{"id":"0a400989-8fa1-480a-aceb-ca6eb3d7d0e5","arxiv_id":"2505.08566","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A vision-based large AI model, used offline, refines codebook codewords for CSI feedback in FDD massive MIMO, improving reconstruction accuracy and sum rate without online inference latency.","lead":"This paper proposes offline frameworks that use a large vision model, previously trained on images, to generate improved codebooks for CSI feedback in FDD massive MIMO, moving the heavy computation out of the real-time path. The online feedback remains a simple codebook lookup, and simulations report higher reconstruction accuracy and sum rate than existing codebook-based schemes.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed transfer from image pretraining is not isolated: no same-architecture randomly initialized baseline is run, so the gains in Figs. 6-9 may come from scale or fine-tuning rather than vision transfer.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: the paper never isolates the contribution of image pretraining from model capacity. I agree that this is the key soft spot in the central claim. The paper's own framing in Section IV-B asserts a structural analogy between images and dual-polarized CSI codewords and credits this analogy for the transfer, but the experiments compare LVM4CF only against smaller or differently pretrained architectures. A same-architecture randomly initialized control would settle whether the gain is due to vision pretraining or simply to having a larger, well-tuned model. Because the reader already marked the paper CONDITIONAL on this and related issues, my read does not change the verdict. No additional independent objection rises to the same level: the offline-latency claim is structurally sound, and the missing error bars, code, and data would be addressed by the conditional requirements rather than by a different verdict.","tokens_in":17097,"tokens_out":7897,"duration_ms":90863,"concrete_test":"Train the exact LVM4CF architecture (D=4096, L=2, same CSI preprocessor/output heads) from random initialization on the same UMa and RMa datasets used for SSLCF and MSLCF, with identical epochs, loss, optimizer, and update schedules, and compare cosine similarity and sum rate against the pretrained-initialized version reported in Figs. 5-9. If the random-init version closes the gap, the vision-pretraining claim fails; if the pretrained version retains a clear advantage across B=12/16/20 and both scenarios, the transfer claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism asserted in Section IV-B is that a vision LVM pretrained on images transfers spatial-correlation modeling to CSI codeword refinement. The only comparisons are against FC, a lightweight Transformer, and GPT2-XL. Those baselines differ in architecture and parameter count as well as pretraining, so they cannot separate the effect of image pretraining from model capacity or from the newly added CSI-specific embedding and output layers. LVM4CF is never compared with a randomly initialized version of the same architecture trained on the same CSI data, nor with the same architecture initialized from the image-pretrained checkpoint but with the CSI embedding/output heads reinitialized. Without such an ablation, the phrase 'the pre-trained LVM to transfer its spatial modeling capability' (Section IV-B) is unsupported; the results in Figs. 6, 7, and 8(c) are also consistent with the simpler hypothesis that a larger model fine-tuned on CSI data achieves better codebook refinement regardless of the source of its initial weights. This matters because the paper's stated design methodology, and its claimed advantage over the LLM baseline, relies on image-domain transfer rather than on scale alone.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two offline frameworks, SSLCF and MSLCF, for FDD massive MIMO CSI feedback. Instead of running a large model in real time, a large vision model called LVM4CF (a LLaMA-style backbone pre-trained on images and fine-tuned on CSI data) is used offline to refine each codeword of a conventional RVQ codebook into an environment-aware codebook. In the online stage the UE only feeds back the index of the best codeword and the BS looks it up, so no large-model inference is incurred. SSLCF produces site-specific codebooks, while MSLCF pre-generates a set of environment-specific codebooks and selects one with a low-overhead environment index. The paper details the LVM4CF architecture, two training and inference algorithms (single-side and dual-side), and simulations using QuaDRiGa channels showing cosine similarity and sum-rate gains over conventional codebook, FC, lightweight Transformer, and a GPT2-XL-based LLM baseline.","tokens_in":17273,"tokens_out":7531,"duration_ms":75509,"significance":"The offline-use idea is attractive: it reconciles the expressive power of large pretrained models with the latency constraints of wireless deployments, and the two deployment modes are clearly motivated. The paper is generally well structured, gives explicit algorithms, and evaluates on a standard simulator with open-source pretrained weights, which aids reproducibility. If the vision-transfer mechanism were substantiated, the result would be a useful design principle for codebook-based feedback. However, the current experiments do not isolate the effect