{"id":"188fbbf4-c292-44f1-8bc4-7cbcedd22267","arxiv_id":"2507.07422","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"For task-oriented image classification over wireless links, a static encoder compresses features to 16 transmitted dimensions, while a multi-exit dynamic encoder routes easy inputs to early exits to meet per-device FLOPs budgets.","lead":"This paper designs wireless links that send only the information needed for a machine-learning task, with a version that saves on-device computing by classifying easy inputs early. It matters because phones, drones, and sensors cannot run heavy neural nets, and this is a concrete attempt to fit task-oriented communication to such devices.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The dynamic model's threshold calibration in Algorithm 2 does not realize the exit distribution assumed in Eq. (17), so realized FLOPs can depart from the declared budget B and the accuracy-FLOPs curves are unvalidated.","rationale":"The reader's weakest assumption identified the same load-bearing region: the budget-calibration mechanism of Section V and the possibility that realized exit behavior departs from the calibrated probabilities. My stress-test sharpens this into a concrete internal inconsistency: Algorithm 2 sets thresholds from marginal per-exit confidence-score quantiles, while Algorithm 3 performs sequential early exiting, so the actual exit distribution is the joint distribution of ordered confidence scores, not the marginal distribution. Therefore the expected-FLOPs formula in Eq. (18) does not describe the deployed rule, and Proposition 2's bound applies to a different decision procedure. This is the most load-bearing concern because the dynamic model's central promise, operating within a specified computation budget while improving accuracy, is exactly what this calibration step is supposed to guarantee. If realized FLOPs deviate from B, the accuracy-FLOPs curves in Figs. 6-9 may show accuracy at a nominal budget that is not the actual compute spent. This is fixable by measuring FLOPs and re-calibrating sequentially, but the paper currently reports no measured FLOPs and no ablation showing that realized exit frequencies match P r_k. I do not find a comparably serious flaw in the static model claims: the dimension reduction and FLOPs numbers are internally consistent, and the accuracy comparisons, while not universally favorable in Rayleigh fading at low PSNR, are reported per condition rather than as a single overclaim. The convergence analysis is standard nonconvex SGD and not novel, but it is not load-bearing for the efficiency claims. The reader's conditional verdict remains appropriate: the central efficiency claims are plausible but need the budget-accounting loop closed with measured FLOPs and ideally released code before the dynamic model results can be relied upon.","tokens_in":1343,"tokens_out":1022,"duration_ms":63808,"concrete_test":"For the CIFAR-100 AWGN PSNR=0 dB case, recompute thresholds theta_k via Algorithm 2 on Sval for each budget B in Table IV, then apply Algorithm 3 to Sval and to a held-out test split, recording the empirical exit distribution p_hat_k and realized average FLOPs R_hat = sum_{k=1}^{K} k phi p_hat_k. If for any B the deviation between R_hat and B exceeds 5%, or if p_hat_k differs materially from the intended P r_k of Eq. (17), the calibration loop does not realize the declared budget, and the accuracy-FLOPs curves should be re-plotted against measured FLOPs. Running this on Sval isolates the calibration mismatch from the test-distribution-shift issue.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the dynamic TOC model meets a device computation budget B while improving accuracy rests on two pieces: the exit probabilities P r_k in Eq. (17) and Proposition 2, which bounds expected FLOPs as R = sum k phi P r_k, and the calibration step in Algorithm 2 that converts P r_k into thresholds theta_k. The problem is that Algorithm 3 uses a sequential greedy rule: a sample exits at exit k only if it fails all earlier thresholds and then passes theta_k. Algorithm 2, however, sets each theta_k (by sorting validation confidence scores) to match a target fraction of validation samples whose confidence score at exit k alone exceeds theta_k, i.e., a marginal passing rate. The actual exit rate at exit k is the probability that the sample fails exits 1 through k-1 and passes exit k, which depends on correlations among confidence scores across exits. Except in special cases (e.g., perfectly monotone confidence scores), this realized exit distribution will differ from the P r_k used in Eq. (18) and Proposition 2. Consequently, the realized average FLOPs need not equal R and can exceed B even when Proposition 2 holds for the nominal distribution. The accuracy-versus-budget curves in Figs. 6-9 plot nominal B on the x-axis, not measured FLOPs, and no measured FLOPs are reported to close this loop. This gap directly affects the paper's headline contribution: the dynamic model's 'computational demand' is not verified to be the declared B. The issue is internal to the proposed method, not merely a distribution-shift concern: it should be visible on the validation set itself.