{"id":"24490d4c-c57c-4c17-af30-763f0623cb74","arxiv_id":"2505.15828","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"PG-ZFO combines a prompt-guided decision transformer with zero-forcing beamforming to solve QoE maximization for RIS-assisted digital twin interactions, allegedly generalizing to new scenes without retraining.","lead":"This paper proposes a generative AI method, prompt-guided decision transformer with zero-forcing optimization (PG-ZFO), for allocating wireless resources in digital twin interactions so that user-perceived quality of experience is maximized. The method is designed to work across changing digital twin scenes without retraining, which could matter for real-time networked applications like metaverse or remote education.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Generalization claim rests on an unspecified, mis-cited prompt-acquisition procedure (Sec. III.D.2, [13]); online prompts may require per-scene optimization, making the 'no re-training' claim unsupported and unreproducible.","rationale":"The reader's weakest assumption was essentially this: the prompt acquisition is mis-cited and dataset generation is unspecified. I agree, and I think this is the load-bearing point for the paper's main contribution. PG-ZFO is not claimed to be a better optimizer; it is claimed to generalize across scenes. That claim logically depends on the prompt carrying enough scene-specific structure to let a fixed transformer produce good decisions. If the prompt is obtained by an expensive DRL+ZF solve, the method has not removed the need for per-scene optimization; it has only moved it. If the prompt is a random or cheap trajectory, the paper gives no evidence that it would work. The offline data provenance is the other half of the same issue: MSE training against unstated episodes cannot establish what the transformer has learned. The ZF/water-filling subproblem is standard and the convergence plot is reasonable evidence that training works on the historical scenes, but neither addresses generalization. I therefore concur with the CONDITIONAL verdict; no change in verdict is needed, but the revision should specify both procedures and rerun the ablations above.","tokens_in":10170,"tokens_out":4790,"duration_ms":50312,"concrete_test":"Ask the authors for the prompt-acquisition code and dataset-generation code, or for a precise algorithmic description that replaces the incorrect [13] citation. Then rerun the three unseen-scene evaluations with two variants: (a) prompts produced by the specified DRL+ZF method, and (b) prompts replaced by random exploration trajectories of the same length T*. If variant (a) falls to DF-WP levels, or variant (b) matches variant (a), the gain is attributable to the prompt oracle rather than the transformer. Also record the wall-clock cost of prompt acquisition; if it scales like solving P(i), the 'without re-training' claim is misleading.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that PG-ZFO solves new scene-specific problems without time-consuming re-training (Sec. I). The mechanism is the prompt tau*(i), which is supposed to encode scene-specific information (Sec. III.B.1). In online execution (Sec. III.D.2), the prompt is obtained 'through a deep reinforcement learning (DRL) integrated with ZF-based optimization algorithm based prompt acquisition approach [13].' Reference [13] is Peel, Hochwald, and Swindlehurst (2005), a channel-inversion paper; it contains no prompt-acquisition method. The actual procedure is never described. This matters because if obtaining a prompt for a new scene requires running a DRL+ZF solver on that scene, then per-scene optimization has not been eliminated; the model is just imitating the solver's trajectory in the prompt. If prompts are cheap, the paper does not say how. The offline training set (Sec. III.D.1) is also under-specified: 40 scenes x 100 episodes are collected, but no generating policy or solver is stated. Since training minimizes MSE against the decisions in these episodes, the learned policy can only be as good as the data generator. If the same solver generates both offline labels and online prompts, the reported generalization to unseen scenes (Sec. IV.B) may be self-confirming. The claim may still hold, but the paper as written does not allow it to be checked.