of image pretraining from model capacity, and the dual-side training algorithm contains an ambiguity about parameter reinitialization; these issues need to be resolved before the central claim can be accepted.","major_comments":[{"comment":"The central claim that image-domain pretraining transfers to CSI codeword refinement is not isolated by the experiments. LVM4CF is compared only against FC, a lightweight Transformer, and GPT2-XL, which differ in architecture, depth, parameter count, and pretraining data. Because no same-architecture randomly initialized baseline and no variant with only the CSI-specific embedding and output layers reinitialized are reported, the gains shown in Figs. 6, 7, 8(c), and 9 are also consistent with the simpler hypothesis that a larger model fine-tuned on CSI data yields better codebooks regardless of initialization. Since Section IV-B explicitly attributes the improvement to the structural analogy between images and CSI, an ablation separating pretrained initialization from model capacity is load-bearing and should be added.","section":"Section IV-B, Figs. 6-9"},{"comment":"The instruction 'Reinitialize the trainable parameters in LVM' is ambiguous and potentially undermines the transfer claim in dual-side deployment. If all trainable parameters are reset at each codebook update, the image-pretrained initialization is discarded whenever the codebook is refreshed, which contradicts the premise that pretrained knowledge is being exploited. If only the added heads are reset, the paper should say so explicitly. The surrounding text also says the update mechanism 'guarantees' consistent improvement, but the algorithm only accepts a codebook when the validation loss decreases; that is an empirical selection rule, not a guaranteed monotonic improvement.","section":"Algorithm 2, line 8"},{"comment":"The comparison is restricted to codebook-refinement approaches; no end-to-end DL CSI feedback method such as CsiNet [15], CsiNet+ [19], or TransNet [22] is included. The abstract and introduction claim that the frameworks significantly outperform existing schemes, but the experiments only support superiority over the listed codebook-based baselines. Either add at least one representative end-to-end method under the same channel data and feedback-bit budget, or explicitly limit the claim to codebook-based feedback frameworks.","section":"Section V, baseline set"},{"comment":"No error bars, multiple seeds, or confidence intervals are reported. The conclusion that LVM4CF consistently outperforms the LLM baseline, particularly the vision-versus-text comparison in Figs. 8(c) and 9, would be more convincing with at least three independent runs per configuration and a statement of variance.","section":"Section V, all figures"}],"minor_comments":[{"comment":"References [4] and [33] are assigned the same arXiv identifier (2406.09022), which appears to be an error; the LLM-CSI-feedback reference should be corrected.","section":"References [4] and [33]"},{"comment":"The LLM baseline is described as a '12-layer GPT2-XL,' but GPT-2 XL has 48 layers; if a 12-layer GPT-2 variant was used, it should be named and configured accordingly.","section":"Section V, LLM baseline"},{"comment":"Typos such as 'multiple-intput' in the title and abstract, 'practicale propagation environments' in Section I, 'choose' for 'chosen' in Section V, and 'thei-th toj-th' in the Notation paragraph should be corrected.","section":"Title, Section I, Section V, Notation"},{"comment":"Please clarify whether the reported running time includes only the online codebook lookup and index feedback or also the offline codebook-generation time; the 'no added latency' claim depends on this distinction.","section":"Fig. 8(b)"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: a credible engineering paper that scales offline codebook optimization to a pretrained vision Transformer, with a useful deployment taxonomy. The headline claim that image pretraining is what drives the gain, though, is not actually tested.\n\nWhat's new and good: the offline codebook refinement idea from [34] is upgraded to a large vision model and given a clean treatment. The SS/DS deployment distinction (UE-side RVQ kept in SS, shared enhanced codebook in DS) and the site-specific vs. multi-scenario split are sensible and well explained. Algorithms 1 and 2 are clear, the QuaDRiGa setup is standard, and the comparisons against RVQ, a small FC net, a small Transformer, and GPT2-XL show consistent improvement. The offline framing honestly sidesteps the online-latency objection to LAMs.\n\nSoft spots, in rough order of importance. First, the transfer claim in Section IV-B is unsupported: there is no same-architecture randomly-initialized baseline, nor a variant that reinitializes the CSI-specific heads. The baselines differ in depth, width, and pretraining, so Figs. 6-9 are compatible with the simpler story that a larger fine-tuned model helps regardless of initialization. One ablation would fix this, and it should be required before the transfer narrative is trusted. Second, no error bars or multiple seeds appear anywhere; the curves are clean and the differences at some operating points look small. Third, \"outperforming existing schemes\" is too broad when standard DL autoencoders like CsiNet and its successors are absent from the comparison. They operate online, so not apples-to-apples, but the claim should be scoped. Minor: references [4] and [33] share the same arXiv ID, which looks like a typo, and the GPT2-XL citation points to the base GPT-2 paper.