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes computation-resource-efficient task-oriented communication (TOC) systems with static and dynamic operation modes. The static mode uses a ResNet as the task-oriented encoder and a lightweight channel encoder/decoder to reduce transmitted feature dimensions (e.g., 16 dimensions on CIFAR) and FLOPs, while the dynamic mode uses a multi-exit dynamic neural network with confidence-score thresholds calibrated on a validation set to allocate computation within a device budget B. The authors provide a convergence analysis for SGD training and evaluate on CIFAR-10/100 and ImageNet, reporting accuracy gains over DeepJSCC, VFE, and GatedJSCC baselines. The central claims are that the static model reduces transmitted dimensions, FLOPs, and accuracy loss simultaneously, and that the dynamic model meets device computation budgets while further improving accuracy as the budget grows.","tokens_in":21338,"tokens_out":8143,"duration_ms":82730,"significance":"The paper addresses a timely and practically important problem: making task-oriented communication feasible on resource-constrained devices. The static model's reductions in transmitted feature dimension (to 16 on CIFAR) and FLOPs are clearly demonstrated and are valuable contributions. The dynamic model's idea of early exiting based on task complexity is natural for TOC and could be impactful if the budget mechanism is sound. The convergence analysis is standard nonconvex SGD theory and is correctly applied, though it does not use TOC-specific structure. The main weaknesses are a gap between the threshold calibration and the realized exit distribution in the dynamic model, an inconsistency in the expected-FLOPs formula and its proof, and an overstated claim of simultaneous accuracy improvement. These issues are fixable, but they affect the load-bearing budget-feasibility claim and the accuracy comparison.","major_comments":[{"comment":"The threshold calibration in Algorithm 2 sets each θk to match the marginal per-exit probability Prk, but Algorithm 3 uses a sequential greedy rule: a sample exits at k only if it fails all earlier thresholds and then passes θk. The realized exit probability at exit k is therefore the joint probability P(φ1 < θ1, ..., φk−1 < θk−1, φk ≥ θk), which equals the marginal Prk only under strong assumptions about correlation among confidence scores across exits. Consequently, the expected-FLOPs formula in Eq. (18) and the budget bound in Proposition 2 hold for the nominal distribution, not for the distribution actually realized by Algorithm 3, and the realized computation cost can exceed B even when the nominal bound is satisfied. Since Figs. 6–9 plot the nominal budget B on the x-axis and the paper reports no measured FLOPs for the dynamic model on the test set, the central claim that the dynamic model meets a device computation budget is not verified. Please either revise the threshold calibration to be sequential (e.g., set θk on the subset of validation samples that failed all earlier exits) or report measured average FLOPs on the test set alongside the nominal budget curves.","section":"Section V-C-2, Algorithms 2 and 3, Eq. (17)-(18), Proposition 2"},{"comment":"The denominator in Eq. (18) is printed as ∑_{j=1}^k r^j, but the derivation in Appendix B (Eq. (39) and the following simplification) replaces it with ∑_{j=1}^K r^j, which equals r(1−r^K)/(1−r) and is independent of k. This substitution is unjustified as written: for k < K, ∑_{j=1}^k r^j ≠ r(1−r^K)/(1−r). The final closed form and Lemmas 1–2 correspond to the denominator ∑_{j=1}^K r^j, which is consistent with the definition of Prk in Eq. (17). The printed Eq. (18) should be corrected to the total sum, and the proof in Appendix B should be rewritten to avoid the invalid substitution.","section":"Eq. (18) and Appendix B"},{"comment":"The abstract and Section VII.B claim that the proposed static model outperforms baseline models in transmitted dimensions, FLOPs, and accuracy simultaneously, but Table III shows counterexamples. In Rayleigh fading at PSNR = 0 dB, the proposed static ResNet-20 achieves 34.98% accuracy and ResNet-32 achieves 34.57%, both below DeepJSCC (36.20%) and well below GatedJSCC (53.28%). The claim should be qualified to the operating points where the accuracy advantage actually holds (e.g., higher PSNR levels or AWGN channels), or the comparison should be aggregated in a way that acknowledges these exceptions.","section":"Section VII-B, Table III"}],"minor_comments":[{"comment":"The pseudocode in Algorithm 2 places the 'Sort