{"id":"fafb933c-e85e-4129-ae75-ad300f11caf1","arxiv_id":"2502.06095","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Rateless joint source-channel coding lets a network puncture a coded image to match channel capacity, and a new autoencoder code (RLACS) shows graceful image-quality degradation without channel state information at the source.","lead":"This paper proposes a family of joint source-channel codes that work across many rates at once, so a network can cut the coded image to fit the channel without telling the application the channel state. It also lays out a 6G design that lets slightly damaged bits reach the application, giving graceful quality loss instead of dropped packets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RLACS is trained only at F discrete code lengths; the abstract's 'continuum of rates' is unsupported, and Algorithm 2's null-padding at off-grid puncturing lengths is never seen in training.","rationale":"Reading in good faith, the operational idea — an application JSCC code that a network can puncture based on CSI, with the decoder aware of the puncturing length — is coherent and the aligned-rate experiments (Fig. 6, Fig. 7) support it for the tested modulation grid. The most load-bearing weakness is not the SVBSC abstraction: the test channels already include block fading and imperfect CSI (Sec. IV-E), and Fig. 7 shows modest degradation as ε grows, so the memoryless-BSC training mismatch is partially examined. The unexamined gap is the continuum claim itself. The formal definition requires all L in the range, but training provides only F decoders and never shows a decoder a null-padded prefix. The paper's own granularity discussion (Sec. IV-F) concedes that off-grid operation is what hurts Code 2; Code 3 only works because the grid matches the finite modulation set. Therefore the abstract's 'continuum of coding rates' overstates what is constructed and tested. This is a fixable overclaim rather than a reason to reject the paper: the blueprint and the on-grid results remain useful, and the proposed test can establish how much off-grid interpolation actually costs. Because the reader's verdict is already CONDITIONAL and our concern reinforces the need for revised claims and an additional experiment, the verdict should remain CONDITIONAL (UNCHANGED).","tokens_in":24227,"tokens_out":20216,"duration_ms":208095,"concrete_test":"Take the trained Code 3 (or retrain per Sec. IV-D) and run inference under SVBSC(qo=0.05, ε=0) at puncturing lengths K-L = 160, 192, 224, 320, 384, 448, 576, 640, 704, 832, 896, 960, 1088, 1152, 1216 — values not in {128,256,...,1280} — using Algorithm 2's null padding. Plot image PSNR versus K-L together with the on-grid trained points. If any off-grid point falls more than ~1-2 dB below the linear interpolation of its neighboring on-grid points, the RLACS code is not rateless over the claimed continuum. Repeat for Code 2 at K-L=768 and 1088 to confirm the granularity mechanism. If instead off-grid points interpolate smoothly, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central definition (Def. 2) requires the code to satisfy E[d(s,ŝ)|L]≤d[L] for every L∈{0,...,¯L}, and the abstract promises a continuum of rates. The RLACS implementation, however, trains only F discrete code lengths: iteration i optimizes decoder i for length C_i with no nulls (SVBSC(qo,ε,C_i,0), Sec. IV-A). At inference, whenever K-L falls between two trained lengths, Algorithm 2 pads the received bits with 0.5s up to the next C_i. That null-padded input distribution is never seen during training. Code 3 avoids the problem only by choosing C_i to coincide exactly with the finitely many modulation-induced rates (Sec. IV-E); Code 2's poor low-SNR performance is attributed by the authors to granularity (Sec. IV-F). Thus the 'continuum of rates' claim is not supported by the construction or by the experiments, which only test on-grid rates. The claimed optimality of Code 3 (Sec. IV-E) also rests on comparing with Code 1, another DNN, not a theoretical upper bound.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces rateless joint source-channel coding (rateless JSCC), a framework in which the application encodes a source into a binary codeword and the network adapts the coding rate by puncturing bits from the end of the codeword, without exchanging channel state information with the application. The network operation is formalized as a stabilized vector binary symmetric channel (SVBSC) that keeps the bit-flipping ratio around a target. The authors then propose a deep autoencoder realization, the RLACS code, built from a ladder of encoder/decoder pairs trained sequentially, and evaluate it on CIFAR10 image transmission over a Rician block-fading channel. They report that the rateless code (Code 3) outperforms a non-rateless DNN benchmark (Code 1) by more than 5 dB in PSNR at low SNR, and that the proposed