{"id":"55962b2c-5676-43b8-8d20-a75ec80344ed","arxiv_id":"2508.02048","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"FedSFR lets a subset of federated clients transmit semantic feature vectors that the server reconstructs, improving stability and effectiveness of JSCC image transmission training.","lead":"Researchers propose a federated learning method, FedSFR, that lets some devices send compact semantic features instead of full model updates while training image-transmission networks. The method aims to cut communication costs and improve training stability, with reported gains in image quality and convergence.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The FedSFR gain hinges on server-side feature reconstruction being a low-error proxy for skipped local updates; the text gives no control of reconstruction error, so convergence and quality claims are unverified.","rationale":"The reader's verdict is UNVERDICTED because the full text was unreadable. I isolate the same load-bearing premise the reader identified: server-side feature reconstruction must encode enough information to substitute for skipped local model updates. The abstract promises both an algorithmic gain and a convergence-rate derivation, but it provides no condition controlling reconstruction error, no explicit theorem assumptions, and no experimental details that could be inspected. Because the central contribution is an algorithmic substitution plus a convergence proof, the single point of failure is the approximation error of the reconstruction step. If that error is bounded but not vanishing, the convergence result is at best a neighborhood convergence result, and the comparison against other FL algorithms becomes regime-dependent. I agree with the reader's weakest assumption, and I do not see a basis to move the verdict to ACCEPT or REJECT: there is no evidence of a false claim, but also no inspectable proof. The verdict should remain UNVERDICTED, so no change to the reader's verdict is needed.","tokens_in":883,"tokens_out":674,"duration_ms":47940,"concrete_test":"Obtain the full manuscript and locate the FedSFR convergence theorem. Re-derive the server aggregation update under the explicit assumption that reconstruction error is bounded by a positive constant that does not vanish with the number of rounds. If the stated rate changes from O(1/T) to O(1/T) plus a constant, the claimed convergence improvement holds only when reconstruction error vanishes. Then run the same experiment on the reported image dataset and compare the validation PSNR plateau with standard FedAvg; if FedSFR plateaus strictly worse, the reconstruction is not a faithful proxy for skipped local updates.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that transmitting feature vectors from a subset of clients and reconstructing them at the parameter server stabilizes training, improves image transmission quality, and has a mathematically derived convergence rate. For this to hold, the reconstructed features must be informationally sufficient: the aggregated update formed from them must not diverge from the update obtained from the skipped local gradients. The abstract states no necessary controls: no bound on reconstruction error, no unbiasedness assumption, no condition on the fraction of clients sending features, and no statement of how reconstruction error enters the convergence bound. In particular, if reconstruction error is bounded only by a positive constant, the standard FL convergence bound degrades from O(1/T) to O(1/T) plus a constant, meaning convergence only to a neighborhood. The full text provided to this reviewer is garbled and unreadable, so theorem statements, assumptions, proofs, and experimental tables cannot be inspected. The load-bearing premise is therefore uncheckable in the available material. This is an internal-support gap, not a disagreement with external consensus.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes FedSFR, a federated learning algorithm for image semantic communication. In each round, a subset of selected clients transmits compact semantic feature vectors instead of local model updates to the parameter server, which performs feature reconstruction. The abstract claims that this scheme improves training stability and image transmission quality compared to other algorithms, and that the authors mathematically derive its convergence rate. The abstract contains no experimental setup, dataset, baseline details, or theorem statement. The full text provided for review is an unreadable binary artifact, so none of the technical content could be inspected.","tokens_in":2057,"tokens_out":3956,"duration_ms":45757,"significance":"If the claims are substantiated, FedSFR would be a practical contribution to reducing uplink communication cost in federated semantic communication while preserving or improving convergence. The idea of server-side reconstruction of skipped client information is an interesting middle ground between full gradient transmission and model compression. However, the empirical and theoretical validation is entirely unverifiable from the provided material. The plausibility of the convergence claim hinges on an explicit, quantified bound on the reconstruction error; without such a bound the expected rate would degrade to convergence to a neighborhood rather than to the optimum.","major_comments":[{"comment":"The full-text file submitted for review is not legible; it consists of a binary artifact rather than readable text. As a consequence, the convergence theorem, its assumptions and proof, the experimental setup, result tables, and the algorithm pseudocode cannot be inspected. This is a load-bearing failure, because the abstract's central claims of \"experimental results demonstrate\" and \"mathematically derive the convergence rate\" depend entirely on content that is absent from the reviewable material. A resubmission must include a readable manuscript.","section":"Full text (as provided)"},{"comment":"The convergence-rate claim is stated without any qualifier on the reconstruction error. The abstract says that some clients send feature vectors and the server performs feature reconstruction, but it does not state whether the reconstruction error is assumed to be bounded, vanishing, unbiased, or controlled by the fraction of clients using feature transmission. For the claimed improvement over conventional FL to be theoretically grounded, the paper must state the exact convergence theorem and show how the reconstruction error enters the bound; otherwise the natural expectation is that the rate becomes O(1/T) plus a constant, which would not support the advertised \"improved performance.