{"id":"e1b7941c-fe0c-43ca-9b83-891fabf22c43","arxiv_id":"2505.10834","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"TACO transmits a downsampled VQ-VAE latent plus GradCAM-selected task tokens and reconstructs via latent replacement, with a ground-truth-selected feedback percentage.","lead":"TACO is an image communication system that sends a small context image plus only the parts of the image a downstream task cares about, then refines the receiver's reconstruction on request. If its reported numbers hold, it would offer large bandwidth and latency savings over diffusion-based semantic communications, but the evaluation uses test labels to choose what to send.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"TACO-LSF's headline gains rely on using ground-truth labels at the transmitter to choose the transmitted token percentage per image; without that oracle the 96.49% at 0.65 KB claim is unsupported and may not hold.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the LSF procedure at the transmitter uses ground-truth labels to choose the task-token percentage, making the reported TACO-LSF gains an upper bound over a per-image search rather than a deployable policy. My reading of Section III.A.4 confirms this: the transmitter literally evaluates downstream task performance against ground truth for each candidate percentage before transmitting. This is a test-time label leak that invalidates the central claim of simultaneous superiority in task accuracy, bandwidth, and latency. The concrete test I propose would settle the concern by comparing a fixed-policy TACO against the original image baseline. I therefore agree with the REJECT verdict and would not change it. I also note an additional internal inconsistency (Table I reports FID 11.21 for TACO while Table II reports FID 34.30 for TACO on the same Cityscapes reconstruction task), which further weakens the reconstruction-quality claims, but the label leakage in LSF is the more fundamental issue because it directly undermines the strongest claim of beating the original image.","tokens_in":9057,"tokens_out":6391,"duration_ms":62427,"concrete_test":"Retrain or re-evaluate TACO with LSF disabled or with p fixed to a single value chosen on the training/validation set (e.g., p=30). Then evaluate classification accuracy on a held-out STL-10 test set without any per-image ground-truth label search, and report the resulting accuracy at the corresponding average bandwidth. If the accuracy at a fixed p is not above 94.30%, the reported TACO-LSF superiority over the original image is an artifact of oracle selection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that TACO-LSF achieves 96.49% classification accuracy at 0.65 KB, surpassing the original image's 94.30%, is not supported as a deployable system. Section III.A.4 describes the LSF procedure: for each image, the transmitter tries percentages from {10,20,30,50,70,90,100}, evaluates the downstream task performance of the reconstructed output against ground-truth labels, and selects the percentage whose performance is 'compatible with the ground truth.' This is a per-image oracle search over labels. In a real semantic communication link, the transmitter (e.g., a camera or sensor) does not know the ground-truth class or detection labels; it only has the image and the downstream model. Thus the reported TACO-LSF accuracy and bandwidth are best-case selections, not the behavior of a fixed policy. The bandwidth figure of 0.65 KB is also selected per image over the search grid, so it does not represent an achievable fixed-rate transmission. Without label access, the transmitter would have to fix p a priori, and the claimed advantage over the original image and over JPEG baselines is not established. The no-LSF results (e.g., 85.51% at 1.51 KB for ζ+10%) are below the original image's 94.30%, so the claim that TACO simultaneously maximizes task accuracy, bandwidth efficiency, and reconstruction quality depends critically on the label-leaking LSF selection.