approach degrades gracefully. The second half of the paper is a blueprint for 6G semantic communication system design, including new QoS metrics and a list of open research problems.","tokens_in":24522,"tokens_out":4476,"duration_ms":42936,"significance":"The central idea of commissioning the network with rate control through puncturing, while the application runs a channel-agnostic JSCC code, is a useful conceptual contribution that could help reconcile JSCC with existing network architectures. If the claimed performance holds, the RLACS code would provide a practical way to achieve graceful degradation without CSI at the application, and the progressive-coding interpretation (Fig. 8) is a nice demonstration. Strengths of the paper include the formal definitions of the rateless code and SVBSC, the detailed description of the sequential training procedure, the reproducible training setup, and the breadth of the system-level discussion. The framework also makes a falsifiable prediction: a code optimized over a finer set of puncturing lengths should outperform a coarser code at off-grid rates, which is testable with the proposed architecture.","major_comments":[{"comment":"The paper's abstract claims that the code is 'designed and optimized for a continuum of coding rates,' and Definition 2 requires inequality (3) to hold for every L in {0,...,Lbar}. However, the RLACS implementation trains only F discrete code lengths: training iteration i optimizes encoder/decoder i for input length C_i with no nulls (SVBSC(qo,epsilon,C_i,0)), so the trained decoders have never seen the null-padded inputs produced by Algorithm 2 when K-L falls between two consecutive C_i values. Code 3 avoids the problem only by choosing the C_i to coincide exactly with the finitely many modulation-induced rates (Sec. IV-E), and the experiments in Figs. 6 and 7 only test those on-grid rates. The construction and the evaluation therefore do not support the continuum claim; they support a claim about finitely many rates. Please either train the decoders on the actual distribution of null-padded inputs for all L, or revise the abstract and Definition 2 to describe a finite set of supported puncturing lengths.","section":"Definition 2, Sec. IV-A, Algorithm 2"},{"comment":"The reported results come from a single training run with no error bars, confidence intervals, or multiple seeds. The central empirical claims—that Code 3 outperforms Code 1 by more than 5 dB in PSNR at low SNR and that fine granularity is critical (Code 2 vs. Code 3)—are therefore not shown to be statistically robust. Given that the training procedure includes stochastic optimization and random channel realizations, I ask that the authors report mean and variance over at least several training seeds, or provide error bars on the PSNR curves.","section":"Sec. IV-E, Figs. 6 and 7"},{"comment":"The SVBSC abstraction models the channel as a set of parallel memoryless binary symmetric channels, and the RLACS decoders are trained on i.i.d. bit flips. However, the stated test channel is block-fading Rician with uncoded QAM: during a deep fade, a large fraction of the bits in a frame can be corrupted simultaneously, producing burst errors rather than independent flips. Equation (1) only bounds the probability that the average BER exceeds qo, not the spatiotemporal correlation of the errors. The paper does not test whether the decoder trained on i.i.d. flips remains robust under correlated errors. I recommend adding an experiment with burst-error channels (e.g., block-fading realizations without the averaging implied by the SVBSC model) to validate the bit-pipe abstraction.","section":"Definition 1 and Sec. IV-D"}],"minor_comments":[{"comment":"The effective rate expression 'R = (K-L)/M' appears to contain a typo: M is not defined in this context and should probably be N, the image dimension, to match the earlier definition R = K/N.","section":"Sec. III-B"},{"comment":"The statement that the negligible gap between Code 3 and Code 1 at high SNR 'empirically demonstrates optimality' is too strong: Code 1 is another DNN, not a theoretical rate-distortion upper bound. I suggest rephrasing to say that the rateless code does not sacrifice performance at the maximum rate relative to this particular benchmark.","section":"Sec. IV-E"},{"comment":"There are several typographical errors that should be corrected, e.g., 'opern' (Sec. V), 'ralteless' (Sec. IV-F), 'espectral' (Sec. IV-E), 'ratelss' (abstract), and the notation-table row 'F; P; Pr' is unclear. A careful proofread is needed.","section":"Throughout"},{"comment":"The hand-picked values qo = 0.05 and epsilon = 0.05/0.1/0.01 are acknowledged as not optimized, which is reasonable; however, the paper would be strengthened by a brief sensitivity analysis around these values, since the QoS-metric interpretation in Sec. V-B2 depends on them.","section":"Sec. IV-D"}],"recommendation":"major_revision","confidential_remarks":"The second half of the paper is a vision/blueprint essay that is not technically evaluated; the archival value of the paper hinges on the rateless JSCC results. The discrete-vs-continuum gap and the missing statistical robustness are fixable with additional training/evaluation, so I believe the paper is suitable for further consideration after a major revision. I would also note that the claimed 'continuum' in the abstract currently overstates what is actually constructed and tested."