{"id":"5f46ae15-e67e-4142-b37d-efaf965b3b79","arxiv_id":"2411.18010","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A joint SLM prompt compression and DRL power allocation framework cuts simulated wireless LLM response time by about 17% while keeping reported fidelity near 0.9.","lead":"This paper wires a small language model into user-side prompt compression and uses deep reinforcement learning to set both the compression ratio and wireless transmit power for LLM services. A generalist might read it to see whether on-device prompt trimming can cut latency in mobile LLM services without wrecking answer quality.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (4)'s fidelity and the DRL reward are never defined, so the objective being optimized is not a well-defined function of the state; absent those formulas, the ~0.9 fidelity curves and 17% latency claim cannot be reproduced.","rationale":"The reader's weakest assumption is precisely the one I would attack: the objective f is both the optimization target and the evaluation metric, so its under-specification is not a cosmetic omission. A well-defined f is necessary to solve Eq. (10), to define the DRL state and reward, and to interpret Fig. 5(a). The paper contains no code, no error bars, and no baselines; the two worked examples (44 and 388 tokens) give reductions of 16.4% and 16.5%, consistent with the ~17% headline but without statistical support. Given the missing equations, the central claim does not survive scrutiny. I find no independent support that would change this: no formal verification, no public code, no parameter-free derivation. I agree with the reader's REJECT verdict; no adjustment is needed.","tokens_in":8173,"tokens_out":5731,"duration_ms":53803,"concrete_test":"Run a small end-to-end experiment on MeetingBank transcripts with the described stack (LLMLingua/GPT-Neo-125M compressor, GPT-J-6B responder) at compression ratios κ ∈ {0.25, 0.5, 1.0}: define f by fixing a similarity metric (e.g., cosine similarity of sentence embeddings) for f1/f3 and an explicit f2 that includes BER, then measure f alongside true task accuracy (exact/answer-match against reference responses), and compute the Spearman correlation between f and true accuracy. If the correlation is below ~0.5, the Eq. (10) objective is not a valid proxy for service quality and the 'high fidelity' claim is an artifact of the unspecified metric; if the paper's text alone is insufficient to fix f, that itself confirms the under-specification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that JPPO 'effectively balances service quality with resource efficiency' and reaches fidelity ≈0.9, BER < 0.2, and ~17% latency reduction. The load-bearing object is the fidelity f in Eq. (10), defined only abstractly by Eq. (4) as f = α1 f1 + α2 f2 + α3 f3. Section II-B gives qualitative glosses: f1 and f3 are 'similarity metric' comparisons, f2 is 'token retained ... considering BER,' but no similarity metric, no f2 formula, and no dependence on η(PT) is written. Section III-B's reward is equally underspecified ('maximizes fidelity while minimizing penalties related to BER and power usage'). The state includes current fidelity f_n, which is not observable before action in a real deployment; if it is a simulator artifact, the paper never says how it is computed. Thus the optimization problem (10) is not a well-defined mathematical object, and Figs. 4–5 cannot be checked or reproduced. (Algorithm 1 line 10/Eq. (11) also misstate the Bellman update, but that is secondary to the missing objective.)","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes JPPO, a framework that combines SLM-based prompt compression at the user device with Double DQN-based joint optimization of compression ratio and transmission power for wireless LLM inference services. The stated contributions are an adjustable prompt compression mechanism, a joint optimization problem balancing fidelity, energy, latency, and power, and a DRL solution. Experimental results on a MeetingBank-based dataset report convergence of training reward, average fidelity around 0.9, BER below 0.2, power in the 4–5 W range, and an average response-time reduction of about 17%.","tokens_in":8424,"tokens_out":2550,"duration_ms":23402,"significance":"If the results hold, JPPO addresses a timely and practical problem: reducing the communication and inference latency of long-prompt LLM services over wireless links. The manuscript introduces a sensible system decomposition (SLM compression, wireless transmission, LLM inference) and a concrete latency breakdown with measured end-to-end times for two prompts. The use of real SLM/LLM components (GPT-Neo 125M and GPT-J 6B) is a strength. However, the central optimization objective and the DRL reward are never defined mathematically, several equations are internally inconsistent, and the reported fidelity/BER figures are not compared against any baselines or accompanied by error bars. As a result, the main claims of 'high service fidelity' and 'effectively balances service quality with resource efficiency' cannot currently be verified or reproduced.","major_comments":[{"comment":"The fidelity metric f = α1 f1 + α2 f2 + α3 f3 is not defined as a function of the decision variables. The text states that f1 and f3 are 