{"id":"90ab9f2c-d2ab-4d57-af0e-614b1c1b5e4f","arxiv_id":"2501.10926","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A semantic successive interference cancellation scheme that uses decoded text as side information improves multi-user text transmission quality in simulated multiple access channels.","lead":"This paper extends a deep learning text communication system to multiple users sharing one wireless channel, using the decoded message of the strongest user as a hint to decode the next user. It reports better semantic similarity and BLEU scores than the listed baselines and a cheaper partial retraining method for when users join the network.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Side-information benefit is not tested on uncorrelated or weakly correlated user texts, so the central semantic-SIC gain may be an artifact of the SNLI dataset.","rationale":"The reader's weakest-assumption analysis and my stress-test converge on the same load-bearing concern: the side-information benefit presupposes strong semantic correlation between different users' texts, and the experiments are run only on SNLI, a dataset chosen for that very property. The paper asserts the advantage of semantic SIC without testing the regime in which the side information is uninformative. Since the central claim is about the general advantage of semantic SIC over benchmarks, not just about a particular dataset, the evidence is insufficient as presented. I therefore recommend moving the verdict from CONDITIONAL to REJECT: the claim as stated is not supported by the experimental scope. The concrete proposed control experiment would settle whether the concern actually lands, allowing the verdict to be revisited with broader evidence.","tokens_in":18514,"tokens_out":1136,"duration_ms":14647,"concrete_test":"Repeat the 2+1 and 3+2 user experiments with a control dataset that has weak inter-sentential correlation across users (e.g., Europarl, or a deliberately shuffled/re-topic-grouped SNLI split), holding all other training and channel settings identical. If the gap between 'with SI' and 'without SI' narrows to near zero, the claimed side-information benefit is dataset-specific; if the gap persists, the claim is robust.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that decoding in the semantic domain with already-decoded texts as side information improves over decoding without such side information. The mechanism, IFG_i in Eq. (34)-(35), is only useful if knowing T_1,...,T_{i-1} provides information about T_i. The paper's own dataset rationale states that SNLI is chosen precisely because it exhibits strong inter-sentential connections, and it explicitly contrasts this with other datasets such as Europarl. The head-to-head evidence for the side-information gain is therefore obtained only in the favorable regime where user texts are deliberately correlated; no experiment varies or measures the degree of semantic correlation between users. If users transmit unrelated texts, the IFG mechanism may provide no gain or even degrade decoding, and the claimed advantage over no-side-information baselines could vanish. This matters because the semantic-SIC contribution is conceptually distinct from symbol-level SIC only through this correlation exploitation; without it, the method reduces to conventional SIC plus learned decoders.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a deep learning-enabled semantic communication system for uplink multiple access channels (MAC), extending the single-user DeepSC framework to K users. The approach applies successive interference cancellation (SIC) in the semantic domain: after decoding each user's text, the decoded text is re-encoded and its signal is removed from the received mixture; in addition, an integrated feature generator (IFG) uses previously decoded texts as side information to improve decoding of subsequent users. To handle dynamic user addition, the paper also proposes a pretraining scheme and a partial retraining scheme that updates only the newly added users' networks while reusing existing users' encoders/decoders. Experiments on the SNLI corpus compare the proposed methods with ablated baselines (with/without side information, full/partial retraining) and with traditional Huffman+LDPC and Huffman+Polar schemes, reporting semantic similarity and BLEU scores across several channel conditions.","tokens_in":18749,"tokens_out":9341,"duration_ms":92080,"significance":"If the claims hold, the paper provides a meaningful extension of semantic communication to multi-user networks and introduces a practical retraining mechanism for dynamic user sets. The internal ablations show consistent gains from side information and reduced training time from partial retraining, and the open-source code is a plus. However, the significance is somewhat limited by the evaluation: the traditional baselines do not use any interference cancellation, so the comparison does not isolate the value of the semantic SIC; and the side-information benefit is demonstrated only on SNLI, which was selected for its strong inter-sentential correlations, leaving the behavior on uncorrelated sources untested.","major_comments":[{"comment":"The traditional baselines (Huffman+LDPC and Huffman+Polar