{"id":"f5426876-d4f2-41fb-901a-a30eccc1cbd7","arxiv_id":"2505.09141","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An LLM-based model that fuses historical sensing and communication CSI predicts future communication channels with lower NMSE than non-sensing deep-learning baselines in a ray-traced ISAC street scenario.","lead":"A base station can use its own radar-like echoes, together with past signal measurements, to predict where a moving user's wireless channel is heading. This paper builds a large-language-model predictor that fuses sensing and communication data, claiming lower prediction error than standard deep-learning baselines in a simulated street scene.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The sensing-channel model in Eq. (2) assigns shared scatterers the same Doppler as the communication channel, which is physically implausible for monostatic sensing and may inflate the reported gains.","rationale":"The reader's weakest assumption was that the single Sionna street scene may not generalize to real complex environments. I agree that transferability is unestablished, but the more load-bearing and more concrete problem is inside the synthetic model itself: Eq. (2) assumes identical Doppler parameters for shared paths in monostatic sensing and bistatic communication channels. That assumption is physically questionable for moving scatterers, and the paper provides no Sionna-based verification that the Doppler values actually coincide. Because the entire sensing-assisted gain rests on the sensing channel being temporally informative about the communication channel, an unrealistic Doppler tie can inflate the measured benefit of both sensing assistance and the LLM fusion. The paper has real strengths: the architecture is coherently described, the ablation table includes a no-LLM sensing-augmented configuration, and the simulation setup is more realistic than pure LoS models. I am therefore not proposing rejection; the right outcome is a conditional acceptance or revision requiring a corrected physical model and rerun of the experiments. The proposed test directly settles whether the central empirical claim survives physically correct channel generation.","tokens_in":8096,"tokens_out":8464,"duration_ms":91570,"concrete_test":"Re-generate the Sionna dataset using physically correct round-trip Doppler shifts for monostatic sensing paths, or extract and report the per-path Doppler pairs for the 14 shared paths from Sionna to verify whether the equality in Eq. (2) actually holds. Then retrain the proposed LLM model and the Table I ablations on the corrected dataset and replot Figs. 4-5. If the proposed method's NMSE advantage over the no-sensing or W/o-LLM configuration shrinks materially, the headline claim is an artifact of the Doppler identity; if the gain persists, the concern is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim depends on sensing CSI carrying usable information about future communication CSI, but the paper's only evidence comes from a simulated street scene whose channel model appears to make this information artificially strong. In Section II, Eq. (2) and the discussion after it state that shared-scatterer paths in the sensing and communication channels have the same Doppler parameter f_{d,i}. For a monostatic sensing channel this is not generally true: a moving scatterer or the moving MU (which Footnote 1 explicitly treats as a shared scatterer) produces a round-trip Doppler of about 2v/λ, whereas the communication path has a one-way or bistatic Doppler of v/λ or a different projection. Unless all shared scatterers are static, using the same f_{d,i} makes the sensing phase time series track the communication phase time series far more closely than in a real ISAC environment. Fig. 2(b) verifies only azimuth/elevation overlap, not the joint Doppler-phase relation that determines whether sensing CSI is actually predictive. If the dataset was generated under this shared-f_d assumption, the measured NMSE advantage of sensing assistance may be an artifact; if it was not, the model equations still need correction. This is a concrete internal risk, distinct from the also-real absence of error bars, code, and cross-scenario validation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This letter proposes an LLM-based sensing-assisted channel prediction framework for a multi-antenna OFDM ISAC system. The system uses the previous P slots of communication and sensing CSI to predict the communication CSI in the next Q slots, as formalized in Eq. (3). The proposed architecture consists of an IDFT-based preprocessor, a ConvLSTM channel attention module, a cross-attention fusion module, a pre-trained GPT-2 backbone with frozen attention/feedforward layers, and an output module with de-normalization. The authors evaluate the method on a Sionna ray-traced street scenario with one ISAC BS and one mobile user, comparing against LSTM, Transformer, CNN, and a no-sensing benchmark using NMSE. The reported results show the proposed approach achieves the lowest NMSE across MU speeds and SNRs, and an ablation study indicates that the sensing channel, channel attention, cross-attention, and LLM backbone each contribute to performance. The central claim is that the LLM-based method with sensing assistance significantly outperforms conventional deep learning baselines