{"id":"08705f6f-db8a-4792-8207-9eb2e83f5b41","arxiv_id":"2502.06118","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"ToDMA lets uncoordinated devices share a token codebook and transmit non-orthogonally, then uses a pretrained transformer to repair token collisions from context.","lead":"ToDMA is a new wireless access scheme in which many devices send images simultaneously using shared token codes, and a pretrained AI model repairs the collisions. It trades orthogonal bandwidth for generative source intelligence, promising lower latency for massive IoT networks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The advertised 'better image quality' and 'several times lower latency' are not supported by the paper's own figures and formulas: Orth-Com has higher PSNR at BER=0, and ToDMA latency exceeds Orth-Com at KT=500 for typical BERs.","rationale":"I focused on the strongest claim as stated in the abstract and conclusion, since a load-bearing concern should test exactly what the paper advertises. The paper promises superiority over both orthogonal and non-orthogonal baselines on both latency and image quality. The numerical evidence contradicts this. Fig. 3 plots Orth-Com at BER=0, an idealized assumption, and even under that ideal Orth-Com has higher PSNR and lower LPIPS than ToDMA, so the 'better image quality' part of the claim is false as stated. The latency claim also fails a direct arithmetic check using the paper's own formulas, as detailed above. These are not external generalization worries; they are internal mismatches between the paper's claims and its own reported results. The transformer-generalization concern identified by the reader is valid but less immediate: the simulation is in-distribution for MaskGIT on ImageNet-100, so the demonstrated collision-recovery performance is credible for the stated setup. The more pressing issue is that the headline quantitative claims are not supported by the figures and equations already in the manuscript. The fix is straightforward: either revise the claims to say 'comparable image quality to an error-free orthogonal baseline' and 'lower latency only for large KT', or rerun the comparison against a finite-BER orthogonal baseline that actually uses the paper's own rate formula. For this reason, the reader's CONDITIONAL verdict remains appropriate; the concern does not change the verdict, but it sharpens what must be addressed before acceptance.","tokens_in":9728,"tokens_out":22001,"duration_ms":210262,"concrete_test":"Rerun the comparison with Orth-Com evaluated at the paper's own finite target BER (using the R_Orth formula in Section VI-B to set the QAM rate) rather than BER=0, for KT in {400, 600, 800, 1000} at SNR=25 dB, and report latency, PSNR, and LPIPS for both schemes. If ToDMA is not strictly better than Orth-Com on both metrics in the plotted range, the abstract and conclusion claims must be revised; if it is better, the overclaim is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, as worded in the Abstract and Conclusion, is that ToDMA outperforms both context-unaware orthogonal and non-orthogonal methods in image transmission, achieving lower latency and better image quality. The paper's own simulation results do not support the orthogonal part of this claim. In Fig. 3, Orth-Com is evaluated at BER=0, an idealized error-free baseline; even under this ideal, its PSNR is above ToDMA and its LPIPS is below ToDMA for all plotted K. Thus ToDMA does not achieve 'better image quality' than the orthogonal baseline; the quality advantage holds only against the non-orthogonal Non-Orth Com scheme. The latency claim also fails a direct arithmetic check using the paper's own formulas in Section VI-B. At SNR=25 dB, KT=500, BER=10^-3, R_Orth = (N_s f_s / KT) log2(1 + 1.5/(-ln(5 BER)) SNR) ≈ 199.7 kbps, giving Orth-Com latency 2560/199700 ≈ 12.8 ms, while ToDMA latency = LN/(N_s f_s) = 1024*256/(1024*15e3) ≈ 17.1 ms. So ToDMA is slower at this plotted operating point. The 'several times' latency advantage only emerges for KT well above the plotted range and depends on very low target BERs. These are internal mismatches between the headline claim and the numerical evidence, not robustness concerns about external data distributions. The transformer-generalization issue raised by the reader is real but secondary: the simulation uses MaskGIT on ImageNet-100, which is in-distribution, so the demonstrated TER is credible for the stated setup. The baseline comparison is what a reader checks first, and it currently does not support the strongest claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"ToDMA is a proposed multiple access scheme in which active devices share a tokenizer and a fixed orthonormal modulation codebook, transmit token-modulated codewords non-orthogonally over the same time-frequency resources, and rely on a base station with many antennas to detect active tokens, assign them to devices by CSI, and fill unresolved token collisions using a pre-trained bidirectional masked image transformer. The paper evaluates the scheme on ImageNet-100 image transmission with a VQ-GAN tokenizer and MaskGIT, comparing against a context-unaware non-orthogonal scheme ('Non-Orth Com') and an orthogonal QAM scheme ('Orth-Com') that is simulated at BER=0. The main claimed advantages are lower latency and better image quality than both benchmarks.","tokens_in":10001,"tokens_out":6133,"duration_ms":52458,"significance":"The idea of moving collision resolution from the physical layer into a pretrained generative source model is interesting and potentially relevant for 6G massive