with 64-QAM) are described without any interference cancellation or multi-user detection, whereas the proposed method relies on successive interference cancellation. As a result, the performance differences in Figs. 9–12 could be attributed to SIC itself rather than its semantic aspect. Please include a baseline that performs conventional symbol-level SIC with the same channel coding, or justify why the current comparison is appropriate for the claimed advantage over existing benchmark methods.","section":"Section VI-A"},{"comment":"The side-information benefit is evaluated only on the SNLI dataset, which the authors selected because it exhibits much stronger inter-sentential connections than other corpora. The mechanism in Eqs. (34)–(35) is thus tested exclusively in a high-correlation regime. No experiment varies the degree of semantic correlation between user texts (e.g., using random sentence pairings or a corpus such as Europarl). Without such an experiment, it is unclear whether the 'semantic SIC' advantage persists when users transmit unrelated content, which is a common scenario in multiple access networks. Please add such experiments or explicitly qualify the claims to correlated sources.","section":"Section VI-A"},{"comment":"The tradeoff factors τ_i in the loss function (39) are not specified; it is unclear how they are set or whether they are tuned. Since the partial retraining performance is a central contribution, please provide the values used or a sensitivity analysis to show that the results are not sensitive to these factors.","section":"Section V-C, Eq. (39) and Algorithm 2"}],"minor_comments":[{"comment":"The zero-padding procedure in Eq. (19) pads each sentence to a fixed length N; however, the paper does not specify what happens when a sentence exceeds N words (truncation?) or how the positional encoding handles the padded dummy words. Clarify this in the text.","section":"Section III, Eq. (19)"},{"comment":"The description of the partial retraining decoding order is inconsistent: Section V-C states that after recovering the new user's message, the old users in group G2 are decoded 'as if the new user does not exist', but Algorithm 4 decodes all old users in a single step after canceling the new users' signals. Please reconcile the algorithm with the narrative.","section":"Section V-C and Algorithm 4"},{"comment":"The statement 'we allocate 18 symbols for each word' is not connected to the encoding parameters: in the system model, each word is compressed to a c-dimensional vector and then mapped to c/2 complex symbols. Please explain how 18 symbols per word relates to c and the total block length M.","section":"Section VI-A"},{"comment":"The paper calls the method 'semantic SIC', but the actual interference cancellation in Eqs. (30) and (32) is performed in the complex symbol domain after re-encoding; only the side-information generation (Eq. (35)) is in the semantic embedding domain. Consider clarifying this terminology.","section":"Section IV-B"},{"comment":"The text contains a typo: 'acheive' should be 'achieve' in the description of the Rayleigh fading BLEU results.","section":"Section VI-B, Fig. 11"},{"comment":"Reference [47] (SNLI) lists the arXiv identifier 2201.01389, which duplicates reference [13]; the correct identifier for the SNLI corpus is 1508.05326. Please correct this citation.","section":"References"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Plainly: this is a real extension of DeepSC to the MAC, not a repackaging. The two ideas—performing SIC in the word-embedding domain and injecting previously decoded texts as side information through an integrated feature generator—are clearly described, and the internal experiments (side information vs. none, partial vs. full retraining) support the paper's own claims. The partial retraining protocol is a practical contribution; cutting training time by roughly a third while keeping most of the gain is a useful result if it holds. Code is linked, which helps.\n\nThe soft spots are real but not fatal. The biggest one is the dataset: SNLI is chosen precisely because its sentences are semantically paired (entailment/contradiction), so the side-information gain is demonstrated only in a regime where user texts are correlated by construction. There is no experiment with weak or no correlation, so we don't know if the benefit vanishes, flips negative, or just shrinks. That is a gap, not a contradiction; the paper is honest about the choice, but the abstract's general phrasing overstates the scope. A referee should ask for either a Europarl or random-pair experiment or a clear scoping of the claim to correlated-text applications. The second issue is the baseline comparison: Huffman+LDPC/Polar with 64-QAM are not described as using any interference cancellation, so the comparison is stacked in favor of the SIC-based methods. It shows the semantic system wins, but it doesn't isolate the semantic-SIC advantage. Missing error bars are a minor omission, and the opening claim about 'breaking performance limits from statistical information theory' is not supported by anything in the paper.