\n\nBottom line: the framework is plausible and worth having in the record, but the transfer mechanism is unproven and the evaluation needs tightening. Deserves a serious referee; I'd send it out expecting major revisions.","headline":"Plausible offline LVM codebook-refinement framework, but the vision-transfer claim is untested without a same-architecture random-init ablation.","tokens_in":17826,"tokens_out":3101,"would_cite":false,"duration_ms":31108,"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":"Offline AI models sharpen MIMO feedback codebooks at zero latency.","keywords":["CSI feedback","massive MIMO","codebook optimization","large vision model","offline inference","frequency-division duplex","channel state information","LVM4CF"],"falsifier":"Train LVM4CF from scratch with the same architecture, data, and loss but with random initialization instead of image pre-training; if its cosine similarity and sum rate match the pre-trained version, then the image-transfer premise is not what carries the result.","tokens_in":16891,"feed_emoji":"📡","tokens_out":5284,"duration_ms":50000,"temperature":0.7,"pith_summary":"This paper proposes a way to put large AI models into frequency-division duplex (FDD) massive MIMO channel feedback without paying their usual cost: instead of running a large model in real time on every channel vector, the model works offline to convert an ordinary RVQ codebook into an environment-tailored codebook. The refined codebook is then used at the base station, and optionally at the user, for the same low-bit feedback procedure, so the online path adds no inference latency or computational overhead. The paper claims that its model, LVM4CF, a vision-backbone network pre-trained on images and fine-tuned on CSI, attains higher cosine similarity between reconstructed and true channels and higher sum rate than conventional RVQ, FC, Transformer, and LLM baselines across site-specific and multi-scenario deployments. This matters because it offers a standard-compatible route to harvesting large-model representation power under tight latency constraints.","feed_headline":"Offline AI models sharpen MIMO feedback codebooks at zero latency","feed_subtitle":"A vision-model codebook refiner beats RVQ, transformer, and LLM baselines in cosine similarity and sum rate.","key_machinery":"The load-bearing mechanism is the offline codebook-refinement map $g_i = f_{\\mathrm{LVM4CF}}(c_i;\\Theta)$, which takes the $i$-th codeword of the conventional RVQ codebook and returns a unit-normalized refined codeword trained against the loss $\\mathcal{L}_\\rho = -\\mathbb{E}_{\\mathcal{H}_{\\text{batch}}}\\{|g^H h|/(\\|g\\|\\|h\\|)\\}$. This map is built from a vision transformer pre-trained on image tokens and fine-tuned on CSI, exploiting the claimed structural analogy between spatial correlations in images and the dual-polarized antenna array correlations in CSI codewords. The same map powers both frameworks, and the two training algorithms differ only in whether the refined codebook is used at one side or shared and iteratively updated at both sides.","core_discovery":"The central claim is that codebook-based CSI feedback can be treated as an offline codeword-refinement problem and that a large vision model is the right tool for it. LVM4CF takes each codeword from the conventional random-vector-quantization codebook, splits it into its two dual-polarization subvectors, embeds them, passes them through the pre-trained transformer layers, and outputs a refined codeword; the model is fine-tuned to maximize the expected cosine similarity between the refined codeword and the true channel. Two frameworks use this network: SSLCF fine-tunes it on site-specific CSI to produce one customized codebook, while MSLCF fine-tunes it on several environment types to produce a library of codebooks selected by an environment index. In dual-side mode the refined codebook is shared by base station and user and is iteratively updated during training; in single-side mode the user keeps the conventional codebook while the base station alone uses the refined one. The reported simulations show the refined codebooks improving both cosine similarity and downlink sum rate over the baselines, with the gain growing at high SNR.","pith_inferences":["The paper does not isolate pre-training from capacity: a randomly initialized LVM4CF trained on the same CSI data is not compared, so part of the reported gain could come from the architecture alone rather than from image transfer.","If the image-to-CSI transfer claim is right, the gain should scale with the diversity of the pre-training image data; a direct test would train variants on more or less visual data and track cosine similarity.","The dual-side iterative update can be viewed as a functional fixed-point search over codebooks; its convergence is empirically validated but not theoretically guaranteed, so a formal convergence analysis would be a natural next step.","The environment-index mechanism of MSLCF could be extended to online environment detection, letting the base station switch codebooks as the propagation environment changes."],"forward_implications":["In single-side