access, and the system model is presented clearly enough to be simulated. The paper's central quantitative claims, however, are currently overstated: the orthogonal baseline is evaluated in an idealized error-free regime, and the reported figures and latency formulas do not support the advertised quality and latency advantages over that baseline. If the claims are appropriately weakened and the orthogonal comparison is made resource-matched, the core architecture and the observed TER trend remain a useful contribution to the semantic multiple access literature.","major_comments":[{"comment":"The claim that ToDMA 'outperforms context-unaware orthogonal and non-orthogonal communication methods ... achieving lower latency and better image quality' is contradicted by the paper's own results. In Fig. 3(b) and 3(c), the Orth-Com baseline, evaluated at BER=0, achieves higher PSNR and lower LPIPS than ToDMA for every plotted value of K; and using the paper's formulas with the stated parameters (SNR=25 dB, KT=500, BER=10^-3), the Orth-Com latency is roughly 12.8 ms while the ToDMA latency is about 17.1 ms. The headline claim must be weakened or the orthogonal benchmark must be redefined as a resource-matched practical scheme.","section":"Abstract, Section III, Fig. 3, Section VI-B"},{"comment":"The random access model says that the BS does not know which devices are active, but the token assignment step in Section V-C assumes the CSI of the K active devices is known. The paper gives no activity detection or channel estimation procedure that would provide this CSI in a grant-free setting, so the receiver design does not yet implement the claimed uncoordinated massive access scenario.","section":"Section IV-B and Section V-C"},{"comment":"The Orth-Com baseline in Fig. 3 is not a fair or practical comparison: it is simulated at BER=0, i.e., with perfect error-free transmission, whereas ToDMA and Non-Orth Com operate at finite SNR with a fixed spreading factor L=Q. The authors should instead compare against an orthogonal scheme consuming the same total bandwidth and energy, or clearly state that Orth-Com is an idealized upper bound and not a competing scheme to be outperformed.","section":"Section VI-A, Fig. 3"},{"comment":"The masked-token prediction component is a pretrained MaskGIT model trained on natural images, and the simulations only use ImageNet-100 images, i.e., in-distribution data. The paper does not evaluate the sensitivity of the TER advantage to out-of-distribution sources or to collision rates that exceed the model's completion ability; this should be discussed and, ideally, experimentally characterized before claiming a general token-domain multiple access framework.","section":"Section V-D"}],"minor_comments":[{"comment":"The notation is not always consistent: the frame is called 'TODMA' in the Section IV heading and 'ToDMA' elsewhere; the abstract contains a stray space in 'T oDMA'.","section":"Throughout"},{"comment":"The threshold Th is introduced as a design parameter and set to 2σ^2 in simulation, but no sensitivity analysis is provided; a brief discussion of how the performance depends on Th would improve reproducibility.","section":"Section V-C"},{"comment":"The assumption that ToDMA and Non-Orth Com have perfect estimation of the token set Pn in Fig. 4 is stated in the caption but not in the main text; this should be part of the simulation description.","section":"Section VI-A, Fig. 4"},{"comment":"The latency comparison ignores beacon overhead, tokenization/detokenization delay, and transformer inference time at the base station, all of which are relevant for a complete latency comparison.","section":"Section VI-B"},{"comment":"The concept of 'semantic orthogonality' is illustrated only with examples; a formal definition or measurable criterion would strengthen the paper, since this notion is central to the claimed collision-mitigation capability.","section":"Section II"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The paper has a clearly specified architecture and a plausible TER trend, but the authors should be asked to reconcile the abstract and conclusion with their own figures and formulas before publication. The skeptical reviewer's arithmetic checks out, and the 'Orth-Com' baseline is ideal in a way that makes the comparison unfair. I recommend a major revision with specific instructions to reword the claims and provide a resource-matched orthogonal baseline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read it. The core idea is real: let devices share a VQ-GAN tokenizer and an orthonormal modulation codebook, transmit non-orthogonally, and repair token collisions with a pretrained bidirectional transformer using candidate token sets and CSI-based assignment. I haven't seen that exact assembly in the multiple-access literature, and the simulation supports it: against the context-unaware non-orthogonal scheme (random filling), ToDMA's TER/PSNR/LPIPS advantage is large and consistent as K grows. That part is credible and worth building on.\n\nThe soft spots are mostly in the advertised claims, not in the architecture. The abstract and conclusion say ToDMA outperforms context-unaware orthogonal and non-orthogonal methods with lower latency and better image quality. That is not what Fig. 3 shows. Orth-Com is plotted at BER=0, an idealized error-free baseline, and it has higher PSNR and lower LPIPS than ToDMA for every K. So the quality claim only holds against the non-orthogonal baseline. The latency claim also fails a direct check with the paper's own formulas at SNR=25 dB, KT=500, BER=10^-3: Orth-Com is about 12.8 ms, ToDMA 17.1 ms. The 'several times' advantage only appears at total device counts well above the paper's setup or at very low target BERs. These are internal mismatches between the headline and the numbers, and they need to be fixed before publication, either by softening the claims or by adding a resource-matched comparison (e.g., same total bandwidth and channel use, and a practical coded orthogonal scheme rather than an error-free ideal).