'similarity metric' comparisons and that f2 is 'token retained ... considering BER,' but no similarity metric (e.g., cosine similarity, BLEU, ROUGE) is specified, no formula for f2 is given, and the dependence on η(PT) is never written. Since Eq. (10) maximizes f(κ, η(PT)), the optimization problem is not a well-defined mathematical object. Figures 4 and 5(a) therefore cannot be checked or reproduced without the missing definitions.","section":"Section II-B, Eq. (4)"},{"comment":"The DRL reward function in Section III-B is described only qualitatively: the agent 'maximizes fidelity while minimizing penalties related to BER and power usage.' No explicit reward formula is provided. The reader cannot determine what scalar objective the Double DQN is actually optimizing, how the penalties are weighted relative to fidelity, or how the fidelity components enter the reward. This is a load-bearing gap because the convergence curves in Fig. 4 and the fidelity/BER/power curves in Fig. 5 depend entirely on the unspecified reward design.","section":"Section III-B, Reward definition"},{"comment":"The Q-learning update is algebraically inconsistent. Algorithm 1 line 10 sets y = R' + [α rt+1 + γ max_a Q(st+1, a) − Q(st, at)], which is not a valid target value: it adds the current Q-value to the TD error, and it uses α as a multiplier on rt+1 while Eq. (11) defines α as the learning rate. Equation (12) then writes the loss with a target r + μ max Qtarget but omits the Double DQN current-network action selection that Eq. (13) claims to implement. These inconsistencies make it impossible to reproduce the training procedure or to know which update was actually run.","section":"Section III-B, Algorithm 1 and Eq. (11)"},{"comment":"There are no baseline comparisons or error bars. The 17% latency improvement is supported only by two examples (44 and 388 tokens) plus a statement that 'evaluation across our test dataset showed that the overall performance improvement fluctuates around 17%,' with no standard deviation or number of test samples. The fidelity and BER curves in Fig. 5 are averages over episodes but no variance is reported. Given that the fidelity f used in Fig. 5(a) is the same quantity used as the DRL objective, the fidelity result is in part a circular measure of success; independent task-accuracy evaluation is not provided.","section":"Section IV, Numerical Results"}],"minor_comments":[{"comment":"The state vector is written as '[fn(ηn), γn, ]' with a trailing comma and missing third element; this appears to be a typographical error.","section":"Section III-B, State definition"},{"comment":"Equation (9) defines total delay as tSLM_e + tLLM_e + tt, but Eq. (6) groups both SLM and LLM GPU time under Ee; the relationship between the two time components and the encoding energy is not clearly explained.","section":"Section II-D and Eq. (9)"},{"comment":"The objective notation f(κ, η(PT)) introduces η(PT) as the BER, but no BER model (e.g., Q-function of SNR) is given anywhere, so the optimizer's dependence on PT through η is not explicit.","section":"Section III-A, Eq. (10)"},{"comment":"The row 'Total test runs range [1, 10]' is unclear, and the number of episodes per test run is listed as 10,000 while Fig. 5 extends to 12,000 episodes; please reconcile.","section":"Section IV, Table I"},{"comment":"The MeetingBank-transcript dataset citation is missing (shown as '[?]'), which prevents reproducibility of the experimental setup.","section":"Section IV, Dataset reference"},{"comment":"The symbol α denotes both the fidelity weights in Eq. (4) and the learning rate in Eq. (11); this notation clash is confusing and should be resolved.","section":"Eqs. (4) and (11)"},{"comment":"The algorithm input mentions 'target privacy (ϵ, δ)', but ϵ is later used as the exploration rate and δ is never defined; please correct the terminology.","section":"Algorithm 1, Input line"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is compact and reads like a workshop/conference extension; for a journal venue, the missing mathematical definitions of the fidelity metric and the DRL reward are disqualifying in the current form. The latency reduction measurement is the most credible part of the paper, but it is presented without statistical support. The inconsistent Q-learning update in Algorithm 1/Eq. (11) suggests the authors may need to re-run and re-report all training results after correcting the update rule. I would encourage a thorough revision rather than outright rejection, because the underlying system concept is relevant and the use of real SLM/LLM models is a plus."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper's real contribution is the combination, not any single component: SLM-based prompt compression (LLMLingua-style) plus DRL-driven transmit power allocation for wireless LLM serving. That joint problem is new, and the latency measurement is an independent number: 56.1s to 46.9s and 85.3s to 71.2s on two prompt lengths, roughly 17% less. That part is concrete and worth taking seriously.\n\nWhat the paper does well: it identifies a real bottleneck (long prompts over wireless links), and the system model in Section II is clearly laid out. The Double DQN formulation is a standard and reasonable tool for the joint control problem. The example in Fig. 3 is useful.