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies QoE-aware resource allocation in an RIS-assisted digital twin interaction system where the DT model evolves uncertainly, giving rise to a growing set of scene-specific optimization problems P(i). The authors formulate each per-scene problem to maximize the sum of user QoE over time slots by jointly choosing RIS phase shifts, receive/transmit beamforming, rendering resolution, and computing allocation. They reformulate each P(i) as an MDP and propose PG-ZFO: a prompt-guided decision transformer that outputs phase shifts, resolutions, and computing allocations, while a ZF-based algorithm computes the beamforming matrices in closed form. The central claim is that, once trained on historical scenes, PG-ZFO solves new unseen scene-specific problems without retraining by conditioning on a scene-specific prompt. Simulations compare PG-ZFO with a rigid optimization method and a prompt-free decision transformer, reporting superior QoE.","tokens_in":10434,"tokens_out":9366,"duration_ms":84183,"significance":"If the central claims are substantiated, the paper contributes a practical architecture for handling the combinatorial difficulty of scene-specific wireless optimization problems, and the prompt-guided transfer idea is relevant to the broader GAI-for-networking literature. The problem formulation is novel in its explicit treatment of joint uplink and downlink DT interaction QoE with uncertain evolution, and the ZF and water-filling subroutines in Section III.C are standard and executable. However, the key mechanism for generalization (the prompt) and the offline training data are not specified, and one of the optimization subproblems contains an equivalence error. Because of these gaps, the current evidence in Figures 2-4 does not yet establish the \"no re-training\" generalization claim. With the missing details supplied and the derivation corrected, the approach could be a credible contribution; in its present form the paper is not fully reproducible or verifiable.","major_comments":[{"comment":"The online prompt acquisition procedure is never described. The text states that the prompt tau*(i) is obtained \"through a deep reinforcement learning (DRL) integrated with ZF-based optimization algorithm based prompt acquisition approach [13]\", but reference [13] is a 2005 channel-inversion paper (Peel et al.) and contains no such method. Since the prompt is the mechanism that carries scene-specific information (Section III.B.1) and the paper's central claim is generalization to unseen problems without retraining (Section I), the paper must specify how the prompt is computed for a new scene, including the DRL formulation, the data used for that DRL, and its computational cost. If prompt acquisition requires per-scene optimization with DRL+ZF, the \"without time-consuming re-training\" claim is misleading.","section":"III.D.2, reference [13]"},{"comment":"The offline training dataset D_tra is under-specified. The paper states that for |I_tra|=40 scenes, 100 episodes per scene are \"collected\", but it never states which policy or solver generated the episodes, nor how the prompt subset D_star and the trajectory subset D_diamond are divided within each D(i_tra). Because training minimizes the MSE against the decisions in D_tra (Section III.D.1), the learned transformer can only replicate the behavior of whatever generated the dataset. If the same DRL+ZF solver is used both to generate offline labels and to acquire online prompts, the generalization results in Section IV.B may be self-confirming. Please state the generation policy, the prompt/trajectory division rule, and whether the dataset is available for reproducibility.","section":"III.D.1"},{"comment":"Problem (4) is not equivalent to problem (3). After zero-forcing, the downlink SINR of user k is p_k / (sigma^2 v_k), so problem (3) minimizes the sum over k of a_k / log2(1 + p_k/(sigma^2 v_k)) with a_k = (omega_iota_k * xi * E_k) / b, which is a sum of ratios of the form (positive constant) / log(1+gamma). Problem (4), however, maximizes the sum of b * log2(1 + p_k/sigma^2) / (xi * E_k * omega_iota_k), i.e., a sum of logarithms weighted by reciprocal constants. These two objectives are not equivalent for multiple users; the optimizers can differ. Moreover, the printed water-filling solution with the form p_k = max{1/rho - v_k sigma^2, v_k p_k^min} corresponds to an objective containing v_k inside the logarithm, which is not what Eq. (4) displays. The derivation must be corrected, or the water-filling step justified explicitly as a heuristic rather than a solution to problem (3).","section":"III.C, Eqs. (3)-(4)"}],"minor_comments":[{"comment":"The symbol G is used for both the number of transformer decoder layers (Section III.B.2) and the number of input tokens per mini-batch (Section III.D.1); Table I lists G=12 and G=16 without clarifying which is which. Please use distinct symbols or add a clarifying note.","section":"Table I"},{"comment":"The sentence \"dividing it into T_i = 1/F_i time slots with T_i = {1,2,...,T_i}\" abuses notation by using T_i for both a scalar cardinality and a set. Please use separate symbols, e.g., T_i and \\mathcal{T}_i.","section":"II.A"},{"comment":"In the benchmark description, the text introduces DF-WP and then states \"SDF does not consider the augmentation of prompts\", but the acronym SDF is