\n\nThe transformer-generalization concern is secondary but real: the gain comes from MaskGIT completing masked tokens for ImageNet-100, which is in-distribution. If the source distribution or collision pattern goes outside what the model can complete, the advantage degrades. A quick out-of-distribution test or a plot versus mask ratio would tell us how robust the idea is. Also missing: error bars and code. The threshold Th is the one free parameter and seems sensible.\n\nBottom line: novel, clearly presented, and the central mechanism is demonstrated. It deserves peer review, but the authors should be pushed to align claims with evidence and to add the baseline and robustness experiments. I'd cite it as related work; whether I'd bring it to reading group depends on whether we want to discuss how to fairly benchmark generative multiple-access schemes.","headline":"ToDMA is a genuinely new token-domain multiple-access framework, but its abstract overclaims: the orthogonal baseline beats it on image quality at BER=0, and the latency advantage doesn't show up at the plotted operating point.","tokens_in":10621,"tokens_out":3145,"would_cite":true,"duration_ms":28027,"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":"Devices that collide on the same wireless channel can still recover their images via masked-token prediction.","keywords":["token communications","semantic multiple access","masked token prediction","non-orthogonal multiple access","generative semantic communications","massive random access","image transmission","ToDMA"],"falsifier":"Re-run the ToDMA receiver on images drawn from a domain the pretrained MaskGIT model has not seen, such as medical or satellite imagery, using the same VQ-GAN tokenizer; if the token error rate after masked-token prediction is no better than random selection from the candidate set, the claimed collision-mitigation gain is a property of the pretraining distribution rather than of the multiple-access scheme itself.","tokens_in":9466,"feed_emoji":"🖼️","tokens_out":8577,"duration_ms":75913,"temperature":0.7,"pith_summary":"The paper is trying to establish that multiple wireless devices can share the same time-frequency resources without coordination if the receiver is allowed to repair token collisions with a pretrained generative model. In the proposed token-domain multiple access (ToDMA), every device uses the same tokenizer and the same orthonormal modulation codebook, so the overlapped signal at the base station is a structured mixture of tokens and channel vectors. The receiver detects which tokens were active, assigns them to devices by channel similarity, and hands the leftover collided positions to a bidirectional transformer, which fills them using context and semantic orthogonality. If this works, massive uncoordinated uplink access becomes a source-completion problem rather than an orthogonal-resource-allocation problem, cutting latency while keeping image quality.","feed_headline":"Colliding wireless tokens get repaired by a transformer","feed_subtitle":"Devices share one uncoordinated uplink; a masked-token model restores the images at lower latency.","key_machinery":"The load-bearing object is the shared orthonormal modulation codebook U and the projection identity it enables: because $U^H U = I_Q$, the receiver's projection $\\widehat{H}_n = U^H Y_n$ recovers the equivalent channel matrix $H_n = \\sum_k b_{k,n} h_k^T$ plus noise. That single operation turns an overlapped multiuser signal into a per-time-slot list of token-and-channel-vector pairs, which drives the whole receiver. Token detection extracts rows of $\\widehat{H}_n$ whose energy passes a threshold; token assignment matches those channel vectors to known device CSIs; and the unassigned residual tokens become a candidate set that shrinks the transformer's prediction search from the full codebook size Q to the small collision set. The second machinery is the pretrained bidirectional transformer, which scores each candidate token against the surrounding context, exploiting what the paper calls semantic orthogonality: the idea that distinct sources' token sequences are mutually predictable and contextually separable.","core_discovery":"The central claim is that token collisions in a non-orthogonal multiple access channel are not wasted information: they can be resolved by a pretrained masked-token transformer. Concretely, with K active devices sharing an orthonormal codebook U, the received signal at slot n takes the form $Y_n = U H_n + Z_n$, so projecting onto U gives an estimate of the equivalent channel matrix $H_n$, whose nonzero rows carry both the transmitted token and the device-specific channel vector. The receiver uses these pairs to assign each token to the device with the closest CSI; when several devices transmit the same token, the leftover tokens form a candidate set and the affected positions are marked [MASK]. A pretrained bidirectional transformer, the MaskGIT-style ViT, then picks the most contextually consistent candidate for each masked position. Simulations on ImageNet-100 at SNR = 25 dB show that this pipeline keeps token error rate, PSNR, and