{"id":"111a7ca7-b026-4f66-891f-aee5b538c1ad","arxiv_id":"2506.23511","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A multi-level convolutional autoencoder for finite-blocklength AWGN channels claims comparable or better bit error rates than TurboAE-MOD and polar codes, with per-level exhaustive codebook tests and SNR-adaptive rate by layer removal.","lead":"This paper introduces a multi-level autoencoder that splits messages into 16-bit blocks, encodes each block with its own neural network, and sums the resulting signals for transmission. The design is meant to offer rate adaptation by dropping encoder/decoder layers and to allow exhaustive testing of all codewords within each block.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Per-level exhaustive testing does not validate the joint code: successive subtraction propagates earlier-level errors, and the paper offers no bound or joint test.","rationale":"The reader's conditional verdict targets exactly the right spot. The paper's contribution is defined by exhaustive codebook validation and by rate adaptation through layer removal. Both claims depend on the successive decoder behaving as a clean superposition code. But the decoder is not a linear interference canceller with genie-aided subtraction; it re-encodes its own estimates. Any mapping error at level 1 changes the residual that level 2 sees, and this distortion is dataset-dependent and not captured by testing 2^16 level-1 codewords while presumably holding other levels' contributions fixed (the paper is not explicit about what the other levels transmit during per-level tests). The phrase 'performance guarantees for the whole architecture' in Section III-A is never followed by a guarantee, an error-propagation bound, or a joint-frame simulation. The aggregated 'Total BER' in Table III is an average of per-level bit-error rates; it does not equal frame error probability and does not reveal whether errors are independent or cascading. This matters because the abstract and conclusion use exhaustive per-level testing as the basis for 'comparable, or even superior results' against polar codes and TurboAE-MOD. If error propagation is severe, the true end-to-end reliability at rates 0.75 and 1.0 could be much worse than the reported average BER suggests. The proposed check is decisive and feasible: joint random-message FER with and without genie-aided subtraction isolates the propagation mechanism. I therefore agree with the reader's conditional verdict; no change is needed. The paper can address this by adding a joint-FER experiment, a bound on error propagation, or at minimum an explicit statement that exhaustive testing covers per-level mappings only and that joint reliability is empirically evaluated on a sample.","tokens_in":7835,"tokens_out":3413,"duration_ms":35546,"concrete_test":"Run an end-to-end evaluation on random joint messages at SNR=0 dB: sample 10^5 messages uniformly from {0,1}^{16L} for L=2 and L=4, transmit through the trained MLAE, and record frame error rate (FER). Compare with the FER predicted from per-level exhaustive BERs under an independence/no-propagation assumption (1 - ∏(1 - p_i), p_i from exhaustive per-level block-error rates). In the same runs, substitute the true level-1 codeword instead of the decoded one before subtraction and measure level-2 BER. If the measured joint FER exceeds the independence prediction, or if the level-2 BER drops substantially with the true codeword, then per-level exhaustive testing alone does not validate the full multi-level code and the paper's central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that MLAE can be exhaustively validated by testing all 2^B codewords per level. Section III-A states that each level can be tested separately 'while being able to provide performance guarantees for the whole architecture as we shall discuss later,' but no such guarantee appears anywhere in the paper. With L active levels, the true codebook has 2^{BL} messages (2^64 for L=4), and the paper never tests a joint message nor bounds error propagation in the successive subtraction decoder. Decoder i subtracts the re-encoded output of decoder i-1; if that estimate is wrong, the residual for later levels contains an uncompensated interference term that is not modeled in any per-level BER measurement. Table III reports only per-level BERs averaged into a 'Total' BER, which is an arithmetic mean of bit-error rates, not a frame-error rate, and it cannot detect correlated or cascading errors. The abstract's 'exhaustively test 2^B possible codewords for each encoder/decoder level' is literally true but is used to support reliability claims for the full multi-level code, for which it is insufficient. Thus the load-bearing premise—that per-level exhaustive testing validates the complete system—is unsubstantiated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a multi-level convolutional autoencoder (MLAE) for joint channel coding and modulation over AWGN channels. Each of L levels encodes B bits into n complex symbols, and the transmit signal is the sum of per-level codewords. The receiver performs successive decoding: each decoder operates on a residual signal obtained by subtracting the re-encoded estimates of previously decoded levels. The main claimed advantages are (i) the ability to exhaustively test all 2^B per-level codewords, in contrast to sampling-based evaluation in prior learned-code work; (ii) SNR-adaptive operation by removing layers without retraining; and (iii) BER comparable to or better than TurboAE-MOD and PAC/polar codes at blocklength n=64. Results are