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript introduces TACO, a semantic image communication framework built on a pretrained VQ-VAE. The transmitter sends a downsampled context latent and a set of task-salient latent tokens selected by GradCAM, and the receiver fuses the two. The paper also proposes 'local semantic feedback' (LSF), where the transmitter tries several task-token percentages and chooses the one that yields the best downstream accuracy. Evaluation on Cityscapes, STL-10, and Flickr reports large improvements over diffusion-based semantic communication and JPEG/JPEG2000 baselines, including 96.49% classification accuracy at 0.65 KB, higher than the 94.30% accuracy of the original image at 59.93 KB, and reconstruction time of 0.0015 s.","tokens_in":9313,"tokens_out":9516,"duration_ms":98231,"significance":"If the experimental claims were valid, TACO would be a useful contribution: the context-plus-task-token decomposition is simple, the reconstruction speed gain is large, and the paper demonstrates adaptation to a new task via receiver feedback. No code or repeated-seed statistics are provided, and the core performance claim depends on an oracle-like use of ground-truth labels at the transmitter during LSF. The paper is therefore better read as a promising architecture with an unsupported headline evaluation.","major_comments":[{"comment":"The LSF protocol selects the transmitted token percentage per image by evaluating downstream accuracy against ground-truth labels. The text says the transmitter 'evaluate[s] the downstream task performance' and keeps the percentage whose performance 'is compatible with the ground truth.' In a real system the transmitter (e.g., a camera) does not have the class or detection ground truth, so the reported TACO-LSF accuracy of 96.49% at 0.65 KB is an oracle-selected upper bound, not the result of a fixed transmission policy. The bandwidth is also per-image selected, so it is not an achievable fixed rate. The no-LSF rows in Table III (e.g., ζ+10% gives 85.51% at 1.51 KB) fall below the original image's 94.30%, so the central claim that TACO beats the original image depends entirely on this label-leaking selection. The authors should either design a selection rule that uses only transmitter-side information (e.g., downstream confidence on the reconstructed context) and report that policy's performance, or clearly present LSF as an oracle upper bound and re-derive the system-level claims.","section":"Sec. III.A.4 and Tables III-IV"},{"comment":"Tables III-V report single-run accuracy, mAP, and mIoU values without error bars or repeated seeds. Because LSF performs a per-image search over the test set, a single split can overfit to the ground-truth labels used in selection. At minimum the tables need means and standard deviations over at least three independent runs or bootstrap intervals, and the conclusions should be drawn from the resulting intervals.","section":"Tables III-V"},{"comment":"The paper contains no channel model or channel coding: token indices are assumed to arrive error-free, and the 'Bandwidth' columns do not account for coding overhead, the cost of transmitting the mask M in Eq. (7), or bit errors. Section V explicitly lists channel noise as future work. As written, the contribution is a noiseless source-coding system, and claims about 'semantic communications' and 'ultra-high bandwidth efficiency' should be narrowed accordingly, or the experiments should include a noisy-token channel with error correction.","section":"Sec. V and Tables II-IV"},{"comment":"Tables III and IV compare TACO only against JPEG and JPEG2000 for classification and object detection; no learned image codec or existing deep semantic-communication baseline appears in these tables, so the abstract's claim of outperforming 'existing work' is not supported on these tasks. The comparisons in Table I are also uneven: TACO is a single feed-forward pass, yet its 'Steps' entry is listed as 'NA,' while the text claims it outperforms diffusion models in number of steps; the number of inference steps should be stated explicitly.","section":"Tables III and IV and Table I"}],"minor_comments":[{"comment":"The candidate percentage sets are inconsistent: Sec. III.A.3 lists {10,20,30,50,70}, Sec. III.A.4 lists {10,20,30,50,70,90,100}, and the Figure 2 caption says [10,20,50,100]; please unify them.","section":"Sec. III.A.3, Sec. III.A.4, Fig. 2"},{"comment":"'DATAEST' is a typo for 'DATASET' in the table title.","section":"Table III"},{"comment":"The phrase 'ˆx is unnecessarily to be in the format of original data' should read 'ˆx does not need to be in the format of the original data.'","section":"Abstract and Sec. II.B"},{"comment":"The row 'Diff-GO+(n=1024,L=1024) ... +GO-EV AE(n=1024,L=16)' appears to be a formatting error; please correct the baseline name and parameters.","section":"Table II"},{"comment":"The 'Steps' entry for TACO is listed as '-'; since the paper's latency argument depends on TACO being single-step, report the exact number of forward passes.","section":"Table VI"},{"comment":"The paper does not provide a code or data release statement; adding one would improve reproducibility.","section":"Throughout"},{"comment":"The