LPIPS close to an error-free orthogonal baseline as K grows, while cutting latency by a factor that grows with the total device population. The paper concludes that ToDMA outperforms context-unaware orthogonal and non-orthogonal schemes in both image quality and latency.","pith_inferences":["Because the latency expression contains no dependence on $K_T$, the scheme's advantage should grow with the size of the device population; the paper only simulates up to $K=80$ active devices, so extrapolating to much larger K is an untested consequence.","The collision-repair gain is really a property of the source prior: ToDMA should transfer to text, audio, or video tokens only insofar as a pretrained bidirectional model can predict the collided tokens, so the fundamental limit is the source's conditional predictability rather than the channel's collision statistics.","A natural extension the paper does not test is iterative refinement: have the transformer emit confidence scores and lock in the most certain masked positions first, then re-predict the remaining ones; this could push the operating point to higher collision rates."],"forward_implications":["With ToDMA, the latency of an uplink transmission becomes independent of the total number of devices $K_T$, since the frame duration is $L N / (N_s f_s)$ rather than a share of the bandwidth divided among all devices.","Image quality stays close to an ideal orthogonal baseline even as the number of active devices K grows, because the transformer repairs the vast majority of collided tokens instead of leaving them as random errors.","The token error rate becomes a modality-agnostic performance metric, and the residual candidate set reduces the prediction complexity from the full codebook size to the number of collided tokens.","The same transmitter-receiver structure applies to any source with a pretrained tokenizer and a pretrained bidirectional token model, making the scheme a joint source-channel coding approach for generative semantic communication."],"supporting_citations":[{"why":"Supplies the pretrained MaskGIT vision transformer used to fill masked token positions.","marker":"[7]"},{"why":"Supplies the VQ-GAN tokenizer that maps each 256x256 image to 256 tokens from a 1024-entry codebook.","marker":"[8]"},{"why":"Provides the ImageNet-100 dataset on which all image-transmission simulations are run.","marker":"[39]"},{"why":"Defines the variable-rate adaptive QAM model whose rate and latency equations are used to build the Orth-Com benchmark.","marker":"[40]"},{"why":"Establishes the massive random-access scenario of many devices with sporadic activity that ToDMA targets.","marker":"[31]"}],"fun_headline_variants":["Transformer resolves token collisions in shared wireless uplink","Colliding tokens are repairable by a masked-token model","Semantic orthogonality lets devices share one channel without scheduling","ToDMA: untangling overlapping transmissions with a transformer","Token-domain MA: transformer lifts the collision bottleneck"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole scheme depends on the pretrained transformer correctly guessing the tokens lost to collisions for the actual images being sent, and the paper demonstrates this only on ImageNet-100 at one SNR and one collision regime.","fun_headline_variants_meta":{"raw":{"variants":["Transformer resolves token collisions in shared wireless uplink","Colliding tokens are repairable by a masked-token model","Semantic orthogonality lets devices share one channel without scheduling","ToDMA: untangling overlapping transmissions with a transformer","Token-domain MA: transformer lifts the collision bottleneck"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000193,"raw_usage":{"total_tokens":1352,"prompt_tokens":946,"completion_tokens":406,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":326}},"tokens_in":562,"tokens_out":406,"duration_ms":4404,"temperature":1.0,"reasoning_tokens":326,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T16:41:34.538290+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the ToDMA receiver on images drawn from a domain the pretrained MaskGIT model has not seen, such as medical or satellite imagery, using the same VQ-GAN tokenizer; if the token error rate after masked-token prediction is no better than random selection from the candidate set, the claimed collision-mitigation gain is a property of the pretraining distribution rather than of the multiple-access scheme itself.","supporting_citations":[{"cited_title":"MaskGIT: Masked generative image transformer,","cited_arxiv_id":null,"evidence_quote":"Supplies the pretrained MaskGIT vision transformer used to fill masked token positions."},{"cited_title":"Taming transformers for high- resolution image synthesis,","cited_arxiv_id":null,"evidence_quote":"Supplies the VQ-GAN tokenizer that maps each 256x256 image to 256 tokens from a 1024-entry codebook."},{"cited_title":"Imagenet large scale visual recognition chal- lenge,","cited_arxiv_id":null,"evidence_quote":"Provides the ImageNet-100 dataset on which all image-transmission simulations are run."},{"cited_title":"Variable-rate variable-power MQAM for fading channels,","cited_arxiv_id":null,"evidence_quote":"Defines the variable-rate adaptive QAM model whose rate and latency equations are used to build the Orth-Com benchmark."},{"cited_title":"Unsourced multiple access: A coding paradigm for massive random access,","cited_arxiv_id":null,"evidence_quote":"Establishes the massive random-access scenario of many devices with sporadic activity that ToDMA targets."}],"review_version":1}