reported as per-level and total BER at 0 dB and 2.5 dB for rates from 0.25 to 1 bit per transmission.","tokens_in":8150,"tokens_out":4098,"duration_ms":38866,"significance":"If validated, MLAE would be a useful modular learned coding scheme: per-level exhaustive testing is a genuine improvement over the small-sample evaluation typical of learned codes, and the rate-adaptation-by-layer-removal property is practically attractive. The paper also provides external benchmarks against TurboAE-MOD and PAC/polar codes, which strengthens the empirical comparison. However, the central validation claim—that per-level exhaustive testing suffices to certify the full multi-level code—is not supported, and the error-propagation behavior of the successive decoder is never analyzed. These issues must be resolved before the reliability claims can be accepted.","major_comments":[{"comment":"The text promises that per-level testing can 'provide performance guarantees for the whole architecture as we shall discuss later,' but no such guarantee or joint analysis appears in the rest of the paper. With L active levels of B bits each, the true message space has 2^{BL} elements (2^64 for the L=4 configuration), and the reported exhaustive testing covers only the 2^16 per-level subspaces. Successful per-level decoding does not imply reliable joint decoding because decoder i subtracts the re-encoded estimate of level i-1; when that estimate is wrong, the residual contains an uncompensated interference term that per-level BER measurements do not capture. The authors should either supply a bound on error propagation or report joint frame-error rates over a large random sample of full messages, and explain why exhaustive joint testing is impossible as they do for TurboAE-MOD.","section":"Section III-A"},{"comment":"The 'Total' BER in Table III is defined as the arithmetic mean of per-level BERs. This metric is not a frame-error rate and cannot detect correlated or cascading failures across levels. In a successive decoder, an error in an early level can wipe out later levels; an average of per-level BERs will systematically understate such events. I recommend reporting the joint FER for the multi-level message, or at least the conditional BER of level i given correct decoding of levels 1 through i-1.","section":"Table III and Section IV-B"},{"comment":"The training SNR is never disclosed, although the model is claimed to operate at various SNRs and Fig. 2 reports only 0 dB and 2.5 dB. The reader cannot judge whether the reported gains reflect the chosen training point or a genuine SNR-adaptive property. Please state the training SNR explicitly, report results at additional SNR values, and provide error bars or multiple-seed statistics, since the BER values span six orders of magnitude and are based on finite Monte Carlo realizations.","section":"Section IV-A"},{"comment":"The training set repeats each of the 2^16 codewords only 3 times with different noise realizations, while testing uses 2^10 realizations per codeword. The paper asserts that this 'insures the system's reliability,' but no analysis or ablation shows that three noise realizations per codeword suffice to learn decoders that generalize across the test noise distribution, particularly for the higher levels that see residual interference. At minimum, a sensitivity study on the number of training noise realizations is needed before this claim is supportable.","section":"Section IV-A"}],"minor_comments":[{"comment":"The title misspells 'Multi-Level' as 'Mutli-Level'; please correct it.","section":"Title"},{"comment":"The caption refers to 'MLEA' while the rest of the paper uses 'MLAE'; please unify the acronym.","section":"Figure 1 caption"},{"comment":"The expressions '3 × 216' and '210' should be typeset as 3 × 2^16 and 2^10 to avoid confusion; as printed they look like the decimal numbers 216 and 210.","section":"Section IV-A"},{"comment":"The abstract and conclusion state 'improved reliability' and 'comparable, or even superior performance,' but Fig. 2 shows cases where TurboAE-MOD achieves lower BER (e.g., at high rates for 2.5 dB); please qualify the wording to match the data.","section":"Abstract and Section IV"},{"comment":"The sentence 'sum them and divide by L' defines the total BER, but it is unclear why equal weighting is appropriate when levels exhibit very different BERs (e.g., Level 2 at 0.11 vs. Level 4 at 0.02); a weighted or joint metric would be more informative.","section":"Section IV-B"}],"recommendation":"major_revision","confidential_remarks":"The paper is essentially an empirical system proposal; the main blocking issue is the gap between the advertised exhaustive-validation guarantee and the actual per-level validation. If the authors can supply joint frame-error results and an error-propagation analysis, the contribution could become acceptable for the journal. No concerns about authorship or attribution; my recommendation is driven solely by the missing load-bearing evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick take: this is a reasonable incremental paper on neural channel coding. The new bit is the multi-level architecture: split a message into B-bit blocks, each with its own CNN encoder/decoder, sum the encoder outputs, and decode successively with subtraction. That's a clean idea, and the per-level exhaustive test (all 2^16 codewords per level, 2^10 noise realizations each) is a real methodological improvement over the usual random-sampling evaluation in this literature.