{"id":"c2c60aa1-62cc-4722-9de8-998a106278d6","arxiv_id":"2508.07958","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":3,"one_line_summary":"An adaptive, separately coded scheme with logistic-regression distortion surrogates and convex rate/power optimization is claimed to outperform deep JSCC and SSCC over parallel Gaussian channels.","lead":"This paper proposes an adaptive source-channel coding scheme for semantic communications that keeps neural semantic source coding separate from conventional digital channel coding while tuning rates and power per channel. The authors claim it beats both deep joint source-channel coding and separate coding baselines in simulation, while staying compatible with existing digital systems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Supplied full text is a different paper (arXiv:2508.07956), so the ASCC abstract's technical claims cannot be checked; the load-bearing surrogate-fidelity assumption remains unverified.","rationale":"The reader's UNVERDICTED verdict is appropriate: the strongest claim cannot be assessed because the supplied full text is an unrelated paper. The reader's identified weakest assumption, the fidelity of the logistic-regression surrogates, is indeed the main technical fragility visible in the abstract. However, the immediate blocker is the body/abstract mismatch, which prevents any substantive technical check. I partially agree with the reader: the surrogate-fidelity concern is real but secondary to the unverifiability caused by the incorrect full text. No technical conclusion about the correctness of the ASCC scheme can be drawn from the available material, so the verdict should remain unchanged.","tokens_in":12250,"tokens_out":2432,"duration_ms":29993,"concrete_test":"Obtain the actual full text of arXiv:2508.07958 and perform an independent replication of the surrogate-based optimization on a held-out set of E2E distortion measurements over parallel Gaussian channels. Specifically, compare the true distortion achieved by the SCA-optimized rates/powers against both deep JSCC and SSCC baselines; if the logistic surrogates mispredict distortion by more than the reported margins, the central outperformance claim fails. Also verify that the supplied full text matches the target paper.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that ASCC outperforms deep JSCC and SSCC by fitting logistic-regression surrogates for E2E distortion (as a function of source coding rate and BER) and BER (as a function of SNR and channel coding rate), then solving a weighted-sum distortion minimization via SCA. This claim depends critically on the surrogates being accurate enough that the optimized rates/powers are near-optimal for the true system. In the finite-blocklength regime over parallel Gaussian channels, distortion can depend on source statistics, codec architecture, per-channel error correlation, and power allocation coupling; the abstract provides no error analysis, no validation of the logistic fits, and no comparison details. Moreover, the supplied full text is arXiv:2508.07956 (a RAG/WebFilter paper), not the ASCC paper, so none of the derivations, simulations, or baseline implementations can be inspected. This is not an internal inconsistency, but it makes the outperformance claim unverifiable from the provided materials.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript as submitted consists of an abstract for a paper titled 'Adaptive Source-Channel Coding for Semantic Communications' (arXiv:2508.07958) and a full-text body that is actually a different paper, 'Careful Queries, Credible Results: Teaching RAG Models Advanced Web Search Tools with Reinforcement Learning' (arXiv:2508.07956). The abstract proposes ASCC: DNN-based semantic source coding is kept separate from conventional digital channel coding, and source/channel coding rates and powers are allocated over parallel Gaussian channels by fitting logistic-regression surrogates for E2E distortion and BER and then solving a weighted-sum distortion minimization via successive convex approximation (SCA). The abstract claims simulations show outperformance over deep JSCC and SSCC in single- and parallel-channel settings.","tokens_in":12334,"tokens_out":4385,"duration_ms":52693,"significance":"Conditional on the claims being true, the practical compatibility with existing digital systems and the explicit treatment of rate and power adaptation over parallel channels would be a useful contribution to the semantic-communications literature. The paper does not ship reproducible code, machine-checked proofs, or a parameter-free derivation; all support is deferred to simulation results that are not described in the submitted materials. Because the supplied full text is unrelated to the abstract, the significance cannot presently be evaluated beyond this conditional statement.","major_comments":[{"comment":"The submitted full text is a different paper (the WebFilter/RAG paper), not the ASCC manuscript. Consequently none of the load-bearing elements can be inspected: problem formulation, logistic-regression fitting and validation, SCA derivation/convergence, simulation setup, baseline implementations, or numerical results. The final-sentence claim that ASCC 'outperforms typical deep JSCC and SSCC schemes' is therefore unsupported in the submitted materials.","section":"Full text (arXiv:2508.07956) vs Abstract"},{"comment":"E2E data and semantic distortions are modeled as functions of source coding rate and BER, and BER as a function of SNR and channel coding rate. In finite-blocklength parallel Gaussian channels, distortion can also depend on source statistics, codec architecture, per-channel error correlation, and the coupling between power allocation and error probability. The abstract reports no validation of the logistic fits, no error analysis, and no sensitivity study. If the fits are not accurate, the SCA solution minimizes the wrong objective and