\"","section":"Abstract"},{"comment":"The selection rule for choosing which clients transmit features is not described. The number or fraction of feature-transmitting clients, whether the choice is random or based on channel state or data quality, and whether it changes across communication rounds, all affect both the communication savings and the convergence behavior. Without this specification, the claimed stability and effectiveness gains cannot be reproduced or compared with baselines.","section":"Abstract"}],"minor_comments":[{"comment":"The abstract would benefit from a brief definition of \"stability\" in this context: is it variance of the loss across clients, monotonicity of the training curve, or robustness to client participation patterns?","section":"Abstract"},{"comment":"The title and abstract mention \"semantic communication\" and \"feature vectors\" but do not state the type of semantic features (e.g., deep encoder outputs, learned codes) or the reconstruction loss used; one sentence clarifying the feature representation would help the reader.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The submitted manuscript file appears corrupted; if the actual submission is a normal PDF, the authors should be asked to resubmit a readable copy. The abstract alone is insufficient to evaluate the technical claims. I also note that the contribution, as described in the abstract, does not clearly distinguish itself from existing federated learning schemes with communication-efficient updates (e.g., gradient compression or local feature transmission); the authors should sharpen the novelty statement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read on 2508.02048. The abstract describes FedSFR, an FL algorithm where some selected clients send compact feature vectors to the parameter server, which reconstructs the missing local updates. That is a plausible and genuinely new combination of existing pieces for JSCC image transmission, and the claim of stabilizing FL while cutting uplink cost is worth taking seriously. The idea is coherent on its face.\n\nWhat the paper does well, as far as the abstract shows: it names a real problem (JSCC modules go stale and need FL updates), proposes a concrete mechanism (feature reconstruction at the PS), and claims both experiments and a convergence-rate derivation. The combination appears novel in this specific context. I see no red flag in the novelty claim itself.\n\nThe soft spot is exactly what the stress-test note says. Everything load-bearing depends on reconstructed features acting as a low-error proxy for the local updates they replace. The abstract does not state any bound on reconstruction error, any unbiasedness condition, or any restriction on what fraction of clients can send features. Without that, a convergence bound will either degrade to O(1/T) plus an additive constant—meaning convergence to a neighborhood—or it will rely on assumptions that need to be stated. That may be fine in a full paper; many FL analyses do exactly this. But from the abstract it is unverifiable.\n\nI also have to flag the practical problem: the full text I was given is garbled and unreadable. I cannot check the theorem statements, the assumptions, the experimental setup, or the numbers. This is not evidence of a flaw in the work, but it is a total block on verification. The reader's UNVERDICTED verdict is the right one.\n\nWho is this for: people working on JSCC image transmission and FL-based model updates. It is a subfield paper, not a foundational one. If the full text delivers the promised experiment and a convergence theorem whose assumptions are honestly stated, it is a reasonable contribution to that subfield and deserves review. I would not cite it myself until I can see the full text.\n\nRecommendation: ask the authors for a clean, readable version and then send it to a serious referee. The abstract alone clears the bar for interest, and the proposed mechanism is concrete enough to be judged on its merits. But someone needs to verify the reconstruction-error control before the stability and convergence claims can be believed.","headline":"FedSFR is a plausible new combination of feature reconstruction and federated learning for image semantic communication, but the garbled full text leaves the load-bearing reconstruction-error assumptions and experiments unverifiable.","tokens_in":2526,"tokens_out":1806,"would_cite":false,"duration_ms":23047,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes FedSFR, a federated learning algorithm for image semantic communication in which selected clients send smaller semantic feature vectors instead of local model updates and the parameter server reconstructs the missing…","keywords":["semantic communication","federated learning","joint source-channel coding","image transmission","feature reconstruction","parameter server","convergence analysis","JSCC"],"falsifier":"A decisive experiment would train FedSFR on a non-IID image dataset and compare server-side reconstructed features against the ground-truth features computed directly from client encoders; if reconstruction error grows with client heterogeneity and the converged image quality falls below standard federated averaging at the same communication budget, the central claim would be weakened.","tokens_in":1678,"feed_emoji":"📡","tokens_out":2571,"duration_ms":33644,"temperature":0.7,"pith_summary":"The