ϕj,k in descending order' step inside the per-sample loop, but threshold setting requires sorting the confidence scores for each exit across all validation samples. Please clarify the loop structure so that the sorting and threshold selection are performed per exit over the entire validation set.","section":"Algorithm 2"},{"comment":"The text accompanying Table IV states that 'neither the baseline models nor the proposed static TOC model show an increase in accuracy' as the budget grows, but the table shows the static model accuracy increasing from 55.58% (ResNet-20 at 50M FLOPs) to 64.94% (ResNet-32 at 70M FLOPs) when a larger network becomes affordable. The intended point is that accuracy is constant for a fixed architecture once its FLOPs requirement is met; please reword to avoid the contradiction.","section":"Section VII-C-2, Table IV"},{"comment":"All results are reported as averages over 5 trials without standard deviations or confidence intervals. Given the non-monotonic accuracy across ResNet depths, including error bars or variance information would strengthen the comparisons.","section":"Section VII-A-3"},{"comment":"The convergence analysis is a standard nonconvex SGD bound that does not use any TOC-specific structure. The authors should state explicitly that the channel and task modules are subsumed in the objective function, so the analysis applies to the whole end-to-end system, not just the encoder.","section":"Section VI"},{"comment":"The PSNR definition PSNR = 10 log10(P/σ²) uses signal power P and noise power σ²; please clarify whether P denotes peak signal power or average power, and specify the normalization used in the simulations.","section":"Eq. (29)"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely problem and the static-model results are solid and clearly presented. The main risk is the unverified budget guarantee for the dynamic model: the calibration-vs-testing gap and the missing measured FLOPs should be addressed before publication. The equation error in the expected-FLOPs derivation is also a correctness issue in the proof. These items are fixable, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you work on task-oriented communication or early-exit inference. The genuinely new piece is applying multi-exit dynamic networks to TOC, with a validation-calibrated budget interface: you set a FLOPs budget B and get exit thresholds that are supposed to make expected compute track B. The static model is also a straightforward but effective combination: ResNet encoder plus a small linear channel codec, transmitting 16-dimensional features on CIFAR-100 with 40.81M FLOPs for ResNet-20 and accuracy above DeepJSCC, VFE, and GatedJSCC at most SNRs.\n\nThe dynamic results are plausible and the accuracy-versus-budget curves on three datasets suggest the idea has legs. The paper is clearly written, baselines are appropriate, and the complexity analysis in the appendix is properly documented. The convergence section is standard nonconvex SGD theory and shouldn't be counted as a contribution, but it is correctly handled.\n\nThe soft spot is load-bearing. The budget mechanism does not actually guarantee the advertised FLOPs. Algorithm 2 sets each θ_k by matching a marginal pass rate at exit k on validation data. Algorithm 3 then exits greedily: a sample passes exit k only after failing all earlier exits. The realized exit distribution depends on correlations among confidence scores across exits, so it can differ substantially from the P r_k used in Eq. (18) and Proposition 2. That means realized FLOPs can exceed B even when the nominal bound holds. No measured FLOPs are reported anywhere, and the x-axes in Figs. 6–9 are the declared budget B, not measured compute. This is not a distribution-shift quibble; it should be visible on the validation set itself. A reasonable fix is to calibrate thresholds jointly to the sequential rule, or report measured FLOPs and adjust r accordingly.\n\nMinor issues: no error bars despite averaging over five trials, no code released, and the Rayleigh fading result at 0 dB for ResNet-20 (34.98%) falls below DeepJSCC (36.20%). The similar-distribution assumption is stated but not stress-tested.