deployment, operators can improve CSI reconstruction while leaving user equipment on the existing RVQ codebook, preserving backward compatibility.","In dual-side deployment, the shared refined codebook yields the largest cosine similarity and sum-rate gains, at the cost of delivering the codebook once to users.","MSLCF reduces that delivery cost to a short environment index, making multi-scenario deployment feasible without retraining at each site.","Because all large-model computation happens offline, adding a large vision model does not increase online feedback running time, as the paper's timing comparison shows.","At high SNR, where interference dominates, the accuracy of the refined codebooks translates into widening sum-rate gains over conventional feedback."],"supporting_citations":[{"why":"Establishes the AI-enhanced codebook-based CSI feedback baseline (FC) that this work extends with a large vision model.","marker":"[34]"},{"why":"Supplies the pre-trained large vision model backbone and pretrained weights that LVM4CF adapts for CSI codeword refinement.","marker":"[50]"},{"why":"Provides the VQGAN tokenizer used in the vision-model pre-training pipeline that produces the image token sequences.","marker":"[51]"},{"why":"Defines the random vector quantization codebook used as the conventional baseline and as the starting codebook for refinement.","marker":"[54]"},{"why":"Defines the Transformer architecture used both as a lightweight baseline and as the attention backbone inside the large vision model.","marker":"[29]"},{"why":"Provides the GPT2-XL large language model baseline that the paper compares against to argue that vision pre-training transfers better to CSI.","marker":"[30]"},{"why":"The QuaDRiGa channel simulator used to generate the site-specific and multi-scenario CSI datasets for training and evaluation.","marker":"[52]"},{"why":"Defines the 3GPP urban and rural channel scenarios (UMa, RMa) used to create environment-specific training and test data.","marker":"[53]"}],"fun_headline_variants":["Offline vision models refine CSI codebooks at zero latency","LVM4CF: offline codebook refinement for MIMO feedback","Extract best codewords, discard rest: offline AI for CSI","Zero-latency CSI feedback via offline large AI models","Site-specific codebook tuning with large vision models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that spatial correlations learned from natural images transfer to dual-polarized CSI codewords, so that image pre-training, rather than model size or training procedure, is what gives LVM4CF its edge.","fun_headline_variants_meta":{"raw":{"variants":["Offline vision models refine CSI codebooks at zero latency","LVM4CF: offline codebook refinement for MIMO feedback","Extract best codewords, discard rest: offline AI for CSI","Zero-latency CSI feedback via offline large AI models","Site-specific codebook tuning with large vision models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000173,"raw_usage":{"total_tokens":1348,"prompt_tokens":1083,"completion_tokens":265,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":699,"completion_tokens_details":{"reasoning_tokens":181}},"tokens_in":699,"tokens_out":265,"duration_ms":2886,"temperature":1.0,"reasoning_tokens":181,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:51:00.981217+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train LVM4CF from scratch with the same architecture, data, and loss but with random initialization instead of image pre-training; if its cosine similarity and sum rate match the pre-trained version, then the image-transfer premise is not what carries the result.","supporting_citations":[{"cited_title":"AI-enhanced codebook-based CSI feedback in FDD massive MIMO,","cited_arxiv_id":null,"evidence_quote":"Establishes the AI-enhanced codebook-based CSI feedback baseline (FC) that this work extends with a large vision model."},{"cited_title":"Sequential modeling enables scalable learning for large vision models,","cited_arxiv_id":null,"evidence_quote":"Supplies the pre-trained large vision model backbone and pretrained weights that LVM4CF adapts for CSI codeword refinement."},{"cited_title":"Taming transformers for high- resolution image synthesis,","cited_arxiv_id":null,"evidence_quote":"Provides the VQGAN tokenizer used in the vision-model pre-training pipeline that produces the image token sequences."},{"cited_title":"MIMO broadcast channels with finite-rate feedback,","cited_arxiv_id":null,"evidence_quote":"Defines the random vector quantization codebook used as the conventional baseline and as the starting codebook for refinement."},{"cited_title":"Attention is all you need,","cited_arxiv_id":null,"evidence_quote":"Defines the Transformer architecture used both as a lightweight baseline and as the attention backbone inside the large vision model."},{"cited_title":"QuaDRiGa: A 3- d multi-cell channel model with time evolution for enabling virtual field trials,","cited_arxiv_id":null,"evidence_quote":"The QuaDRiGa channel simulator used to generate the site-specific and multi-scenario CSI datasets for training and evaluation."},{"cited_title":"Study on channel model for frequencies from 0.5 to 100 GHz,","cited_arxiv_id":null,"evidence_quote":"Defines the 3GPP urban and rural channel scenarios (UMa, RMa) used to create environment-specific training and test data."}],"review_version":1}