paper is trying to establish that federated learning can be made more efficient and more stable for image semantic communication by replacing some client updates with compact feature vectors that a server-side module reconstructs. The motivation is practical: JSCC-based image codecs rely on a shared knowledge base that grows stale, so they need periodic retraining with distributed data. FedSFR claims that letting a subset of selected clients transmit semantic features instead of full local updates uses the limited communication budget more wisely, and that reconstructing these features at the parameter server makes training steadier and yields better images than conventional federated approaches. If true, this gives a bandwidth-friendly way to keep learned image communication systems fresh over time. The authors also derive a convergence rate to back the improvement claim.","feed_headline":"Feature reconstruction lets federated learners skip full updates","feed_subtitle":"A server-side reconstruction step keeps training stable and preserves image transmission quality.","key_machinery":"The key object is the semantic feature vector, a compact representation extracted by the JSCC encoder, which selected FL clients send to the parameter server instead of their full local update information. The parameter server integrates a feature-reconstruction module that recovers missing semantic structure from these vectors, allowing the aggregated model to benefit from participating clients without receiving their complete local gradients. This pairing of selective feature transmission with server-side reconstruction is what carries the argument: it trades a small amount of per-client information for a large reduction in communication, while the reconstruction step is what keeps the aggregated model stable.","core_discovery":"The central claim is that FedSFR, a federated learning algorithm built for training JSCC modules in image semantic communication, outperforms conventional FL by having some clients send compact semantic feature vectors rather than local update information, while the parameter server runs feature reconstruction to incorporate those clients' contributions. Even though these clients omit their gradient-based updates, the reconstructed features carry enough semantic content to keep the global model on track, reducing communication load and stabilizing the training process. The paper reports experimental gains in both stability and transmission quality and validates the approach theoretically with a convergence-rate analysis.","pith_inferences":["If the reconstruction is cheap enough, the same idea could generalize to other semantic communication modalities, such as audio or video, by replacing the image encoder features with modality-specific semantic features.","A likely design tradeoff, which the paper does not fully explore, is tuning how many clients send features versus full updates in each round; the convergence analysis suggests there should be a stable operating region that balances bandwidth savings and model accuracy.","A direct testable extension would be to measure performance under heterogeneous client data and noisy channels, checking whether the reconstructed features remain sufficient when the global data distribution differs strongly across clients."],"forward_implications":["Communication cost per round drops because selected clients send small feature vectors instead of full local updates.","The server-side feature reconstruction can absorb information from more clients within the same bandwidth budget, making federated training more stable.","Image transmission quality of the jointly trained JSCC codec improves relative to conventional federated learning baselines.","The derived convergence rate provides a theoretical guarantee that the reconstructed-feature approach does not sacrifice learning progress.","The scheme could reduce the frequency of costly full-update rounds, extending the usable lifetime of the shared knowledge base in image semantic communication systems."],"supporting_citations":[],"fun_headline_variants":["Federated learning that trades gradients for feature vectors","Server-side feature reconstruction stabilizes federated training","Send compact semantic features, not full gradient updates","Federated image comms: skip gradients, reconstruct features","Feature reconstruction lets FL clients send less and converge"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central hope is that the compact semantic features sent by the subset of clients contain enough information for the server's reconstruction to stand in for the clients' full model updates, so the shared model still converges to a good image codec.","fun_headline_variants_meta":{"raw":{"variants":["Federated learning that trades gradients for feature vectors","Server-side feature reconstruction stabilizes federated training","Send compact semantic features, not full gradient updates","Federated image comms: skip gradients, reconstruct features","Feature reconstruction lets FL clients send less and converge"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000515,"raw_usage":{"total_tokens":2426,"prompt_tokens":800,"completion_tokens":1626,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":416,"completion_tokens_details":{"reasoning_tokens":1551}},"tokens_in":416,"tokens_out":1626,"duration_ms":15498,"temperature":1.0,"reasoning_tokens":1551,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:10:58.346930+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive experiment would train FedSFR on a non-IID image dataset and compare server-side reconstructed features against the ground-truth features computed directly from client encoders; if reconstruction error grows with client heterogeneity and the converged image quality falls below standard federated averaging at the same communication budget, the central claim would be weakened.","supporting_citations":[],"review_version":1}