and the no-sensing benchmark.","tokens_in":8408,"tokens_out":3676,"duration_ms":36853,"significance":"If the results are robust, the paper introduces a novel and potentially impactful idea: adapting a pre-trained text LLM to fuse sensing and communication CSI for channel prediction, which is relevant to ISAC and 6G systems. The architecture is clearly described, the problem formulation in Eq. (3) is sound, and the ablation study is useful for showing the value of individual modules. However, the empirical support is currently limited to a single simulated street scene, the statistical evidence for 'significantly outperforms' is absent, and the sensing channel model in Eq. (2) contains a physically questionable assumption that may overstate the predictive value of sensing CSI. These issues affect the load-bearing claim of the paper, and therefore the manuscript needs substantial revision before its conclusions can be accepted.","major_comments":[{"comment":"The sensing channel model in Eq. (2) assigns the same Doppler frequency shift f_{d,i} to shared-scatterer paths as the communication channel in Eq. (1). For a monostatic sensing channel, the round-trip Doppler of a moving scatterer is twice the one-way Doppler (or generally differs by the bistatic geometry), so shared scatterers, including the MU explicitly treated as a shared scatterer in Footnote 1, do not have identical f_{d,i} in the two channels. If the Sionna dataset was generated under this shared-f_{d,i} assumption, the sensing CSI would track the communication phase nearly perfectly, artificially inflating the measured gains of sensing assistance. The authors should either correct the sensing channel model to use physically consistent Doppler shifts, or confirm that the Sionna simulator computed separate Doppler shifts for sensing and communication paths and report the actual Doppler relationship of the 14 shared paths identified in Fig. 2(b). Fig. 2(b) verifies only azimuth/elevation overlap, not the Doppler-phase relation that determines whether sensing CSI is predictive.","section":"Section II, Eq. (2)"},{"comment":"The claim that the proposed approach 'significantly outperforms' baselines is not supported statistically. Results are presented without error bars, confidence intervals, or multiple random seeds, and the test set contains only 600 samples; differences such as NMSE 0.407 vs. 0.515 in Table I could fall within run-to-run variation. The authors should provide repeated training runs with variance estimates and, if 'significantly' is meant in a statistical sense, a formal significance test.","section":"Section IV, Figs. 4-5 and Table I"},{"comment":"All deep learning baselines (LSTM, Transformer, CNN) receive only communication CSI, while the proposed method receives both communication and sensing CSI. This conflates the effect of sensing information with the effect of the LLM architecture. A fair evaluation requires giving at least one conventional DL baseline the same sensing CSI input, for example through a simple concatenation or attention fusion, so that the gain from sensing assistance can be isolated from the gain from the backbone choice. Without this, the comparison does not demonstrate that the sensing data, rather than the LLM itself, is responsible for the reported improvement.","section":"Section IV, baselines"},{"comment":"The entire evaluation relies on a single Sionna ray-traced street scene with one BS, one MU, and a particular set of 16 communication and 26 sensing paths, of which 14 are shared. The paper's conclusion that the method works in 'realistic and complex wireless environments' is not justified without additional scenes with different geometries, richer multipath statistics, or real measurement data. The reader has no evidence that the measured NMSE advantage transfers beyond this one geometry, which is a crucial limitation given that the method is motivated by complex environments.","section":"Section IV, simulation scenario"}],"minor_comments":[{"comment":"There is a typo: 'conventioanl' should be 'conventional'.","section":"Section III-B"},{"comment":"The sensing channel in Eq. (2) uses the same steering vector a(θ_i, φ_i) for both transmit and receive; for a UPA, the receive steering vector may differ depending on angle of arrival, and this should be clarified or justified.","section":"Section II, Eq. (2)"},{"comment":"The notation X_{c,n}, X_{c,n,τ}, X_{s,n}, X_{s,n,τ} is confusing because τ appears as a subscript without prior definition; explicitly stating the matrix shapes and the role of τ would improve readability.","section":"Section III-A"},{"comment":"The description says that Doppler frequency shifts are applied to all paths based on vehicle and MU speeds, but it does not state whether the same Doppler value is used for the sensing and communication versions of shared paths; this needs to be clarified in light of Eq. (2), as it directly affects the validity of the experimental results.","section":"Section IV, dataset generation"},{"comment":"Training details such as learning rate, number of epochs, batch size, optimizer, and computational cost are not reported for any of the methods, making the experiments difficult to reproduce or compare against. Adding these details would strengthen the paper.","section":"Section IV"}],"recommendation":"major_revision","confidential_remarks":"The