{"id":"bb8a6ff2-6a37-4db7-894d-642e2b0e3595","arxiv_id":"2505.18637","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Proposes a standardized semantic coding workflow (tokenization, reorganization, quantization) that extracts general semantic representations for both human- and machine-centric semantic communication, with a case study on wireless image transmission.","lead":"This paper argues that ordinary neural feature compression is not enough for semantic communication, and proposes a 'semantic coding' pipeline of tokenization, token merging, and optional quantization to extract general semantic representations. The authors show a proof-of-concept wireless image transmission system combining this pipeline with a diffusion generator, claiming better low-rate reconstruction than a classical deep JSCC baseline.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Fig. 5 empirical claim is confounded: the semantic-coding system decodes with Stable Diffusion while the neural-coding baseline uses a non-generative JSCC decoder, so the comparison does not isolate the contribution of semantic coding.","rationale":"The reader's stated weakest assumption is that token-embedding similarity preserves semantic relevance, but I judge the more load-bearing weakness to be the confounded empirical comparison. Even if token merging perfectly preserved semantics, the case study would still not demonstrate the paper's central claim because the semantic and neural-coding systems differ in decoder capability. The reader's rationale does mention 'a confounded baseline' and 'the effect of the generative prior,' so there is partial agreement, but the single most load-bearing concern should be the missing matched-decoder control rather than the semantic-similarity assumption. The paper's conceptual framework is coherent and the use of off-the-shelf ToMe on a frozen ViT is a useful, parameter-free illustration; however, the advertised performance advantage is the key quantitative evidence and it is currently under-specified. A conditional acceptance requiring a rate-matched, decoder-matched ablation is the appropriate outcome. I do not recommend rejection because the framework itself may still be valuable even if the specific case study needs revision; I also do not recommend unconditional acceptance because the current Fig. 5 comparison does not isolate the mechanism the paper claims to validate.","tokens_in":10825,"tokens_out":3842,"duration_ms":33376,"concrete_test":"Run a matched-decoder ablation: fix one receiver architecture for both systems. For example, take the neural-coding JSCC decoder and attach it to the semantic tokens, and separately take the Stable Diffusion generator and condition it on neural-coding features, keeping total bitrate identical including token quantization and side information. Recompute the Fig. 5 rate-PSNR and rate-FID curves. If semantic coding no longer dominates all CBR points in both metrics, the claimed advantage should be attributed to the generative prior or to unequal rate accounting rather than to semantic coding.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's strongest advertised result, Section III.D and Fig. 5, is that with 10-30 tokens the semantic-coding system 'performs better across all CBR regimes' than neural coding. As reported, this comparison cannot support that claim. The semantic pipeline transmits ViT/ToMe tokens and reconstructs with a Stable Diffusion generator (Sections III.C-III.D), while the neural-coding baseline is deep JSCC [14] with a learned decoder and no generative prior. The two systems therefore differ in two independent ways: semantic tokenization/reorganization versus plain feature coding, and generative-prior decoding versus direct reconstruction. A generator can improve FID and perceived realism even when the transmitted representation carries less task-relevant content, and a non-generative decoder can be handicapped on perceptual metrics by design. The paper provides no control using the same generator on neural-coding features, nor the same non-generative decoder on semantic tokens. It also does not specify how 10 or 30 tokens are quantized, dimensioned, and mapped to channel symbols at each CBR value, so the claim of 'better across all CBR regimes' may partly reflect unequal bit budgets. Without a matched-decoder, rate-matched comparison, the central empirical evidence that neural coding is not semantic is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper argues that general neural coding, i.e., feature extraction and compressive coding with deep networks, is not sufficient for semantic communication because it captures statistical patterns rather than semantic meaning. The authors propose 'semantic coding' as a standardized workflow consisting of tokenization, reorganization, and optional quantization, and they position it as a pre-processing module before compressive and transmissive coding. They outline a transmitter-side general representation and receiver-side task-specific decoding, and they report a case study on wireless image transmission (Section III.D) in which a ViT/ToMe tokenization pipeline with 10-30 tokens per image and a Stable Diffusion decoder is claimed to outperform a deep JSCC baseline across all tested CBR regimes.","tokens_in":11047,"tokens_out":5702,"duration_ms":48344,"significance":"The conceptual contribution is a useful clarification: separating entropy-oriented feature compression from context-aware semantic representation could help structure future semantic communication systems. The paper explicitly marks the reorganization mechanism as a hypothesis, which is honest, and it provides a concrete albeit limited demonstration. If the framework and the case study were rigorously validated, the paper would provide a valuable reference architecture for the field. However, the current empirical evidence is