never defined and appears to be a typo for DF-WP.","section":"IV.A"},{"comment":"Reference [11] is incomplete: \"Language models are few-shot learners\" lacks venue, year, and page or article identifier.","section":"References"},{"comment":"The text refers to an \"unseen problem P(i+1)\" but I is an index set and the notation is not formally defined. Please introduce a clear notation such as P_new for the held-out problem and state how its prompt is obtained.","section":"IV.B, Fig. 2"},{"comment":"The paper reports averages over 1000 runs but shows no error bars or confidence intervals; please add them or state that the differences are statistically significant, since some of the claimed improvements appear modest.","section":"IV.B, Figs. 3-4"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope, but the gap between the claimed contribution and the specified implementation is substantial. The mis-citation of reference [13] as a DRL-based prompt-acquisition method is a serious correctness and reproducibility issue that should be resolved during revision. The authors should also be prepared to demonstrate that the per-scene prompt acquisition cost does not dominate the avoided retraining cost. If the missing details and the optimization error in Section III.C are addressed, the paper could become a reasonable systems contribution; in its current form it is not yet verifiable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — worth a read if you work on RIS or digital twin resource allocation, but the headline claim needs more work before it can be trusted.\n\nWhat is actually new: the problem setup — QoE maximization for RIS-assisted digital twin interaction where the twin's evolution creates an unbounded sequence of scene-specific problems — is not in the cited literature. The overall architecture (prompt-guided decision transformer for phase shifts, resolution, compute; closed-form ZF/water-filling for beamforming) is a reasonable extension of existing decision-transformer work. The ZF and water-filling derivations in Sec. III.C are standard and, as far as I checked, correct. The paper also honestly places this as a within-subfield advance, not a breakthrough.\n\nThe soft spots are all about the generalization claim. The offline dataset D_tra is described only as 40 scenes x 100 episodes, with no generating policy or solver. Since training minimizes MSE against those episodes, the learned policy is bounded by whatever produced them. The prompt acquisition in Sec. III.D.2 is worse: it says prompts are obtained 'through a deep reinforcement learning integrated with ZF-based optimization algorithm based prompt acquisition approach' and cites [13], which is Peel et al. 2005, a channel-inversion paper. No DRL or prompt-acquisition method is described. If obtaining a prompt for a new scene requires running a DRL+ZF solver on that scene, then per-scene optimization has not been eliminated; the transformer just imitates the solver in the prompt. If prompts are cheap, the paper should say how. Also, constraint feasibility of the transformer's outputs is never addressed, and the evaluation uses three unseen scenes, no error bars, and no ablation where the prompt is replaced by a random trajectory.\n\nThese are specific, addressable problems, and the underlying idea may well hold. But as written, the central claim that PG-ZFO solves new scenes 'without time-consuming re-training' is not independently checkable, and the mis-citation needs fixing.\n\nUseful for: readers in the RIS/metaverse resource-management area who want a template for combining decision transformers with closed-form subproblems. It deserves a serious referee — an editor should send it out, but the review should push hard for dataset-generator and prompt-acquisition details, constraint handling, and tighter evaluation.\n\nMy call: conditional accept after major revision, and I would not cite it in its present form.","headline":"Plausible combination of decision transformer and ZF beamforming, but the advertised 'no re-training' generalization hinge is underspecified and mis-cited.","tokens_in":10965,"tokens_out":1812,"would_cite":false,"duration_ms":17242,"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":"The paper claims that a prompt-guided decision transformer plus zero-forcing beamforming can solve new digital-twin scenes without retraining.","keywords":["quality of experience","reconfigurable intelligent surface","digital twin","decision transformer","generative AI","zero-forcing beamforming","resource allocation","generalization"],"falsifier":"Train PG-ZFO on the 40 historical scenes, then evaluate it on held-out scenes whose parameters (user positions, personalized weights, Rician factors, or RIS location) are drawn from distributions with decreasing overlap with the training set. If the sum QoE drops to the level of the