object-detection experiments use a dataset referred to only as 'Flickr,' but the cited reference [19] is a sentence-to-image retrieval dataset; please specify the exact dataset, annotations, and detection training protocol.","section":"Sec. IV.A.3"}],"recommendation":"reject","confidential_remarks":"For the editor: the paper's headline result rests on an oracle-based selection in Sec. III.A.4; even if the authors reframe LSF as an upper bound, the central claim of outperforming the original image would need to be re-established under a fixed-rate policy. The paper also has no channel model, which is a scope concern for a communications journal. I recommend rejection rather than major revision because the invalid selection protocol is the paper's main contribution, not a peripheral experimental detail."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, you should know two things about TACO before reading: (1) it contains a genuinely new lightweight latent-level semantic communication architecture, and (2) its headline results are undermined by a label-leaking selection procedure that the authors describe openly. The transmitter, in Sec. III.A.4, tries task-token percentages {10,20,30,50,70,90,100}, runs the full receiver pipeline for each, checks downstream accuracy against ground-truth labels, and keeps the percentage that scores best. That is per-image oracle search, not a policy. The widely quoted result — TACO-LSF 96.49% accuracy at 0.65 KB vs 94.30% for the original 59.93 KB image — is the optimum of that search. Without the oracle, the best fixed-percentage result is 85.51% at 1.51 KB, below the original image. The bandwidth number is also per-image selected, not a fixed rate. So the central claim that TACO simultaneously improves accuracy, bandwidth, and latency does not survive.\n\nWhat is genuinely new: the combination of a downsampled context latent from the same VQ-VAE encoder, GradCAM-selected task tokens, and latent replacement at the receiver is not in the prior work I know. The no-LSF ablations show a sensible monotone tradeoff — more task tokens, more accuracy, more bits — and the reconstruction quality claim on Cityscapes (LPIPS 0.056, FID 11.21, 0.0015s on an A100) is striking, though the comparison against SPADE/OASIS/SMIS is apples-to-oranges because those take semantic maps, not images.\n\nSoft spots beyond the oracle: no error bars or repeated seeds; no channel coding or noise model despite the communication framing; a numeric inconsistency in the paper's own tables — FID 11.21 in Table I and Table VI vs FID 34.30 for TACO in Table II; and \"Steps\" is NA for TACO, so claiming superiority in number of regeneration steps is rhetorical. No code or data are provided, so the numbers cannot be independently checked.\n\nThis paper is for the semantic-communication community, and the architectural idea deserves a serious look. But as submitted, the empirical case for TACO-LSF is not supported. I would send it to peer review in the hope that the authors can remove the oracle, fix the rate, add a channel model, and report repeated runs. A fixed-policy version would be a solid contribution; the current version overclaims.","headline":"A genuinely new lightweight VQ-based semantic communication architecture, but its headline LSF results are invalid because the transmitter uses ground-truth labels to select per-image token percentages.","tokens_in":9889,"tokens_out":3092,"would_cite":false,"duration_ms":30609,"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 splitting an image into a compact context latent and a small set of task-salient tokens lets a single lightweight VQ-VAE beat diffusion-based semantic communication on reconstruction quality, bandwidth, and latency.","keywords":["semantic communication","VQ-VAE","task adaptation","GradCAM","local semantic feedback","latent fusion","bandwidth efficiency","image classification"],"falsifier":"Run TACO without LSF at a fixed 10% token share on STL-10 and measure classification accuracy against the original image's 94.30%; if it falls below that, the headline gain depends entirely on the oracle label-based search rather than on the latent context/task fusion itself.","tokens_in":8840,"feed_emoji":"📡","tokens_out":6704,"duration_ms":58603,"temperature":0.7,"pith_summary":"The paper sets out to show that goal-oriented communication does not need expensive diffusion models at the receiver. It argues that any image can be split into two complementary latent pieces — a low-resolution context encoding the whole scene and a small set of task-specific tokens selected