\n\nWho this is for: researchers in semantic/TOC and resource-efficient edge inference. It deserves peer review, but the authors should be asked to close the budget-accounting loop before it is relied on. I'd send it to a serious venue with major-revision expectations.","headline":"Static model is solid; dynamic budget claim needs a fix—Algorithm 2 calibrates thresholds marginally but the greedy exit rule changes realized FLOPs.","tokens_in":21953,"tokens_out":3878,"would_cite":true,"duration_ms":43248,"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":"Task-oriented communication can be made computation-efficient by a multi-exit encoder that spends more compute only on hard images, while a static variant already transmits 16 features and beats stronger baselines on accuracy, transmitted…","keywords":["task-oriented communications","semantic communications","dynamic neural networks","multi-exit networks","computation budget","image classification","wireless transmission","FLOPs"],"falsifier":"Run the trained dynamic model on a test set deliberately shifted from validation, e.g., blurred images or a different class distribution, while keeping the validation-calibrated thresholds fixed, and measure the realized average FLOPs per batch; if it exceeds the budget $B$, the budget-guarantee mechanism fails. Also count FLOPs between consecutive exits directly; if they are not roughly equal, the $[\\varphi, K\\varphi]$ range used in Proposition 2 is not accurate.","tokens_in":20792,"feed_emoji":"📡","tokens_out":7402,"duration_ms":72978,"temperature":0.7,"pith_summary":"Task-oriented communication (TOC) normally spends heavy compute to extract and transmit only task-relevant features, which is hard for phones, UAVs, and sensors. This paper proposes two TOC modes: a static encoder for devices with no computation cap, and a dynamic multi-exit encoder for devices with a computation budget. The static mode compresses transmitted features to 16 dimensions and cuts FLOPs while improving classification accuracy relative to DeepJSCC, VFE, and GatedJSCC on image tasks. The dynamic mode estimates each image's difficulty from a confidence score and exits early for easy images, so average compute stays within a budget while accuracy rises as the budget grows. The paper also proves an $O(1/\\sqrt{T})$ convergence rate for SGD training of the end-to-end system.","feed_headline":"99% fewer transmitted features and higher accuracy","feed_subtitle":"A static encoder sends 16 features with fewer FLOPs; a dynamic variant adapts compute to image difficulty under budget.","key_machinery":"The load-bearing mechanism is a multi-exit dynamic neural network used as the task-oriented encoder. Each input image gets a confidence score at each exit, defined as the softmax value of the predicted-class logit; easy images pass a validation-calibrated threshold at an early exit, while hard images proceed to deeper exits. The network is structured with horizontal depth and vertical feature scales, dense connectivity to stabilize propagation, and exits assumed evenly spaced with $\\varphi$ FLOPs between consecutive exits. A simplified channel encoder/decoder with linear layers compresses features to 16 dimensions before wireless transmission, and exit thresholds are chosen on a validation set so that expected FLOPs $\\sum_k \\Pr_k C_k$ respect the device budget $B$.","core_discovery":"The paper's central claim is that computation efficiency and task accuracy do not have to be traded off in TOC. With ResNet-20 as the static task-oriented encoder on CIFAR-100, the proposed system transmits 16-dimensional features, uses about 40.81M FLOPs, and reaches 55.58% accuracy at PSNR 0 dB in AWGN, against DeepJSCC's 2048 dimensions, 50.25M FLOPs, and 39.22% accuracy; VFE uses more FLOPs and GatedJSCC uses 256 dimensions. The dynamic model operates under a device budget $B$ and, at $B \\ge 60$M FLOPs, reaches 62.26% accuracy, exceeding both static baselines, because simple images exit early and complex images go deeper. Accuracy generally grows with budget up to a point, after which deeper exits can hurt.","pith_inferences":["Editorial inference: the paper's Section V assumption that training, validation, and test sets share a similar distribution is what makes Algorithm 2's validation-calibrated thresholds enforce the budget at test time; under distribution shift the realized FLOPs could exceed $B$, so deployed systems should monitor exit-rate statistics.","Editorial inference: the same multi-exit encoder could double as a channel-adaptive transmission mechanism, e.g., skipping late exits under poor SNR, but the paper does not test this.","Editorial inference: the even-exit-spacing assumption could be relaxed by measuring per-exit FLOPs directly, and a version with uneven exits would need a modified budget formula."],"forward_implications":["If the static claim is right, task-oriented systems can transmit 16 features instead of 2048 with higher accuracy and fewer FLOPs on small-image classification.","The dynamic model gives a single trained network that can serve devices with different computation budgets by only changing exit thresholds.","Accuracy-budget curves improve with PSNR, meaning the same computation budget buys higher accuracy when the wireless channel is cleaner.","The convergence proof implies that end-to-end training of the encoder, channel codec, and inference block stabilizes after enough SGD iterations."],"supporting_citations":[{"why":"Supplies the DeepJSCC baseline that motivates the comparison; its 2048 transmitted dimensions anchor the 99.22% reduction