prompt-free decision transformer as the distance from the training scenes grows, the claimed generalization to unseen scenes is falsified. A second check is to rerun offline training with episodes generated by a clearly specified baseline policy; if results change materially, the reported generalization depended on the hidden data-generation rule.","tokens_in":9955,"feed_emoji":"📡","tokens_out":6602,"duration_ms":66144,"temperature":0.7,"pith_summary":"Digital twins mirror physical systems that keep changing, so every change creates a new optimization problem that an operator would otherwise have to solve from scratch. This paper claims that one offline-trained model can handle this stream of new scenes without retraining. The proposed approach, PG-ZFO, treats each scene as a Markov decision process, encodes the scene in a short prompt made of past good decisions, and lets a decision transformer generate the RIS phase shifts, rendering resolutions, and computing allocations while a zero-forcing step derives the beamformers. The authors' simulations show that PG-ZFO outperforms a fixed DRL policy and a prompt-free decision transformer on unseen scenes, and they attribute this to the prompt carrying scene-specific information the transformer can reuse.","feed_headline":"One offline AI model solves new digital-twin scenes without retraining","feed_subtitle":"The method pairs a prompt-guided transformer with zero-forcing beamforming to keep QoE high as the twin evolves.","key_machinery":"The load-bearing object is the prompt-guided decision transformer, an extension of a standard decision transformer. Its prompt is a sequence of returns-to-go, state, and action tuples sampled from the scene's historical trajectories; returns-to-go are the remaining reward until the scene's maximum, so the prompt implicitly encodes the scene-specific reward function and transition probabilities. At execution, the transformer takes the prompt plus the most recent trajectory, embeds both through trainable linear layers, processes them with stacked masked self-attention decoders, and emits the continuous decisions. The zero-forcing routine with water-filling closes the loop by deriving the beamforming matrices and power allocation in closed form, avoiding the curse of dimensionality in the action space.","core_discovery":"The paper's central claim is that the QoE-maximizing resource allocation problem for RIS-assisted digital twin interaction, which grows into many scene-specific problems as the twin evolves, can be solved by a single offline-trained model with no per-scene retraining. PG-ZFO reformulates each scene-specific problem as an MDP, builds a prompt from a short decision-making trajectory with returns-to-go that encodes the scene's reward structure and dynamics, and uses a causal transformer to output the phase-shift matrix, rendering resolution, and computing allocation. A zero-forcing based algorithm then supplies the uplink and downlink beamforming matrices, and a water-filling step sets downlink powers. In the reported simulations, the sum of user QoE increases as the training loss converges and stays above both benchmarks, which the authors present as evidence that the prompt-guided model generalizes to problems it never saw during training.","pith_inferences":["An extension the authors do not test is distribution shift: the paper evaluates only a few representative scenes, so a stress test sweeping scene parameters far beyond the 40 training scenes would show whether the claimed generalization is bounded by prompt relevance or is genuinely broad.","Because the prompt itself is a short expert trajectory, the method's ceiling depends on the quality of the prompt-acquisition solver; a stronger prompt-generation policy would likely increase PG-ZFO's gains without retraining.","The same architecture could be applied to other evolving resource-allocation problems, such as edge-computing task offloading or metaverse rendering, where each new environment state is a new optimization instance and the prompt can encode that instance's parameters.","A practical check for reproducibility is to require the episode-generation rule for the offline dataset to be stated explicitly, since the paper does not describe the policy that produced its training episodes."],"forward_implications":["If PG-ZFO generalizes as claimed, an operator can train once on historical digital twin scenes and then apply the same model to every new scene that appears as the physical entity evolves, avoiding repeated optimization or model retraining.","The prompt mechanism gives a concrete way to transfer scene-specific information between offline data and online execution, so the method is not just a policy learned on one scene.","Decomposing the action space—the transformer for phase shifts, resolution, and computing, and zero-forcing for