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the network-side puncturing idea is a genuinely useful contribution to the semantic-communication argument. The paper gives JSCC a way to keep graceful degradation while preserving a binary application-to-network interface, which is the practical blocker in that field. The sequential training scheme that makes latent bits monotonically important is a clean workaround to the training instability of stochastic TailDrop. The on-grid results (Code 3 vs Code 1, more than 5 dB at low SNR) are believable as a directional demonstration.\n\nWhere it earns credit: the SVBSC abstraction, as a stabilized bit pipe with a BER and puncturing length, is a clean division of labor between network and application. The blueprint section is qualitative but honest about what would need to change in real radios and transport protocols (UDP-Lite, BER-based QoS, multi-hop issues). I also appreciate the explicit admission in Sec. IV-F that granularity matters; Code 2's low-SNR gap is not hidden.\n\nThe main soft spot is the 'continuum of rates' claim. Definition 2 and the abstract promise every L in the range, but the RLACS training only sees F discrete lengths. At inference, Algorithm 2 pads off-grid cases with 0.5s up to the next trained C_i, a distribution the decoder never saw in training. Code 3 sidesteps this by aligning C_i with the finite set of modulation-induced rates. So what the paper actually demonstrates is a rateless code over a finite, pre-agreed set of rates; the continuum language overstates it. That should be fixed by either restricting the claim or showing the padding works.\n\nThe second real concern is the SVBSC i.i.d.-flip assumption under block fading. A deep fade corrupts most bits in a frame simultaneously. The scheduler bounds the probability that the average BER exceeds qo, but it does not make the per-frame errors i.i.d. The decoder is trained on i.i.d. flips, and the burst-correlated case is not tested. This may or may not be fatal in practice, but the paper should at least address it. Also, no error bars or multiple seeds for the PSNR curves makes the 5 dB headline number fragile.\n\nOverall, the paper is a solid design blueprint with a plausible first demonstration. It deserves a serious referee. I'd read it as a system proposal, not as an information-theoretic result; the optimality claims are not derived. If the authors add code, error bars, a burst-error channel test, and scale back the continuum wording, I'd be comfortable citing it.","headline":"A useful architectural idea for JSCC over a binary network interface, but the continuum claim is not backed by the construction.","tokens_in":25025,"tokens_out":3717,"would_cite":true,"duration_ms":36691,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A24","94A29","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Rateless joint source-channel coding hands the network the job of adapting the transmission rate, so the application never needs channel state information.","keywords":["rateless joint source-channel coding","semantic communications","rate-adaptive link","stabilized bit pipe","SVBSC","RLACS code","autoencoder JSCC","6G semantic networking"],"falsifier":"Take the trained RLACS Code 3 and test it over a channel with the same average bit-flip rate $q_0=0.05$ but with errors generated in correlated bursts, such as a Gilbert-Elliott model or a block-fading channel with the same marginal BER; if the PSNR at low SNR drops materially relative to the i.i.d. SVBSC training channel, then the SVBSC abstraction's sufficiency for RLACS is refuted.","tokens_in":24016,"feed_emoji":"📡","tokens_out":7650,"duration_ms":66794,"temperature":0.7,"pith_summary":"This paper attempts to resolve the main practical obstacle to joint source-channel coding: the application usually does not know the channel state, while the network does. The proposed fix is a rateless JSCC code that is optimized for a continuum of coding rates, together with a network-side operation that punctures bits out of the codeword to fit the channel capacity and stabilizes the bit-flipping ratio around a target value. From the application's point of view the link becomes a controlled bit pipe, and the network absorbs all rate adaptation. The paper demonstrates the idea with an autoencoder code family called RLACS and reports that, on CIFAR10 image