not sufficient to establish the central claim that this token-based pipeline outperforms neural coding, because the comparison is confounded and under-specified.","major_comments":[{"comment":"The comparison between the proposed semantic-coding system and the neural-coding baseline is confounded by the decoder design. The semantic pipeline uses a Stable Diffusion generator for content recovery, while the baseline [14] is a non-generative JSCC decoder, so any improvement in FID or perceived realism could be attributed to the generative prior rather than to semantic tokenization and reorganization. The paper does not provide a control experiment using the same generator on neural-coding features or a non-generative decoder on semantic tokens. Without such matched decoders, the reported 'performs better across all CBR regimes' does not isolate the contribution of semantic coding.","section":"Section III.D, Fig. 5"},{"comment":"The rate and bit-budget matching between the two systems is not specified. The paper does not state how 10 or 30 tokens are quantized (if at all), how token vectors are dimensioned, or how many channel symbols each CBR value corresponds to for each system. Since CBR is defined as the ratio of transmitted signal dimension k to original dimension m (footnote 4), the claim that semantic coding is better 'across all CBR regimes' is only meaningful if both systems spend the same number of channel symbols per source pixel; the manuscript provides no evidence for this.","section":"Section III.D, Fig. 5"},{"comment":"The load-bearing assumption that 'proximate token embeddings maintain high contextual relevance and typically represent similar meanings or concepts' is presented as a hypothesis ('It is hypothesized that token reorganization works due to its concise contextual representation in concept categorization'). The entire semantic coding pipeline depends on this assumption, but the case study provides no direct validation for downstream tasks, no ablation showing that similarity-based merging preserves task-relevant semantics, and no analysis of when it fails. This needs either a formal argument or substantially more empirical support.","section":"Section II.B.3"},{"comment":"The experimental reporting is incomplete: the dataset is not named ('large-scale image datasets'), no error bars, standard deviations, or multiple random seeds are reported, and the PSNR/FID values are not presented numerically for each CBR level. The visual comparison is shown for a single rate (0.0417 cpp). As a result, the claimed advantage cannot be assessed for statistical significance or reproducibility.","section":"Section III.D"}],"minor_comments":[{"comment":"The throughput claim (524 images/s, Top-1 74.06%) appears to be on a classification task, but the dataset and hardware are not specified, and the comparison is to vanilla ViT-B/16 rather than to the neural-coding baseline. Please clarify.","section":"Section III.D"},{"comment":"The sentence 'The overall JSCC framework for wireless transmission is trained end-to-end on large-scale image datasets' is ambiguous: it should state whether the JSCC encoder and decoder are trained on semantic tokens or on raw images, and how the off-the-shelf ViT/ToMe tokenizer is integrated with a trainable framework.","section":"Section III.D"},{"comment":"The example of reducing 196 tokens to 10-30 tokens is illustrative, but it would be helpful to state how the number of tokens is chosen per image and whether it is adaptive or fixed.","section":"Section II.B.3"},{"comment":"For reproducibility, it would be helpful to cite the exact implementations of the metrics and the generator; the current references are to surveys and to the original model papers.","section":"Section III.B and Table I"}],"recommendation":"major_revision","confidential_remarks":"The paper reads partly as a position/tutorial article and partly as a research article. The case study, as currently presented, is not at the level of rigor expected for the strong empirical claim in the title and abstract. If the journal is willing to accept a primarily conceptual contribution, the authors should soften the empirical claims and emphasize the framework; otherwise, they should substantially strengthen the experimental design, including matched decoders, rate matching, dataset specification, and statistical reporting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper gives the semantic-communication community a clean, usable definition of semantic coding as tokenization, reorganization, and optional quantization. That's a real contribution, and the authors are honest that some of it is hypothesis (Section II.B.3). The case study, though, doesn't support the strong \"better across all CBR regimes\" claim. The semantic system ends with a Stable Diffusion generator, while the neural-coding baseline is plain deep JSCC with a learned decoder. Those are two independent differences, so the comparison isn't isolating semantic coding. The paper also doesn't say how 10-30 tokens are quantized or mapped to channel symbols at each CBR value, so the bit budgets might not be equal. No error bars, no dataset name, no multiple runs. The stress-test note I saw lands: the confound is real.\n\nWhat the paper does well: it steps back and gives a principled decomposition of the coding pipeline, and it correctly distinguishes general-purpose neural feature compression from semantic representation. The three-step pipeline is a plausible standardization. The throughput number (524 im/s on ViT-B/16 with 74% top-1, ~5% drop) is interesting, though not the core of the wireless story. The discussion of consistency vs realism and the metric table are useful orientation for newcomers.