beamforming—keeps the high-dimensional optimization tractable while preserving a joint QoE objective.","The QoE model combining Weber-Fechner perception with round-trip latency could be reused in other interactive services that need joint subjective and objective quality guarantees."],"supporting_citations":[{"why":"Provides the decision-transformer sequence-modeling architecture that PG-ZFO extends with scene prompts.","marker":"[12]"},{"why":"Motivates prompt-based few-shot adaptation, the mechanism PG-ZFO uses to encode scene-specific information.","marker":"[11]"},{"why":"Supplies the channel-inversion zero-forcing derivation used to compute receive and transmit beamforming matrices.","marker":"[13]"},{"why":"Establishes the Weber-Fechner logarithmic law used to model subjective perception quality in the QoE objective.","marker":"[9]"},{"why":"Frames digital twin services and the need to track evolving physical entities, motivating the uncertain-evolution setting.","marker":"[1]"},{"why":"Defines the rigid optimization benchmark that represents a fixed DRL policy without prompt-based generalization.","marker":"[14]"}],"fun_headline_variants":["One AI model adapts to new digital-twin scenes without retraining","Prompt-guided transformer keeps QoE high as digital twin evolves","Zero-forcing plus AI prompt solves evolving twin QoE without retraining","No retraining needed: prompt-guided transformer handles evolving digital-twin QoE","One AI model solves unseen digital-twin QoE problems"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The promise of no-retraining generalization rests on the offline dataset containing episodes that are representative of future scenes and generated by a competent policy, yet the paper does not specify how those episodes were produced.","fun_headline_variants_meta":{"raw":{"variants":["One AI model adapts to new digital-twin scenes without retraining","Prompt-guided transformer keeps QoE high as digital twin evolves","Zero-forcing plus AI prompt solves evolving twin QoE without retraining","No retraining needed: prompt-guided transformer handles evolving digital-twin QoE","One AI model solves unseen digital-twin QoE problems"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001141,"raw_usage":{"total_tokens":4737,"prompt_tokens":945,"completion_tokens":3792,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":3697}},"tokens_in":561,"tokens_out":3792,"duration_ms":27358,"temperature":1.0,"reasoning_tokens":3697,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:21:25.552980+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train PG-ZFO on the 40 historical scenes, then evaluate it on held-out scenes whose parameters (user positions, personalized weights, Rician factors, or RIS location) are drawn from distributions with decreasing overlap with the training set. If the sum QoE drops to the level of the prompt-free decision transformer as the distance from the training scenes grows, the claimed generalization to unseen scenes is falsified. A second check is to rerun offline training with episodes generated by a clearly specified baseline policy; if results change materially, the reported generalization depended on the hidden data-generation rule.","supporting_citations":[{"cited_title":"Decision transformer: Reinforce- ment learning via sequence modeling,","cited_arxiv_id":null,"evidence_quote":"Provides the decision-transformer sequence-modeling architecture that PG-ZFO extends with scene prompts."},{"cited_title":"Language models are few-shot learners","cited_arxiv_id":null,"evidence_quote":"Motivates prompt-based few-shot adaptation, the mechanism PG-ZFO uses to encode scene-specific information."},{"cited_title":"A vector-perturbation technique for near-capacity multiantenna multiuser communication-Part I: channel inversion and regularization,","cited_arxiv_id":null,"evidence_quote":"Supplies the channel-inversion zero-forcing derivation used to compute receive and transmit beamforming matrices."},{"cited_title":"Logarithmic laws in service quality perception: where microeconomics meets psychophysics and quality of experience,","cited_arxiv_id":null,"evidence_quote":"Establishes the Weber-Fechner logarithmic law used to model subjective perception quality in the QoE objective."},{"cited_title":"Digital twins: A survey on enabling technologies, challenges, trends and future prospects,","cited_arxiv_id":null,"evidence_quote":"Frames digital twin services and the need to track evolving physical entities, motivating the uncertain-evolution setting."},{"cited_title":"A three-party hierarchical game for physical layer security aware wireless communications with dynamic trilateral coalitions,","cited_arxiv_id":null,"evidence_quote":"Defines the rigid optimization benchmark that represents a fixed DRL policy without prompt-based generalization."}],"review_version":1}