the claimed gains over deep JSCC/SSCC would not transfer to the true system. This is the central load-bearing assumption of the paper.","section":"Abstract: surrogate models"},{"comment":"The weighted-sum E2E distortion minimization is stated without specifying the weights, the constraint set, or convergence guarantees for SCA. Even if SCA converges, it yields a stationary point of the surrogate problem; without a bound on surrogate error or a comparison of the surrogate-optimal solution against the true E2E distortion, the optimality/outperformance claim is not established.","section":"Abstract: SCA objective"}],"minor_comments":[{"comment":"The baselines 'typical deep JSCC and SSCC schemes' are not named; please identify them so the comparison is reproducible.","section":"Abstract"},{"comment":"Please define the 'parallel Gaussian channels' model (independent vs correlated, blocklength, CSI availability) in the abstract or introduction.","section":"Abstract"}],"recommendation":"uncertain","confidential_remarks":"The supplied full text does not match the abstract; this looks like a submission/review-system error rather than a technical flaw in the ASCC work. I would recommend returning the manuscript to the authors before assigning a technical recommendation, or treating the current submission as incomplete."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe supplied package has a serious discrepancy: the abstract is about ASCC for semantic communications over parallel Gaussian channels, but the attached full text is a different paper on RAG web-search training. I can only assess the abstract, and based on that, the scheme is a reasonable engineering extension: DNN semantic source coding separate from conventional digital channel coding, with logistic-regression surrogates for distortion and BER, then rate/power allocation via SCA. The deployment-compatibility goal is legitimate, and the problem formulation is sensible.\n\nCredit where due: the abstract is clear, the architecture is distinct from end-to-end deep JSCC, and the choice of a weighted-sum E2E distortion objective is a natural way to trade off data and semantic quality. If the surrogates are accurate, the approach could be practically useful.\n\nThe soft spots are real but I can't measure them without the actual paper. The load-bearing assumption is that distortion depends only on source coding rate and BER, and BER only on SNR and channel coding rate. In the finite blocklength regime, per-channel error correlation, source statistics, and the codec architecture can all matter. The abstract gives no validation of the surrogate fits, no error bars, and no dataset description. That makes the outperformance claim over deep JSCC and SSCC unverified from what I received. Also, the abstract cites no prior work, so the novelty claim is uncheckable.\n\nThe full-text mismatch is the dominant problem. It blocks any substantive review. Whether it's an arXiv pipeline error or something else, the editor should not desk reject the abstract on the merits, but should request the correct manuscript before sending out. If the real paper delivers what the abstract promises, a knowledgeable referee could check the surrogate validation and the baseline fairness.\n\nBottom line: this idea deserves a look, but not on the current package. I would not cite it until I can read the actual simulation and math. Bring it to the reading group if a corrected manuscript shows up.\n\nRecommendation: send to peer review after requiring the correct full text.","headline":"Plausible semantic-comm scheme in the abstract, but the supplied full text is a different paper, so the technical claims are uncheckable from this package.","tokens_in":12981,"tokens_out":3137,"would_cite":false,"duration_ms":34436,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A15","94A29"],"pacs":[],"model":"deepseek-v4-flash","headline":"Adaptive semantic source-channel coding outperforms deep JSCC and SSCC while remaining compatible with standard digital systems.","keywords":["semantic communications","joint source-channel coding","separate source-channel coding","adaptive rate allocation","parallel Gaussian channels","successive convex approximation","logistic regression","end-to-end distortion"],"falsifier":"Take a deployed semantic codec and a set of parallel Gaussian subchannels at fixed SNRs, compute the true end-to-end data and semantic distortion for a grid of source rates, channel rates, and power splits, and compare against the logistic-regression predictions. The paper's claim would fail if two allocations whose predicted distortion ordering is clear reverse under measured distortion, or if the surrogate's residuals are large exactly in the finite-blocklength regime that motivates the work.","tokens_in":11984,"feed_emoji":"📡","tokens_out":5566,"duration_ms":62064,"temperature":0.7,"pith_summary":"Semantic communication normally pits two designs against each other: deep joint source-channel coding (JSCC) treats the whole link as one learned encoder/decoder and achieves low end-to-end distortion but is tied to a specific codec and channel model, while separate source-channel coding (SSCC) is standard-compatible but loses performance in short blocks. This paper argues for a middle path. It proposes adaptive source-channel coding (ASCC), in which a DNN-based semantic source codec is kept separate from conventional digital channel coding, and the source rate, channel coding rate, and transmit power are adaptively chosen per parallel Gaussian subchannel. The adaptation is made tractable by regression surrogates: end-to-end data and semantic distortion are modeled as functions of source rate and bit error ratio, and bit error ratio as a function of signal-to-noise ratio and channel rate. The paper then solves the weighted sum of data and semantic distortions by successive convex approximation and reports that ASCC outperforms typical deep JSCC and SSCC schemes