claim.","marker":"[3]"},{"why":"Supplies the VFE baseline with a variational information bottleneck; its 64-dim transmission and FLOP count are the main efficiency comparisons.","marker":"[13]"},{"why":"Supplies the GatedJSCC baseline and its 256-dim transmission, used for accuracy and FLOP comparisons.","marker":"[10]"},{"why":"Supplies the ResNet family used as the static task-oriented encoder and the convolution-only FLOP simplification convention.","marker":"[9]"},{"why":"Supplies the multi-scale dense network design that underlies the dynamic TOE's horizontal/vertical structure.","marker":"[8]"},{"why":"Supplies dense connectivity for the dynamic encoder and the empirically chosen equal exit-loss weights.","marker":"[28]"},{"why":"Supplies the ImageNet dataset used to validate large-scale dynamic model performance.","marker":"[7]"},{"why":"Supplies the CIFAR-10 and CIFAR-100 datasets used in the main experiments.","marker":"[32]"},{"why":"Supplies the stochastic-gradient smoothness and bounded-noise assumptions and the nonconvex SGD convergence template used in Theorem 1.","marker":"[30]"},{"why":"Supplies the Lipschitz-continuity and bounded-variance assumptions underlying the $O(1/\\sqrt{T})$ convergence proof.","marker":"[31]"}],"fun_headline_variants":["Task-oriented comms cut features 99% and lift accuracy","Dynamic encoders adapt compute to image complexity","Resource-constrained TOC: static and dynamic encoders beat baselines","62% accuracy with 60M FLOPs using dynamic exits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise, stated in Section V, is that training, validation, and test data share a similar distribution, so confidence thresholds calibrated on the validation set keep the test-time average computation spend within the device budget.","fun_headline_variants_meta":{"raw":{"variants":["Task-oriented comms cut features 99% and lift accuracy","Dynamic encoders adapt compute to image complexity","Resource-constrained TOC: static and dynamic encoders beat baselines","62% accuracy with 60M FLOPs using dynamic exits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000692,"raw_usage":{"total_tokens":3129,"prompt_tokens":938,"completion_tokens":2191,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":2121}},"tokens_in":554,"tokens_out":2191,"duration_ms":18551,"temperature":1.0,"reasoning_tokens":2121,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:42:09.214834+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained dynamic model on a test set deliberately shifted from validation, e.g., blurred images or a different class distribution, while keeping the validation-calibrated thresholds fixed, and measure the realized average FLOPs per batch; if it exceeds the budget $B$, the budget-guarantee mechanism fails. Also count FLOPs between consecutive exits directly; if they are not roughly equal, the $[\\varphi, K\\varphi]$ range used in Proposition 2 is not accurate.","supporting_citations":[{"cited_title":"Deep joint source-channel coding for wireless image transmission,","cited_arxiv_id":null,"evidence_quote":"Supplies the DeepJSCC baseline that motivates the comparison; its 2048 transmitted dimensions anchor the 99.22% reduction claim."},{"cited_title":"Learning task-oriented communication for edge inference: An information bottleneck approach,","cited_arxiv_id":null,"evidence_quote":"Supplies the VFE baseline with a variational information bottleneck; its 64-dim transmission and FLOP count are the main efficiency comparisons."},{"cited_title":"Semantic commu- nications for image recovery and classification via deep joint source and channel coding,","cited_arxiv_id":null,"evidence_quote":"Supplies the GatedJSCC baseline and its 256-dim transmission, used for accuracy and FLOP comparisons."},{"cited_title":"Densely connected convolutional networks,","cited_arxiv_id":null,"evidence_quote":"Supplies dense connectivity for the dynamic encoder and the empirically chosen equal exit-loss weights."},{"cited_title":"Imagenet: A large-scale hierarchical image database,","cited_arxiv_id":null,"evidence_quote":"Supplies the ImageNet dataset used to validate large-scale dynamic model performance."},{"cited_title":"Krizhevsky et al., Learning multiple layers of features from tiny images, 2009","cited_arxiv_id":null,"evidence_quote":"Supplies the CIFAR-10 and CIFAR-100 datasets used in the main experiments."},{"cited_title":"Optimization methods for large-scale machine learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the stochastic-gradient smoothness and bounded-noise assumptions and the nonconvex SGD convergence template used in Theorem 1."},{"cited_title":"Stochastic first-and zeroth-order methods for nonconvex stochastic programming,","cited_arxiv_id":null,"evidence_quote":"Supplies the Lipschitz-continuity and bounded-variance assumptions underlying the $O(1/\\sqrt{T})$ convergence proof."}],"review_version":1}