reconstruction, the fine-grained rateless code outperforms a non-rateless benchmark by more than 5 dB in PSNR at low SNR while degrading gracefully as the channel worsens. If successful, the framework offers a way to realize semantic communication services over conventional networks without redesigning the network from scratch.","feed_headline":"Rateless code beats fixed JSCC by 5 dB at low SNR","feed_subtitle":"RLACS coding lets the network adjust the rate by puncturing bits, so the application never needs channel state information.","key_machinery":"The central object is the pair formed by the rateless JSCC code definition and the SVBSC link abstraction, with RLACS as the concrete realization. A rateless code is defined as a tuple $(f,r,g)$ such that for every puncturing length $L$ in $\\{0,\\dots,\\bar L\\}$ the expected distortion stays at or below a promised value $d[L]$; the network chooses the modulation order $M$ through Eq. (1) so that the probability the bit-flip ratio exceeds $q_0$ is at most $\\varepsilon$. The RLACS machinery is an autoencoder ladder trained sequentially: at iteration $i$, all previous encoder-decoder pairs are frozen and only pair $i$ is optimized against the reconstruction loss over a training channel with $q_0$ and no instability, while decoder $i$ consumes the output of all pairs $1,\\dots,i$. The residual attention architecture provides the learned feature hierarchy, and a VIMCO estimator supplies gradients for the binary latent space.","core_discovery":"On the paper's own terms, the central claim is that the performance advantage of joint source-channel coding can be obtained while preserving the modular separation of application and network: the application only needs to provide a code that maintains a target distortion for every puncturing length the network might choose, and the network only needs to turn the wireless channel into a stabilized vector binary symmetric channel with bit-flip probability around $q_0$ and stability parameter $\\varepsilon$. The formal definition is a $(K,N,\\bar L,d)$ rateless code satisfying a distortion bound for every $L$, with null bits inserted at the decoder where the network punctured. The RLACS implementation realizes this with an autoencoder ladder in which earlier trained modules are frozen while a new module is trained for each additional segment of code length; the first bits therefore learn to carry the most reconstruction-critical information. The reported experiments show Code 3, whose segment sizes match the modulation grid, staying near the non-rateless Code 1 at high SNR and beating it by more than 5 dB in PSNR at low SNR, with graceful rather than cliff-like degradation as SNR falls.","pith_inferences":["The paper trains RLACS on i.i.d. bit flips; a likely untested limit is that correlated burst errors from deep fades violate the SVBSC independence premise, and decoding quality may drop more than the paper's $\\varepsilon$ analysis suggests.","The same bit-pipe abstraction should extend to multi-hop links: each hop could puncture up to a pre-agreed limit, with the bottleneck hop setting the end-to-end rate; this is implied by the blueprint but not experimentally tested.","Because rateless training assigns higher importance to early bits, the same architecture could be trained with semantic or task-specific losses, not only MSE, making it a candidate for effectiveness-oriented 6G services.","A hybrid that uses classic LDPC or Polar codes inside the network to stabilize the BER before RLACS decoding could combine the strengths of both; the paper lists this as future work, so it remains an open testable direction."],"forward_implications":["Applications can be made channel-agnostic: the network performs rate adaptation by puncturing, so no CSI routing back to the application is needed.","Lossy media should escape the cliff and leveling-off effects of packet-ized, error-free networking, because damaged and punctured bits become controlled erasures and errors rather than whole packet drops.","Finely matched puncturing granularity matters: Code 3, whose segment sizes align with the modulation constellation grid, clearly beats the coarser Code 2 at low SNR.","The code family is extensible: adding modules for higher modulation orders requires no retraining of existing modules, which suits incremental deployment.","The link abstraction introduces new QoS metrics, bit-flip ratio, stability $1-\\varepsilon$, maximum puncturing $\\bar L$, and average puncturing, that replace or supplement block-error-rate targets."],"supporting_citations":[{"why":"Provides the DeepJSCC wireless image transmission baseline that RLACS is compared against, setting the context of the reported PSNR gap.","marker":"[35]"},{"why":"Supplies the multi-level reliability interface baseline and Figure 10 comparison, and motivates unequal-importance bit protection.","marker":"[32]"},{"why":"Underpins the