where a downstream model actually looks — and that a single pretrained VQ-VAE can fuse those pieces into a reconstruction that serves both predefined tasks and newly requested ones. On the reported benchmarks this recipe outperforms diffusion-based semantic communication in reconstruction quality (LPIPS 0.056, FID 11.21), uses far less bandwidth, and reconstructs in about 1.5 milliseconds per image. The strongest reported result is that with a local semantic feedback step, TACO reaches 96.49% classification accuracy on STL-10 while transmitting 0.65 KB, higher than the 94.30% accuracy of the original image transmitted at 59.93 KB. If the claim holds, semantic communication becomes a lightweight latent-selection problem rather than a generative sampling problem.","feed_headline":"At 0.65 KB, this codec beats the original image","feed_subtitle":"A lightweight VQ-VAE sends context plus saliency tokens, beating diffusion-based semantic communication.","key_machinery":"The load-bearing object is the shared VQ-VAE latent space with a compression factor of 16 and a fixed codebook of 8192 entries, used for both the context and the task tokens. Two mechanisms select what enters that space: a four-times downsampling of the image produces the context latent, and GradCAM, a gradient-based saliency map from the downstream model, produces a mask whose coordinates pick the task-specific latent embeddings. The fusion step is the latent-level mixup $z_r = (1-M)\\odot z_u + M\\odot z_i$, where $M$ marks the selected coordinates, $z_u$ is the latent of the upsampled context, and $z_i$ carries the transmitted task tokens. This mixup is what lets a single decoder rebuild the image, and the local semantic feedback (LSF) loop is what chooses how many coordinates to send.","core_discovery":"The central claim is that context and task information can be separated in the latent space of one VQ-VAE and recombined at the receiver without retraining the network. The transmitter downsamples the image by a factor of four and encodes the small image to obtain a context latent; it also runs GradCAM on the original image to find the pixels the downstream task model attends to and maps those pixels to the corresponding latent coordinates. The receiver decodes the context latent, upsamples it back to full resolution, re-encodes it to recover a full latent, and replaces the selected coordinates with the transmitted task tokens. A local semantic feedback loop at the transmitter tries different shares of task tokens, picks the smallest share whose reconstruction wins on the target metric, and in doing so reports a 96.49% classification accuracy at 0.65 KB, above the original image's 94.30% at 59.93 KB. The paper also reports that the same procedure lets a receiver switch from classification to object detection after one short feedback round.","pith_inferences":["The LSF gains depend on evaluating reconstructions against ground-truth labels at the transmitter; a label-free proxy such as prediction entropy or reconstruction uncertainty would make the method applicable where labels are unavailable, and testing that proxy is a direct next step.","Because the context/task split lives in a spatially aligned VQ latent, the same design should transfer to video, audio, or point clouds whenever such a quantized latent exists; the paper only demonstrates images.","The finding that a 0.65 KB reconstruction can beat the original image suggests the method is doing semantic denoising, removing distracting background; a stress test under channel noise would show whether that benefit survives in a real link, which the paper leaves to future work.","One could replace the grid search over {10, 20, 30, 50, 70, 90, 100} with a learned selector that predicts the right token share per image, reducing the feedback cost further."],"forward_implications":["Reconstruction drops from thousands of diffusion steps to a single VQ-VAE decode, about 0.0015 seconds per image on an A100 in the paper's measurement.","Bandwidth can be pushed below the raw image size by two orders of magnitude while downstream accuracy rises, because only task-relevant tokens are transmitted.","Adding more task tokens raises classification accuracy in the reported tables, so the transmitted rate can be tuned to a receiver's quality-of-service requirement.","A receiver can switch goals with one round of coordinate feedback rather than retraining, as shown in the classification-to-detection scenario.","The rate bound tightens from $R_i + R_c$ to $R_{fb} \\le R_i$, meaning the feedback mechanism removes the redundancy of sending context tokens that