\n\nI disagree with the stress-test note if it implies the work is invalid: the mechanism is plausible and the internal comparisons are consistent. But the concern is a legitimate scope limitation. The paper would still earn a serious referee: the idea is novel enough, the experiments are reproducible in principle, and the partial retraining result is useful. I'd send it to review, and I'd expect a revised version to either add the uncorrelated-text experiment or narrow the claims accordingly. For a reading group on semantic communications, yes, bring it up.","headline":"Real extension of DeepSC to the MAC with semantic side-information SIC, but the side-information gain is only shown on deliberately correlated SNLI texts, and the traditional baselines lack SIC.","tokens_in":19235,"tokens_out":2806,"would_cite":true,"duration_ms":31691,"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":"Successive interference cancellation for multi-user channels can be moved from the digital symbol domain into the semantic word-embedding domain, where decoded text from one user helps decode the next.","keywords":["semantic communication","successive interference cancellation","multiple access channel","DeepSC","Transformer","word embedding","side information","partial retraining"],"falsifier":"Run the same system on a MAC whose users transmit semantically unrelated texts, for example sentences sampled independently from different topics so there is no entailment-style pairing between users, and compare the with-side-information and without-side-information variants under BERT similarity and BLEU. If the side-information variants do not beat the no-side-information baselines, the claimed benefit rests entirely on the dataset's inter-sentence correlations and does not hold for general user traffic.","tokens_in":18343,"feed_emoji":"📡","tokens_out":8688,"duration_ms":78461,"temperature":0.7,"pith_summary":"The paper proposes a successive interference cancellation (SIC) scheme that works in the semantic domain, extending deep-learning semantic communication from a single point-to-point link to a multi-user multiple access channel (MAC). In classic SIC, once a user's message is decoded, only their signal is subtracted from the received waveform; here, the decoded text itself is kept and used as side information to help decode the next user's text, exploiting semantic correlations between users' messages. The paper also contributes a pretraining and partial retraining scheme that lets new users join an existing MAC without retraining all of the old encoders and decoders from scratch. The reported experiments claim that the side-information variants achieve higher BERT-similarity and BLEU scores than full retraining without side information, and far outperform classic digital coding baselines using 64-QAM.","feed_headline":"Decoded texts become side information for multi-user decoding","feed_subtitle":"Semantic SIC turns recovered sentences into side information and cuts retraining time by 30%.","key_machinery":"The load-bearing mechanism is a neural module called the integrated feature generator (IFG). For user $i$'s sentence, it takes the current compressed word-semantic vector $\\hat{r}^i_j$ together with the recovered compressed vectors of previously decoded users $\\tilde{r}^1_j, \\ldots, \\tilde{r}^{i-1}_j$ and produces $g^i_j = \\text{ReLU}(\\theta_i(\\pi_i(\\hat{r}^i_j), \\Omega^i_1(\\tilde{r}^1_j), \\ldots, \\Omega^{i-1}_i(\\tilde{r}^{i-1}_j)) + \\hat{r}^i_j)$, which the autoencoder decoder converts into the word-semantic vector used to generate the next text. This fusion is what changes SIC from a pure subtraction operation into a side-information operation. Around this, the training scheme uses a joint cross-entropy loss, a per-user pretraining stage, and a partial retraining algorithm that freezes the existing users' encoders and decoders when new users join, so only the new users' networks and the shared decoder are updated.","core_discovery":"The paper's central claim is that SIC for a multiple access channel can be performed in the semantic domain rather than only in the digital symbol domain. After the strongest user's text is recovered, the usual SIC step is kept, namely reconstructing and subtracting that user's symbols from the received signal, but the recovered text itself is not discarded. Its compressed word-semantic vectors are passed through an integrated feature generator together with the next user's noisy features, and the combined feature is decoded into the next text. The paper argues that because natural-language messages from different users are often semantically related, this side information improves semantic similarity and BLEU scores, and that a partial retraining scheme which freezes old users' networks and trains only the new users' networks preserves most of the benefit while reducing per-iteration training time by more than 30 percent.","pith_inferences":["A testable extension is to modulate the side-information branch of the IFG by an estimated inter-user semantic correlation, letting the network ignore side information when users' messages are unrelated; this would reveal how much of the reported gain is dataset-specific.","The subtract-then-reuse-recovered-content recipe should transfer to image or audio semantic MACs, replacing BERT/BLEU with SSIM or SDR, which the authors themselves list