cliff-effect and leveling-off discussion that motivates replacing error-free packet delivery with graceful degradation.","marker":"[17]"},{"why":"Gives the finite-blocklength suboptimality of separate source and channel coding, justifying joint coding in the lossy regime.","marker":"[21]"},{"why":"Introduces the stochastic-bottleneck rateless autoencoder whose random dropout the paper avoids in favor of sequential training.","marker":"[40]"},{"why":"Provides the VIMCO estimator used to train the binary latent space of the RLACS encoders.","marker":"[44]"},{"why":"Supplies the residual attention network architecture used as the building block of RLACS encoder and decoder modules.","marker":"[43]"},{"why":"Represents the earlier Raptor-code-based rateless hierarchical JSCC that the paper differentiates from true single-rate-matching rateless JSCC.","marker":"[25]"}],"fun_headline_variants":["Rateless JSCC decouples app from channel, beats fixed code by 5 dB","Puncture to adapt: rateless code wins 5 dB at low SNR","RLACS: rateless autoencoder code with graceful low-SNR gains","Modular rateless coding: app needs no channel state info","Blueprint for 6G semantic nets via stable bit-pipe links"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the end-to-end bit channel is well described as a memoryless vector binary symmetric channel whose bit flips are independent with probability around $q_0$; under the stated Rician block-fading channel with uncoded QAM, a deep fade can corrupt many bits in a frame at once, so the training and the graceful-degradation claims rest on an independence assumption that Eq. (1) does not directly guarantee.","fun_headline_variants_meta":{"raw":{"variants":["Rateless JSCC decouples app from channel, beats fixed code by 5 dB","Puncture to adapt: rateless code wins 5 dB at low SNR","RLACS: rateless autoencoder code with graceful low-SNR gains","Modular rateless coding: app needs no channel state info","Blueprint for 6G semantic nets via stable bit-pipe links"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000703,"raw_usage":{"total_tokens":3227,"prompt_tokens":1059,"completion_tokens":2168,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":675,"completion_tokens_details":{"reasoning_tokens":2070}},"tokens_in":675,"tokens_out":2168,"duration_ms":16325,"temperature":1.0,"reasoning_tokens":2070,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T16:47:30.459141+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained RLACS Code 3 and test it over a channel with the same average bit-flip rate $q_0=0.05$ but with errors generated in correlated bursts, such as a Gilbert-Elliott model or a block-fading channel with the same marginal BER; if the PSNR at low SNR drops materially relative to the i.i.d. SVBSC training channel, then the SVBSC abstraction's sufficiency for RLACS is refuted.","supporting_citations":[{"cited_title":"Deep j oint source-channel coding for wireless image transmissio n,","cited_arxiv_id":null,"evidence_quote":"Provides the DeepJSCC wireless image transmission baseline that RLACS is compared against, setting the context of the reported PSNR gap."},{"cited_title":"Multi-level reliability interface for semantic communi cations over wireless networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-level reliability interface baseline and Figure 10 comparison, and motivates unequal-importance bit protection."},{"cited_title":"Joint source–channel coding: Fundamentals and recent progress in practical designs,","cited_arxiv_id":null,"evidence_quote":"Underpins the cliff-effect and leveling-off discussion that motivates replacing error-free packet delivery with graceful degradation."},{"cited_title":"Lossy joint source-channel c oding in the finite blocklength regime,","cited_arxiv_id":null,"evidence_quote":"Gives the finite-blocklength suboptimality of separate source and channel coding, justifying joint coding in the lossy regime."},{"cited_title":"Stochastic bottleneck: Ra teless auto-encoder for flexible dimensionality reduction ,","cited_arxiv_id":null,"evidence_quote":"Introduces the stochastic-bottleneck rateless autoencoder whose random dropout the paper avoids in favor of sequential training."},{"cited_title":"V ariational inference for Mo nte Carlo objectives,","cited_arxiv_id":null,"evidence_quote":"Provides the VIMCO estimator used to train the binary latent space of the RLACS encoders."},{"cited_title":"Residual attention network for image classification,","cited_arxiv_id":null,"evidence_quote":"Supplies the residual attention network architecture used as the building block of RLACS encoder and decoder modules."},{"cited_title":"L ossy joint source-channel coding using Raptor codes,","cited_arxiv_id":null,"evidence_quote":"Represents the earlier Raptor-code-based rateless hierarchical JSCC that the paper differentiates from true single-rate-matching rateless JSCC."}],"review_version":1}