\n\nThe soft spots are load-bearing. The fidelity metric f in Eq. (4) is never concretely defined: f1 and f3 are only described as 'similarity metric' comparisons, f2 as 'token retained considering BER,' but no formula or specific metric (e.g., cosine similarity on which embeddings?) is given. The DRL reward is only described qualitatively ('maximizes fidelity while minimizing penalties related to BER and power usage'). So the optimization problem (10) is not a well-defined mathematical object; Figs. 4 and 5 plot an undefined quantity. Worse, the fidelity used as the optimization objective is the same fidelity reported as evidence of success in Fig. 5(a), so the evaluation is circular. The Bellman update in Eq. (11) is algebraically wrong (it mixes learning rate and discount factor, and line 10 of Algorithm 1 duplicates the reward). There are no baselines, no error bars, and the dataset is referenced only as '[?]'. These are not minor issues; they prevent reproducibility.\n\nTo be fair, the latency reduction does not depend on the fidelity definition, and the wireless part (BER vs. power) is standard. So the paper is not a waste of time, but as written it fails the reproducibility bar for the optimization claims.\n\nWho is this for? Researchers working on wireless LLM serving who want a pointer at a plausible joint design; they should not take the fidelity/BER curves as evidence until the objective is defined. I would not cite it in its current form. Do I want to see it in peer review? Yes, if the authors are pushed to write down the reward and the fidelity metric, add baselines, and correct the DQN update. As is, a serious venue should ask for major revision; a lenient conference might let it through, which would be a disservice.\n\nRecommendation: send it back with a request for the missing formulas, or desk-reject if the venue doesn't allow major revision.","headline":"A plausible joint optimization of prompt compression and wireless power, but the objective and reward are never defined, making the claimed fidelity results unreproducible.","tokens_in":8937,"tokens_out":2989,"would_cite":false,"duration_ms":26023,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Jointly choosing prompt compression and transmit power cuts LLM response time by about 17% in wireless services.","keywords":["prompt compression","SLM","power allocation","Double DQN","wireless LLM services","semantic fidelity","latency reduction","edge intelligence"],"falsifier":"Run the JPPO policy with a concrete fidelity definition, such as cosine similarity of sentence embedding vectors for $f_1$ and $f_3$ and token overlap for $f_2$, then compare the policy's chosen compression and power actions against the actual downstream task accuracy (for example, exact match on the MeetingBank QA answers). If the reported ~0.9 fidelity coexists with a statistically significant drop in task accuracy, the central claim that JPPO preserves service quality is falsified.","tokens_in":7945,"feed_emoji":"⚡","tokens_out":6235,"duration_ms":51581,"temperature":0.7,"pith_summary":"The paper proposes JPPO, a framework that uses a small language model on the user device to compress long prompts and a Double DQN agent to jointly choose the compression ratio and the wireless transmit power. The claimed result is that this joint control keeps service fidelity near 0.9, bit error rate below 0.2, and power around 4 to 5 watts while cutting end-to-end response time by roughly 17% compared with transmitting the full prompt at fixed power. The authors argue this turns the long-prompt bottleneck in wireless LLM services into an adjustable trade-off that can be balanced automatically as channel conditions and prompt content change. A sympathetic reader would care because wireless LLM deployment is currently limited by communication load and device energy, and this is a concrete proposal that addresses both at once.","feed_headline":"Joint prompt-power control cuts wireless LLM response time 17 percent","feed_subtitle":"A Double DQN pairs on-device SLM compression with transmit power to keep fidelity near 0.9 and BER below 0.2.","key_machinery":"The argument is carried by the fidelity metric $f(\\kappa,\\eta(P_T)) = \\alpha_1 f_1 + \\alpha_2 f_2 + \\alpha_3 f_3$, where $f_1$ measures representation similarity between original and compressed prompts, $f_2$ measures token retention after transmission at a given BER $\\eta$, and $f_3$ measures response similarity at the target LLM. This metric turns two very different resources—semantic information in the prompt and radio power—into a single objective, which the Double DQN (with separate current and target Q-networks) optimizes over a discrete action space of compression levels and power levels. The SLM-based compressor (GPT-Neo 125M via LLMLingua) is the tool that makes $\\kappa$ controllable at the user side, and the target LLM (GPT-J 6B) is what consumes the compressed prompt.","core_discovery":"The central claim is that optimizing the compression ratio $\\kappa$ and transmit power $P_T$ together, through a Double DQN whose state is the instantaneous fidelity, SNR, and BER, achieves a better operating point than either compression alone or power control alone. Concretely, the paper reports that on MeetingBank-transcript prompts, the