as future work.","Freezing old users' networks during partial retraining implies an incremental onboarding protocol for IoT devices that speak semantic protocols: new devices can enter the MAC without reconfiguring existing ones, which is a deployment property beyond the tested performance metrics.","If semantic SIC holds up, the effective multiuser limit of such systems depends on inter-user semantic redundancy as much as on power and bandwidth, suggesting a semantic capacity region notion that the paper does not formalize."],"forward_implications":["In a K-user MAC, decoding can proceed in order of descending received power, and each successfully decoded text is fed into the decoding of the next text rather than discarded.","Adding a new user to a trained MAC can be handled by partial retraining that keeps existing encoders and decoders fixed, reducing per-iteration training time by more than 30 percent relative to full retraining with side information.","The reported results imply side information is most valuable in low-SINR and fading conditions, where partial retraining with side information outperforms full retraining without it.","Because the method is built on SIC, the same semantic-side-information idea can be carried to broadcast and interference channels whenever SIC applies, although capacity achievability is not guaranteed there."],"supporting_citations":[{"why":"Supplies the point-to-point DeepSC encoder-decoder framework that the paper extends to the multi-user MAC.","marker":"[14]"},{"why":"Provides the Transformer architecture used for semantic encoding, decoding, and position encoding.","marker":"[30]"},{"why":"Represents the prior two-user semantic non-orthogonal multiple access scheme that this K-user semantic MAC work generalizes.","marker":"[27]"},{"why":"Is the earlier integration of SIC into semantic MIMO that stays in the conventional i.i.d. domain, the contrast for semantic-domain SIC.","marker":"[38]"},{"why":"Defines the BERT-based cosine similarity used as the semantic performance metric.","marker":"[39]"},{"why":"Defines the BLEU n-gram score used as the second semantic accuracy metric.","marker":"[41]"},{"why":"Provides the paired-sentence corpus whose entailment and contradiction structure supplies the inter-sentence semantic correlation that side information exploits.","marker":"[47]"}],"fun_headline_variants":["Semantic SIC: decoded text as side info","Recovered text cuts retraining time by 30%","Multi-user decoding gets a semantic boost","SIC in the semantic domain boosts BLEU","Words as side info: semantic SIC for MAC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The gain from side information depends on users' texts being semantically related, so a decoded sentence from one user genuinely narrows the possibilities for another user's sentence; without that correlation the side information may add nothing or even hurt.","fun_headline_variants_meta":{"raw":{"variants":["Semantic SIC: decoded text as side info","Recovered text cuts retraining time by 30%","Multi-user decoding gets a semantic boost","SIC in the semantic domain boosts BLEU","Words as side info: semantic SIC for MAC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000259,"raw_usage":{"total_tokens":1597,"prompt_tokens":969,"completion_tokens":628,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":553}},"tokens_in":585,"tokens_out":628,"duration_ms":6648,"temperature":1.0,"reasoning_tokens":553,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:49:34.294817+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same system on a MAC whose users transmit semantically unrelated texts, for example sentences sampled independently from different topics so there is no entailment-style pairing between users, and compare the with-side-information and without-side-information variants under BERT similarity and BLEU. If the side-information variants do not beat the no-side-information baselines, the claimed benefit rests entirely on the dataset's inter-sentence correlations and does not hold for general user traffic.","supporting_citations":[{"cited_title":"Deep learning e nabled semantic communication systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the point-to-point DeepSC encoder-decoder framework that the paper extends to the multi-user MAC."},{"cited_title":"Attention is all you need,","cited_arxiv_id":null,"evidence_quote":"Provides the Transformer architecture used for semantic encoding, decoding, and position encoding."},{"cited_title":"Exploiting semantic communication fo r non- orthogonal multiple access,","cited_arxiv_id":null,"evidence_quote":"Represents the prior two-user semantic non-orthogonal multiple access scheme that this K-user semantic MAC work generalizes."},{"cited_title":"Semantic Importance-Aware Based for Multi-User Communication Over MIMO Fading Channels","cited_arxiv_id":"2312.16057","evidence_quote":"Is the earlier integration of SIC into semantic MIMO that stays in the conventional i.i.d. domain, the contrast for semantic-domain SIC."},{"cited_title":"BLEU: A m ethod for automatic evaluation of machine translation,","cited_arxiv_id":null,"evidence_quote":"Defines the BLEU n-gram score used as the second semantic accuracy metric."}],"review_version":1}