\n\nThe soft spots are mostly in the experimental validation. The central conceptual argument is a position statement, not a theorem, and that's fine for this kind of paper as long as it's labeled as such. It mostly is. The self-citations are appropriate here—they're for the specific components being reused. I wouldn't call the lack of formal proofs a flaw; this is an architecture/framework paper.\n\nBottom line: this is a paper worth refereeing, but the experiment needs to be redone with a matched decoder and a proper rate accounting. If the authors fix that, the framework has legs. A reader looking for a definition and a workflow taxonomy will get value from this; a reader looking for evidence that semantic coding beats neural coding should be skeptical until the control is added.","headline":"Useful conceptual framework for semantic coding, but the headline empirical claim is confounded by the generative decoder and shouldn't be taken at face value.","tokens_in":11586,"tokens_out":2498,"would_cite":true,"duration_ms":18972,"reading_group":"yes","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 argues that neural feature compression is not semantic coding, and that a tokenize–reorganize–quantize pipeline can transmit images with 10–30 tokens per image better than a learned feature baseline.","keywords":["semantic communication","semantic coding","neural coding","tokenization","token reorganization","contextual modeling","joint source-channel coding","wireless image transmission"],"falsifier":"Take two objects whose token embeddings are close but whose semantics differ (e.g., a wolf and a husky, or two fine-grained bird species), transmit each at 10 tokens through the same pipeline, and measure whether a receiver fine-tuned to distinguish them still can; if the merged representation cannot separate the two categories, the paper's premise that embedding-space proximity preserves semantic content fails for that input class.","tokens_in":10618,"feed_emoji":"📡","tokens_out":7113,"duration_ms":56597,"temperature":0.7,"pith_summary":"Semantic communication has leaned on deep neural networks to extract and compress features from raw data, but this paper argues that such 'neural coding' is not the same as coding semantics: it captures statistical patterns without grounding them in concepts or context. The paper proposes a standardized three-step procedure it calls semantic coding — tokenization, similarity-based reorganization, and optional quantization — that turns raw data into a compact set of semantic tokens meant to represent the meaning behind the data. The payoff, if correct, is a unified coding workflow: one general semantic representation serves both human-facing reconstruction and machine-facing recognition with only light task-specific adaptation at the receiver, and wireless transmission of just 10–30 tokens per image can beat a conventional learned feature-compression baseline. The article is an attempt to establish a working definition, not just a new network architecture, so that semantic communication can be discussed, compared, and standardized around the idea of conveying meanings rather than features.","feed_headline":"Just 10-30 tokens per image outperform neural coding on wireless links","feed_subtitle":"A token-based semantic coding pipeline shrinks 196 patches to 10-30 tokens and still beats feature compression over the air.","key_machinery":"The load-bearing object is the semantic coding pipeline with its three operations. Tokenization uses an attention-based encoder to turn raw data into discrete tokens, which are the basic learnable units of representation. Reorganization is the step that makes representation compact: tokens are grouped and similar ones are matched and merged based on embedding-space similarity (implemented as bipartite soft matching with eight merges per transformer layer), reducing 196 tokens to roughly 10–30 while aiming to preserve each concept in the image. Optional quantization then maps the merged tokens to discrete codes so that one code can stand for one concept, which also aids robustness and compatibility with generative receivers. The workflow additionally decouples a 'general' semantic encoder from task-specific decoders, so the same compact representation can be conditioned by a generator for reconstruction or a discriminator for recognition.","core_discovery":"The central claim is that deep feature extraction ('neural coding') is necessary but not sufficient for semantic communication, and that the missing piece is a dedicated semantic representation stage. The paper defines semantic coding as tokenization (decomposing raw data into discrete learnable tokens), reorganization (merging nearby tokens in an embedding space, e.g., via bipartite soft matching, so that semantically related tokens collapse into one representative token), and optional quantization (mapping tokens to discrete codebook entries). In the reported case study, a pretrained visual transformer's 196 tokens are reduced to 10 or 30 tokens per image, fed through a joint source-channel coding transmitter over an AWGN channel, and reconstructed by a generative receiver; the paper reports better PSNR and FID than a neural-coding baseline across all tested channel bandwidth ratios, with a particularly large gap at low rates. The result is meant to show that compact, context-aware semantic representations, not merely compressed features, are the right unit of transmission for semantic communication.","pith_inferences":["If embedding-space proximity really tracks semantic synonymy, then the ideal number of tokens for an image should scale with the number of distinct concepts in the scene, which points to variable-rate semantic coding rather than fixed 10- or 30-token budgets.","The same reorganization step could be applied to other modalities (audio, video, point clouds) as long as a suitable attention-based tokenizer exists, turning semantic coding into a generic multimodal front-end for wireless transmission.","A direct way to stress-test