\n\nWhat it does well: the BER curves against TurboAE-MOD and PAC codes are plausible, and the SNR-adaptive trick of dropping layers without retraining is practically interesting. The paper is honest in its conclusion that higher levels face training challenges with power allocation. The comparison to external baselines avoids circularity.\n\nWhere it's soft: the central claim of \"performance guarantees for the whole architecture\" (Section III-A) is never delivered. The per-level exhaustive test does not validate the joint code. With L levels, the message space is 2^(16L), and the successive subtraction decoder means an error at level i corrupts the residual for all later levels. The paper gives no bound on error propagation and never tests a joint message. The \"Total BER\" in Table III is just the average of per-level BERs, which can miss correlated cascading errors. On top of that: no error bars, only two SNR values, training SNR is not stated, and the comparison omits the authors' own prior CNN-AE [22], which is the most relevant baseline. These are all addressable in a revision, but right now the strongest advertised bonus—exhaustive validation—is only true per level, and the paper itself flags the missing guarantee with \"as we shall discuss later\" and then never discusses it.\n\nThe stress-test note gets it right. The reader's conditional verdict is fair, and I'd add that the missing training-SNR disclosure is a minor but annoying omission that prevents reproduction.\n\nBottom line: this deserves a serious referee, but it needs a major revision. The architecture is worth reporting, the per-level exhaustive protocol is useful, and the error-propagation question is interesting enough to justify the referee time. I wouldn't cite it in its current form, but I'd bring it to a reading group for a discussion of what \"exhaustive validation\" should mean for multi-level learned codes.\n\nRecommendation: send it to review, but insist on a bound or a joint-code test, error bars, training details, and a comparison to [22] before acceptance.","headline":"A plausible multi-level neural code with a genuinely useful per-level exhaustive test, but the promised joint-code guarantees never materialize and per-level testing does not validate error propagation.","tokens_in":8615,"tokens_out":1946,"would_cite":false,"duration_ms":19540,"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":"A multi-level autoencoder splits messages into B-bit blocks, exhaustively tests every per-level codeword, and matches or beats polar and TurboAE-MOD baselines at blocklength 64 over AWGN.","keywords":["autoencoders","channel coding","modulation","deep learning","multi-level coding","finite blocklength","successive decoding","exhaustive codebook testing"],"falsifier":"Choose a small configuration such as $B=8$, $L=4$, transmit all $2^{32}$ joint messages through the trained MLAE, decode them successively, and compare the joint frame error rate with what the per-level exhaustive tests predict; if a single early-level error cascades and makes the joint error rate substantially worse than the per-level numbers suggest, the claim that per-level exhaustive testing validates the full code collapses.","tokens_in":7636,"feed_emoji":"📡","tokens_out":9147,"duration_ms":85882,"temperature":0.7,"pith_summary":"The paper proposes a multi-level convolutional autoencoder that performs channel coding and modulation jointly. Its central move is to split each message into $B$-bit blocks, give each block its own encoder–decoder pair, and add all encoder outputs together before transmission; decoding proceeds level by level, subtracting already-decoded contributions from the received signal. Because a single level only sees $B$ bits, that level's codebook has $2^B$ entries, so every codeword of every level can be tested exhaustively instead of sampled. The authors report that this scheme matches or beats TurboAE-MOD and polar (PAC) codes at blocklength $n=64$ over AWGN channels, and that the same trained network can lower its rate by removing encoder–decoder layers, adapting to lower SNR without retraining.","feed_headline":"Tests all 65,536 codewords per level","feed_subtitle":"Multi-level learned codec matches or beats polar codes and TurboAE-MOD at blocklength 64.","key_machinery":"The load-bearing mechanism is the multi-level superposition code with successive subtraction. Each of the $L$ levels has a convolutional encoder that maps its $B$ input bits into $n$ complex symbols, and all level outputs are summed to form the transmitted signal; a global power constraint is applied during training. On the receiver side, each decoder sees a residual signal, estimates its own bits, re-encodes them, and subtracts that re-encoded estimate before the next level decodes, so the effective testable object at any stage is one level's $2^B$-entry codebook rather than the composite $2^{BL}$ codebook. The encoder and decoder are trained jointly with a weighted binary cross-entropy loss, which lets the network learn power allocation across levels as part of the optimization.","core_discovery":"The central discovery is that structuring an end-to-end learned codec as $L$ separately testable levels makes exhaustive numerical validation feasible: each decoder works on a residual left after subtracting re-encoded estimates of earlier levels, so the validation burden drops from $2^{BL}$ joint messages to $2^B$ codewords per level. Using $B=16$ and $n=64$, the paper reports BER values at 0 dB for rates 0.25, 0.5, 0.75, and 1 bits/transmission, with the total BER growing from $9.7\\times 10^{-5}$ at rate 0.25 to $4.27\\times 10^{-2}$ at rate 1. The reported comparisons with TurboAE-MOD and polar PAC codes at 