shared-Doppler issue in Eq. (2) is the most serious technical concern and should be checked before acceptance; if the Sionna simulation used identical Doppler for shared paths, the core empirical claim may be an artifact. The absence of statistical evaluation and the lack of a conventional baseline with sensing CSI further weaken the current evidence. The topic fits the journal's scope, but the manuscript requires substantial additional experiments and model corrections to justify publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper combines sensing-assisted channel prediction with an LLM backbone (GPT-2) and a cross-attention fusion module. That combination is new; the architecture is plausible; the ablations show each component contributes. But the headline 'significantly outperforms' rests on a single simulated street scene, and the model equations contain a physical error that could have loaded the dice.\n\nThe good: converting CSI to delay domain, running ConvLSTM channel attention on communication and sensing branches, fusing them with cross-attention, then feeding GPT-2 with frozen weights is a sensible adaptation. Extending LLM4CP to the sensing-assisted case is worth doing, and the paper documents the setup enough to reproduce the experiments in principle.\n\nThe problems: most seriously, Eq. (2) assigns shared scatterers the same Doppler f_d,i in both the communication and monostatic sensing channels. A monostatic echo from a moving scatterer—or from the MU, which Footnote 1 treats as a shared scatterer—has roughly twice the one-way Doppler. If the Sionna dataset was generated with physical Doppler, the equation is wrong; if it was generated with the same f_d, the sensing phase tracks the communication phase artificially well and the reported gain may be an artifact. Fig. 2(b) verifies angle overlap only, not the joint Doppler-phase relation. This needs to be clarified.\n\nSecond, the deep-learning baselines (LSTM, Transformer, CNN) do not receive sensing CSI, so they cannot isolate whether the gain comes from the extra sensing input or from the LLM design. The 'w/o sensing' ablation is closer, but it's the same architecture without sensing. A sensing-augmented Transformer or LSTM baseline is missing. There are also no error bars, no code/data release, and only one scenario. These are fixable, but they leave the headline claim unsupported in this version.\n\nIs it a serious paper? Yes. It tackles a real problem, the architecture is well-motivated, and the experiments are clearly described. I would not desk-reject it. But it is not ready to be accepted as is; the Doppler issue and the baselines need active work.\n\nRecommendation: send to peer review, expect major revision. The authors have something worth engaging with, and a reviewer can push them to fix the physics and strengthen the comparisons.","headline":"Interesting architecture and a real problem, but a likely physical flaw in the Doppler model plus weak baselines leave the headline gain in question.","tokens_in":8892,"tokens_out":3726,"would_cite":false,"duration_ms":34724,"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":"This paper claims that an LLM adapted with ConvLSTM channel attention and cross-attention fusion can predict future communication CSI from historical communication and sensing CSI with significantly lower NMSE than conventional deep…","keywords":["sensing-assisted channel prediction","integrated sensing and communication","ISAC","large language model","OFDM channel prediction","channel state information","cross-attention fusion","ray-tracing simulation"],"falsifier":"Measure the same predictor on a second simulated or measured environment in which the number of shared dominant paths is much smaller, say fewer than five, while keeping the same LLM architecture and training budget; if the NMSE gap over the no-sensing baseline shrinks to near zero, the paper's stated mechanism is not general.","tokens_in":7915,"feed_emoji":"📡","tokens_out":4327,"duration_ms":40854,"temperature":0.7,"pith_summary":"This paper tries to show that a wireless base station can predict its own future communication channel more accurately by listening to its environment: historical sensing echoes are fed alongside previous communication channel estimates into an LLM-based predictor. The setting is an integrated sensing and communication (ISAC) OFDM system in a rich multi-path street environment, where sensing and communication channels share physical scatterers. The authors claim that adapting a pre-trained text-based LLM, with ConvLSTM channel attention and cross-attention fusion, achieves lower normalized mean square error than deep-learning baselines and than predicting from communication CSI alone. If the claim holds, sensing data become a practical aid for channel acquisition in high-mobility 6G links.","feed_headline":"Sensing data help an LLM predict future wireless channels","feed_subtitle":"Historical sensing echoes plus communication CSI beat deep-learning baselines in a simulated street scene, including at low SNR.","key_machinery":"Eq. (3) is the functional target. The mechanism is a preprocessing stage that converts frequency-domain CSI to delay domain, a channel attention module built from ConvLSTM cells, a cross-attention fusion of communication and sensing features, a frozen pre-trained LLM