overlap with task tokens."],"supporting_citations":[{"why":"Supplies the VQ-VAE architecture, codebook quantization, and training objective that the whole latent transmission scheme is built on.","marker":"[9]"},{"why":"Supplies GradCAM, the gradient-based saliency method that selects which latent embeddings count as task-specific information.","marker":"[14]"},{"why":"Supplies the pretrained VQ-VAE weights shared by the context and image encoders in all experiments.","marker":"[17]"},{"why":"The diffusion-based semantic communication system whose reconstruction quality and latency TACO is compared against.","marker":"[8]"},{"why":"The latent mixture baseline closest to TACO, used for the FID, LPIPS, and reconstruction-time comparison.","marker":"[10]"},{"why":"Defines the diffusion goal-oriented communication setup and the reconstruction-step cost that motivates a non-diffusion alternative.","marker":"[2]"},{"why":"Provides the Cityscapes benchmark used for image reconstruction and bandwidth evaluation.","marker":"[18]"},{"why":"Provides the Flickr benchmark for the object-detection results, where TACO-LSF reports the best mAP.","marker":"[19]"},{"why":"Provides the STL-10 benchmark whose 96.49% classification accuracy at 0.65 KB is the paper's headline result.","marker":"[20]"}],"fun_headline_variants":["Task-aware codec sends 0.65 KB, beats full image","Context and task tokens split for 0.65 KB semantic link","One codec, many tasks: switch by feedback, no retrain","GradCAM-guided tokens hit 96% at 0.65 KB bandwidth","Semantic codec adapts to tasks with context embedding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole LSF result rests on the transmitter being able to evaluate downstream task performance against ground-truth labels when choosing the token percentage.","fun_headline_variants_meta":{"raw":{"variants":["Task-aware codec sends 0.65 KB, beats full image","Context and task tokens split for 0.65 KB semantic link","One codec, many tasks: switch by feedback, no retrain","GradCAM-guided tokens hit 96% at 0.65 KB bandwidth","Semantic codec adapts to tasks with context embedding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000477,"raw_usage":{"total_tokens":2329,"prompt_tokens":875,"completion_tokens":1454,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":1360}},"tokens_in":491,"tokens_out":1454,"duration_ms":10499,"temperature":1.0,"reasoning_tokens":1360,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:02:11.487402+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run TACO without LSF at a fixed 10% token share on STL-10 and measure classification accuracy against the original image's 94.30%; if it falls below that, the headline gain depends entirely on the oracle label-based search rather than on the latent context/task fusion itself.","supporting_citations":[{"cited_title":"Neural discrete representation learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the VQ-VAE architecture, codebook quantization, and training objective that the whole latent transmission scheme is built on."},{"cited_title":"Grad-cam: Visual explanations from deep networks via gradient-based localization,","cited_arxiv_id":null,"evidence_quote":"Supplies GradCAM, the gradient-based saliency method that selects which latent embeddings count as task-specific information."},{"cited_title":"Diff-go+: An efficient diffusion goal-oriented communication system with local feedback,","cited_arxiv_id":null,"evidence_quote":"The diffusion-based semantic communication system whose reconstruction quality and latency TACO is compared against."},{"cited_title":"LaMI-GO: Latent Mixture Integration for Goal-Oriented Communications Achieving High Spectrum Efficiency","cited_arxiv_id":"2412.17839","evidence_quote":"The latent mixture baseline closest to TACO, used for the FID, LPIPS, and reconstruction-time comparison."},{"cited_title":"Diff- go: Diffusion goal-oriented communications with ultra-high spectrum efficiency,","cited_arxiv_id":null,"evidence_quote":"Defines the diffusion goal-oriented communication setup and the reconstruction-step cost that motivates a non-diffusion alternative."},{"cited_title":"The cityscapes dataset for semantic urban scene understanding,","cited_arxiv_id":null,"evidence_quote":"Provides the Cityscapes benchmark used for image reconstruction and bandwidth evaluation."},{"cited_title":"An analysis of single-layer networks in unsupervised feature learning,","cited_arxiv_id":null,"evidence_quote":"Provides the STL-10 benchmark whose 96.49% classification accuracy at 0.65 KB is the paper's headline result."}],"review_version":1}