0 dB and 2.5 dB show MLAE achieving comparable or better reliability in the settings tested, and the architecture is shown to change its coding rate by discarding layers without retraining.","pith_inferences":["Read strictly, the exhaustive-testing claim certifies each level separately; certifying the full transmitter would require a joint test over $2^{BL}$ messages or an error-propagation bound, which the paper does not supply.","The layer-dropping rule makes MLAE a learned rate-compatible code family; comparing it with rate-compatible polar codes at identical rates would show what the learned superposition adds over classical puncturing.","The paper notes that training becomes harder as the number of levels grows because of power allocation; this points to the successive-subtraction residual, not encoder size, as the bottleneck, suggesting an explicit per-level power schedule or an ordered curriculum might let $L$ scale further."],"forward_implications":["Because every level is tested against all $2^B$ codewords over $2^{10}$ noise realizations, the reported BER numbers are statements about the full per-level codebook, not about a small sample.","A single trained MLAE can serve a family of coding rates: removing one or more encoder–decoder layers lowers the rate and lets the same weights work at lower SNR without retraining.","At blocklength $n=64$ over AWGN, MLAE is reported to match or beat TurboAE-MOD and polar PAC codes at 0 dB and 2.5 dB in the tested settings, while being validated far more exhaustively.","Joint training with a global power constraint enables the network to distribute power across levels, which is what allows later decoders to recover their bits from successive residuals."],"supporting_citations":[{"why":"It supplies the finite-blocklength error-probability framework that motivates the short-packet coding problem.","marker":"[3]"},{"why":"It introduces the Turbo autoencoder, the end-to-end neural coding approach that MLAE builds on and compares against.","marker":"[20]"},{"why":"It introduces TurboAE-MOD, the learned joint coding and modulation baseline that MLAE is directly compared with.","marker":"[21]"},{"why":"It provides the prior convolutional autoencoder for finite-blocklength Gaussian channels that this architecture extends.","marker":"[22]"},{"why":"It supplies the theory of multilevel codes and successive decoding that the multi-level structure adapts.","marker":"[23]"},{"why":"It defines the polarization-adjusted convolutional (PAC) code baseline with list decoding used for comparison.","marker":"[25]"}],"fun_headline_variants":["Exhaustive codebook test per level in learned codec","Learned codec tests all 65,536 codes per level","Multi-level autoencoder matches polar and TurboAE-MOD","Layer pruning adapts learned codec to SNR without retraining","No retraining: learned codec prunes layers for new SNR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that checking every one of the $2^B$ codewords inside each level separately is enough to certify the whole multi-level code, even though one wrong decision at an early level changes the residual that every later decoder sees.","fun_headline_variants_meta":{"raw":{"variants":["Exhaustive codebook test per level in learned codec","Learned codec tests all 65,536 codes per level","Multi-level autoencoder matches polar and TurboAE-MOD","Layer pruning adapts learned codec to SNR without retraining","No retraining: learned codec prunes layers for new SNR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00016,"raw_usage":{"total_tokens":1237,"prompt_tokens":956,"completion_tokens":281,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":194}},"tokens_in":572,"tokens_out":281,"duration_ms":3682,"temperature":1.0,"reasoning_tokens":194,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:40:02.307735+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Choose a small configuration such as $B=8$, $L=4$, transmit all $2^{32}$ joint messages through the trained MLAE, decode them successively, and compare the joint frame error rate with what the per-level exhaustive tests predict; if a single early-level error cascades and makes the joint error rate substantially worse than the per-level numbers suggest, the claim that per-level exhaustive testing validates the full code collapses.","supporting_citations":[{"cited_title":"Turbo autoencoder: Deep learning based channel codes for point-to-point com- munication channels,","cited_arxiv_id":null,"evidence_quote":"It introduces the Turbo autoencoder, the end-to-end neural coding approach that MLAE builds on and compares against."},{"cited_title":"Joint channel coding and modulation via deep learning,","cited_arxiv_id":null,"evidence_quote":"It introduces TurboAE-MOD, the learned joint coding and modulation baseline that MLAE is directly compared with."},{"cited_title":"Coding for the gaussian channel in the finite blocklength regime using a cnn- autoencoder,","cited_arxiv_id":null,"evidence_quote":"It provides the prior convolutional autoencoder for finite-blocklength Gaussian channels that this architecture extends."},{"cited_title":"Multilevel codes: theoretical concepts and practical design rules,","cited_arxiv_id":null,"evidence_quote":"It supplies the theory of multilevel codes and successive decoding that the multi-level structure adapts."},{"cited_title":"Polarization-adjusted con- volutional (pac) codes: Sequential decoding vs list decoding,","cited_arxiv_id":null,"evidence_quote":"It defines the polarization-adjusted convolutional (PAC) code baseline with list decoding used for comparison."}],"review_version":1}