learned policy maintains fidelity around $0.9$, BER below $0.2$, power in the range $4$–$5$ W, and reduces first-token response time from $56.1$ to $46.9$ seconds for a 44-token prompt and from $85.3$ to $71.2$ seconds for a 388-token prompt, averaging about $17\\%$ across the test set. The discovery is that a single reinforcement-learned controller can track the fidelity–power trade-off online without retraining the SLM compressor.","pith_inferences":["If the fidelity proxy $f$ actually tracks task accuracy, the same DRL structure could absorb extra decisions such as channel coding rate, number of retransmissions, or whether to run part of the prompt locally.","The 17% figure depends on the unstated similarity metrics for $f_1$ and $f_3$; a concrete definition would let other groups reproduce the result and test whether the policy is truly optimizing answer quality.","The framework treats $\\kappa$ and $P_T$ as independent discrete levels; one could extend it to continuous actions with a policy-gradient method and see if the same ~17% gain persists with finer granularity.","A natural stress test is to run JPPO on a different dataset (for example, code or math prompts) where the SLM's compression may hurt task accuracy more, and check whether the DRL learns to back off compression."],"forward_implications":["If the reported numbers hold, wireless LLM services can run on device-class SLMs with roughly 17% lower latency and no drop in measured fidelity, which directly addresses the long-prompt communication bottleneck.","The JPPO policy generalizes across prompt lengths within the test set, so the controller does not need retraining for each new prompt length.","The framework's power band of 4–5 W suggests it is compatible with battery-constrained mobile devices while satisfying an energy budget $E_{th}$ and latency bound $T_{th}$.","Because the action space is discrete and the state space is low-dimensional, the Double DQN converges within 10,000 episodes, making online deployment in a base station or data center operator plausible.","Constraint (10d) enforces a minimum fidelity floor, so the policy is forced to sacrifice compression or power to preserve service quality when the channel degrades."],"supporting_citations":[{"why":"Supplies the SLM-based prompt compression method (LLMLingua) that JPPO adopts for the coarse-to-fine compression.","marker":"[10]"},{"why":"Provides the fidelity formulation used to build $f_1$, $f_2$, and $f_3$ for the semantic communication quality metric.","marker":"[13]"},{"why":"Gives the Double DQN algorithm that the paper uses to solve the joint optimization problem.","marker":"[15]"},{"why":"Supplies the GPU energy consumption model for SLM and LLM encoding used in the energy constraint.","marker":"[14]"},{"why":"Provides the GPT-Neo 125M and GPT-J 6B models and the Transformers platform used in the simulation experiments.","marker":"[16]"}],"fun_headline_variants":["RL pairs prompt compression and power to slash LLM latency 17%","Double DQN tunes compression and transmit power for wireless LLMs","Joint optimization of prompt length and power speeds LLM responses","On-device SLM plus RL power control cuts LLM response time 17%","Fidelity-aware RL balances prompt compression and radio power"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the paper's 'fidelity' measure, assembled from three components that are never written down, genuinely tracks whether the LLM's final answer to the user is still correct; if the measure is wrong, the controller is optimizing a number that has nothing to do with service quality.","fun_headline_variants_meta":{"raw":{"variants":["RL pairs prompt compression and power to slash LLM latency 17%","Double DQN tunes compression and transmit power for wireless LLMs","Joint optimization of prompt length and power speeds LLM responses","On-device SLM plus RL power control cuts LLM response time 17%","Fidelity-aware RL balances prompt compression and radio power"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00071,"raw_usage":{"total_tokens":3165,"prompt_tokens":882,"completion_tokens":2283,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":2192}},"tokens_in":498,"tokens_out":2283,"duration_ms":15169,"temperature":1.0,"reasoning_tokens":2192,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:36:20.497796+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the JPPO policy with a concrete fidelity definition, such as cosine similarity of sentence embedding vectors for $f_1$ and $f_3$ and token overlap for $f_2$, then compare the policy's chosen compression and power actions against the actual downstream task accuracy (for example, exact match on the MeetingBank QA answers). If the reported ~0.9 fidelity coexists with a statistically significant drop in task accuracy, the central claim that JPPO preserves service quality is falsified.","supporting_citations":[{"cited_title":"LLMCarbon: Modeling the end-to-end carbon footprint of large language models,","cited_arxiv_id":null,"evidence_quote":"Supplies the GPU energy consumption model for SLM and LLM encoding used in the energy constraint."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the GPT-Neo 125M and GPT-J 6B models and the Transformers platform used in the simulation experiments."}],"review_version":1}