3 helps."],"supporting_citations":[{"why":"Defines the weighted sum-rate scalarization that the paper replaces; supplies the baseline whose convex-region limitation motivates P3.","marker":"[10]"},{"why":"Shows the achievable rate region of an RIS-assisted interference channel can be non-convex, justifying the need for a non-convex Pareto mapping.","marker":"[11]"},{"why":"Similarly documents that weighted sum-rate optimization in non-convex rate regions is an outer approximation, the limitation P3 is designed to fix.","marker":"[12]"},{"why":"Provides the conventional weighted sum-rate formulation for RIS-aided networks that the directional max-min scaling method is compared against conceptually.","marker":"[21]"},{"why":"Supplies the autoencoder-based communication idea that the paper adapts for secondary latent-space compression.","marker":"[22]"},{"why":"Represents an RL approach to RIS phase-shift optimization, used as a related baseline for the hierarchical RL design.","marker":"[15]"},{"why":"Documents action-space explosion in continuous RL for RIS systems, the problem the dual compression targets.","marker":"[16]"}],"fun_headline_variants":["One direction vector hits every Pareto-optimal rate","RL for RIS IoT: direction vectors replace beamforming","Autoencoder compresses Pareto space for fast RIS RL","Pareto-aware RL speeds up dense RIS IoT optimization","Direction vector mapping tames RIS IoT resource allocation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each direction vector maps to exactly one Pareto-optimal rate point, so replacing the physical variables with $\\omega$ loses no information even when the rate region is non-convex.","fun_headline_variants_meta":{"raw":{"variants":["One direction vector hits every Pareto-optimal rate","RL for RIS IoT: direction vectors replace beamforming","Autoencoder compresses Pareto space for fast RIS RL","Pareto-aware RL speeds up dense RIS IoT optimization","Direction vector mapping tames RIS IoT resource allocation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000528,"raw_usage":{"total_tokens":2578,"prompt_tokens":1009,"completion_tokens":1569,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":625,"completion_tokens_details":{"reasoning_tokens":1494}},"tokens_in":625,"tokens_out":1569,"duration_ms":10508,"temperature":1.0,"reasoning_tokens":1494,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:10:55.486990+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a two-user rate region that is the union of the downsets of $(1,10)$ and $(10,1)$. This region is non-convex and contains $(1,1)$, but the point $(1,10)$ dominates $(1,1)$. Solving P3 with direction $\\omega=(1,1)$ returns $R_s=1$ and the rate vector $(1,1)$, a dominated point, even though a Pareto point exists with the same first coordinate. This directly contradicts the claimed one-to-one correspondence and can be checked with the paper's own feasibility solver on two users.","supporting_citations":[{"cited_title":"Weighted sum-rate maximization by joint antenna grouping and movable RIS deployment,","cited_arxiv_id":null,"evidence_quote":"Defines the weighted sum-rate scalarization that the paper replaces; supplies the baseline whose convex-region limitation motivates P3."},{"cited_title":"Achievable rate region maxi- mization in intelligent reflecting surfaces-assisted interference channel,","cited_arxiv_id":null,"evidence_quote":"Shows the achievable rate region of an RIS-assisted interference channel can be non-convex, justifying the need for a non-convex Pareto mapping."},{"cited_title":"Achievable rate region of MISO interference channel aided by intelligent reflecting surface,","cited_arxiv_id":null,"evidence_quote":"Similarly documents that weighted sum-rate optimization in non-convex rate regions is an outer approximation, the limitation P3 is designed to fix."},{"cited_title":"Autoencoder- based communications with reconfigurable intelligent surfaces,","cited_arxiv_id":null,"evidence_quote":"Supplies the autoencoder-based communication idea that the paper adapts for secondary latent-space compression."},{"cited_title":"Deep reinforcement learning for practical phase-shift optimization in RIS- aided MISO URLLC systems,","cited_arxiv_id":null,"evidence_quote":"Represents an RL approach to RIS phase-shift optimization, used as a related baseline for the hierarchical RL design."},{"cited_title":"Learning to Focus: CSI-Free Hierarchical MARL for Reconfigurable Reflectors","cited_arxiv_id":"2604.05165","evidence_quote":"Documents action-space explosion in continuous RL for RIS systems, the problem the dual compression targets."}],"review_version":1}