in single- and parallel-channel scenarios while remaining compatible with existing digital systems.","feed_headline":"Adaptive semantic coding beats deep JSCC while staying digital","feed_subtitle":"Neural source coding plus standard channel coding, with rates and power tuned per channel, closes the gap to joint end-to-end designs.","key_machinery":"The load-bearing object is a pair of fitted logistic-regression surrogates. One maps the source coding rate $R_s$ and the bit error ratio $p_b$ to the end-to-end data distortion $D_d$ and semantic distortion $D_s$; the other maps the signal-to-noise ratio and the channel coding rate $R_c$ to the bit error ratio $p_b$. These surrogates turn the coupled source/channel rate and power allocation into a weighted-sum distortion minimization over parallel Gaussian channels, which the paper solves with successive convex approximation. The surrogates carry the adaptation: every time the channel or the source changes, the optimizer re-solves the surrogate problem to pick new rates and powers without r","core_discovery":"On its own terms, the paper's central claim is that the semantic-coding advantage does not require an end-to-end learned transceiver. The authors keep the DNN semantic compressor as a separate source code, pair it with an off-the-shelf digital channel code, and formulate the transmission problem as weighted end-to-end distortion minimization over parallel Gaussian channels, with the source coding rate, channel coding rate, and power per subchannel as variables. Because the true distortion as a function of these variables is hard to optimize, they fit logistic-regression models to approximate data and semantic distortion as functions of source rate and BER, and BER as a function of SNR and ch","pith_inferences":["A natural next test is whether the same two-level regression recipe transfers to other semantic codecs and task metrics such as task success rate or retrieval precision; the paper's architecture does not depend on the specific codec, so transferability is plausible but not demonstrated.","Because the surrogate is trained once per codec, an implicit deployment recipe emerges: collect distortion and BER data, fit the regressions, then reuse the SCA solver at runtime, moving adaptation cost from retraining the DNN to re-solving a small optimization.","The comparison to deep JSCC may be sensitive to the semantic codec's operating range and blocklength; a codec with sharper distortion-rate behavior could violate the logistic surrogate's monotone shape and change the ordering of allocations.","The parallel-channel formulation could connect to classical water-filling intuition: if the surrogate simplifies appropriately, the optimal power and rate allocation might approximate rate-water-filling over subchannel SNRs, giving a closed-form benchmark to test against the SCA solution."],"forward_implications":["The scheme can run on a standard digital physical layer; only the source codec needs to be replaceable, so operational systems could adopt semantic gains without new waveforms.","Because adaptation is per parallel subchannel, the method gives a concrete way to allocate rate and power across frequency or time slots according to measured SNRs.","The weighted-sum objective lets the sender trade bit-accurate data fidelity against task-oriented semantic fidelity by choosing weights, rather than committing to one distortion measure.","If the fitted surrogates are accurate, SCA gives a reproducible offline allocation policy that does not require backpropagation through the channel at deployment.","The claimed gains over deep JSCC are established for the single- and parallel-channel scenarios tested, and those same scenarios define the scope of the comparison."],"supporting_citations":[],"fun_headline_variants":["Adaptive semcom: separate coding with tuned rates and power beats deep JSCC","Digital-friendly semantic coding: rate-power adaptation outshines joint design","Semantic source coding + standard channel coding: adaptive rates beat end-to-end JSCC","Adaptive rate-power tuning makes separate semantic coding beat deep JSCC"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The central assumption is that the fitted logistic-regression surfaces—distortion depending only on source rate and BER, and BER only on SNR and channel rate—are faithful enough that solving them minimizes the true end-to-end distortion of the real codec; if finite-blocklength distortion depends on source statistics, codec architecture, or error correlation across subchannels, the optimized allocation can be wrong even though the surrogate is minimized.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive semcom: separate coding with tuned rates and power beats deep JSCC","Digital-friendly semantic coding: rate-power adaptation outshines joint design","Semantic source coding + standard channel coding: adaptive rates beat end-to-end JSCC","Adaptive rate-power tuning makes separate semantic coding beat deep JSCC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000885,"raw_usage":{"total_tokens":3673,"prompt_tokens":776,"completion_tokens":2897,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":2815}},"tokens_in":520,"tokens_out":2897,"duration_ms":24276,"temperature":1.0,"reasoning_tokens":2815,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T21:44:36.649544+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a deployed semantic codec and a set of parallel Gaussian subchannels at fixed SNRs, compute the true end-to-end data and semantic distortion for a grid of source rates, channel rates, and power splits, and compare against the logistic-regression predictions. The paper's claim would fail if two allocations whose predicted distortion ordering is clear reverse under measured distortion, or if the surrogate's residuals are large exactly in the finite-blocklength regime that motivates the work.","supporting_citations":[],"review_version":1}