the paper's thesis is to measure downstream task performance (e.g., fine-grained classification) on reconstructed images: if a 10-token representation preserves accuracy much better than a 196-token one on some tasks but not others, the semantic-preservation assumption is task-dependent.","The paper's comparison is against a specific neural-coding baseline; a stronger test of the claim that semantic coding is not just better compression would be to compare against an equally optimized learned codec at matched rates and distortions."],"forward_implications":["If semantic coding works as claimed, a semantic communication system can be built on top of a pretrained attention-based tokenizer without retraining the tokenizer, since reorganization and quantization are applied to its outputs.","Reducing an image to 10–30 tokens means the transmitted rate is set by the number of semantic concepts, not by the resolution of the image, so low-rate regimes become the natural operating point.","The same general semantic representation can feed both human-centric tasks (via generative receivers) and machine-centric tasks (via discriminative receivers), with only minimal fine-tuning at the receiver side.","Quantized sparse codes are expected to be more robust to channel noise and adversarial perturbations than continuous feature vectors, because a few discrete codes carry the essential concepts.","The reported throughput and accuracy numbers suggest that merging tokens is also a cheap way to speed up inference, with reduced token counts and only a modest drop in classification accuracy."],"supporting_citations":[{"why":"Supplies the three-level view of communication (syntactic, semantic, effectiveness) that motivates placing a semantic layer in the coding workflow.","marker":"[1]"},{"why":"Provides the attention-based tokenizer whose fixed number of output tokens semantic coding reorganizes.","marker":"[4]"},{"why":"Supplies the bipartite soft matching algorithm used to merge similar tokens in the reorganization step.","marker":"[5]"},{"why":"Supports the claim that very few tokens suffice for reconstruction and generation, motivating the 10–30 token regime.","marker":"[6]"},{"why":"Grounds the consistency-realism trade-off that determines how many tokens and which evaluation metrics to use.","marker":"[9]"},{"why":"Provides the nonlinear transform source-channel coding used as the compressive and transmissive stage after semantic coding.","marker":"[11]"},{"why":"Supplies the latent diffusion generator that reconstructs images from semantic tokens at the receiver.","marker":"[12]"},{"why":"Provides the deep joint source-channel coding scheme used as the neural-coding baseline in the case study.","marker":"[14]"}],"fun_headline_variants":["Token semantics beat neural coding over wireless links","Why neural coding falls short: semantic tokens win over the air","Semantic token pipeline outperforms feature compression on wireless","Neural coding insufficient: token-based semantics lead in transmission","From 196 patches to 10 tokens: semantic coding outperforms neural"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline rests on the assumption that tokens which are close together in embedding space really do mean the same thing, so merging them loses no task-relevant meaning; if similar-looking tokens can stand for different concepts, the compact representation silently drops information the receiver needs.","fun_headline_variants_meta":{"raw":{"variants":["Token semantics beat neural coding over wireless links","Why neural coding falls short: semantic tokens win over the air","Semantic token pipeline outperforms feature compression on wireless","Neural coding insufficient: token-based semantics lead in transmission","From 196 patches to 10 tokens: semantic coding outperforms neural"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000175,"raw_usage":{"total_tokens":1270,"prompt_tokens":912,"completion_tokens":358,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":277}},"tokens_in":528,"tokens_out":358,"duration_ms":3583,"temperature":1.0,"reasoning_tokens":277,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:27:37.080059+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take two objects whose token embeddings are close but whose semantics differ (e.g., a wolf and a husky, or two fine-grained bird species), transmit each at 10 tokens through the same pipeline, and measure whether a receiver fine-tuned to distinguish them still can; if the merged representation cannot separate the two categories, the paper's premise that embedding-space proximity preserves semantic content fails for that input class.","supporting_citations":[{"cited_title":"Recent contributions to the mathematical theory of com- munication,","cited_arxiv_id":null,"evidence_quote":"Supplies the three-level view of communication (syntactic, semantic, effectiveness) that motivates placing a semantic layer in the coding workflow."},{"cited_title":"Token merging: Your vit but faster,","cited_arxiv_id":null,"evidence_quote":"Supplies the bipartite soft matching algorithm used to merge similar tokens in the reorganization step."},{"cited_title":"An image is worth 32 tokens for reconstruction and generation,","cited_arxiv_id":null,"evidence_quote":"Supports the claim that very few tokens suffice for reconstruction and generation, motivating the 10–30 token regime."},{"cited_title":"Rethinking lossy compression: The rate- distortion-perception tradeoff,","cited_arxiv_id":null,"evidence_quote":"Grounds the consistency-realism trade-off that determines how many tokens and which evaluation metrics to use."},{"cited_title":"High- resolution image synthesis with latent diffusion models,","cited_arxiv_id":null,"evidence_quote":"Supplies the latent diffusion generator that reconstructs images from semantic tokens at the receiver."}],"review_version":1}