backbone with learnable normalization and positional embedding layers, and an output head that produces complex channel matrices. The load-bearing element is the cross-attention fusion: it lets the model combine temporal-spatial features from both channels before the LLM backbone processes the sequence.","core_discovery":"The central claim is that the mapping in Eq. (3) can be learned well: given the previous $P$ slots of both communication CSI and sensing CSI, the proposed network predicts the next $Q$ slots of communication CSI with substantially lower NMSE than LSTM, Transformer, CNN, and no-sensing baselines. The empirical basis is a ray-tracing street scene in which 14 of 16 communication paths and 26 sensing paths share the same angles, so the sensing channel carries information about communication scatterers. The paper also claims robustness under noisy historical CSI and shows by ablation that sensing input, channel attention, cross-attention, and the LLM backbone all contribute to the gain.","pith_inferences":["Editorial inference: the number of shared scatterers is the likely control knob; if the method's gain comes from overlapping paths, the NMSE advantage should grow when the ray tracer adds more shared scatterers and shrink when they are removed, so a sweep over $N_0$ would test the mechanism directly.","Editorial inference: because the LLM backbone is frozen and only normalization and embedding layers are tuned, a much smaller transformer with the same cross-attention may capture most of the gain at lower cost; comparing against a same-size untrained transformer would isolate the contribution of pre-training.","Editorial inference: the same fusion idea could be transferred to predictive beamforming or precoder selection without first reconstructing full CSI, since the predictor already outputs future channel matrices."],"forward_implications":["ISAC transmitters can predict future communication CSI from historical sensing and communication CSI with lower NMSE than sensing-free deep-learning predictors, especially at high mobile-user speeds.","Sensing information remains valuable even when the historical CSI is noisy, since the proposed approach keeps the lowest NMSE across the tested SNR range.","Each designed module contributes to the gain: removing sensing input, channel attention, cross-attention, or the LLM backbone increases NMSE.","The shared-scatterer structure of real environments is directly exploitable without explicit scatterer extraction or geometry estimation."],"supporting_citations":[{"why":"Supplies the ray-tracing simulation platform used to generate the simulated street-scene CSI dataset.","marker":"[13]"},{"why":"Defines the LLM-for-channel-prediction paradigm that the proposed approach adapts and extends with sensing assistance.","marker":"[3]"},{"why":"Provides the Transformer baseline and the idea of parallel channel prediction.","marker":"[4]"},{"why":"Gives the ConvLSTM spatiotemporal module used for channel attention.","marker":"[5]"},{"why":"Serves as the pre-trained LLM backbone whose transformer layers are frozen.","marker":"[15]"},{"why":"Supports the quasi-static channel model with slot-varying complex coefficients and stable path parameters.","marker":"[12]"},{"why":"Is a prior sensing-assisted channel recovery work that motivates using environmental sensing for communication channel acquisition.","marker":"[7]"}],"fun_headline_variants":["LLM uses sensing echoes to forecast wireless channels","Sensing-aided LLM predicts future wireless CSI","LLM channel prediction gets boost from sensing data","ISAC helps LLM foresee wireless channel states","Sensing data enhance LLM-based channel prediction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The premise is that one ray-traced street scene, with one base station, one moving user, and 14 shared dominant paths between the sensing and communication channels, faithfully represents the 'realistic and complex wireless environments' where the stated gains are meant to hold. If real environments have fewer shared scatterers or larger distribution shift, the NMSE advantage may not transfer.","fun_headline_variants_meta":{"raw":{"variants":["LLM uses sensing echoes to forecast wireless channels","Sensing-aided LLM predicts future wireless CSI","LLM channel prediction gets boost from sensing data","ISAC helps LLM foresee wireless channel states","Sensing data enhance LLM-based channel prediction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000227,"raw_usage":{"total_tokens":1421,"prompt_tokens":846,"completion_tokens":575,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":462,"completion_tokens_details":{"reasoning_tokens":503}},"tokens_in":462,"tokens_out":575,"duration_ms":5220,"temperature":1.0,"reasoning_tokens":503,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:38:13.703530+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the same predictor on a second simulated or measured environment in which the number of shared dominant paths is much smaller, say fewer than five, while keeping the same LLM architecture and training budget; if the NMSE gap over the no-sensing baseline shrinks to near zero, the paper's stated mechanism is not general.","supporting_citations":[{"cited_title":"Spatio- temporal neural network for channel prediction in massive MIMO- OFDM systems,","cited_arxiv_id":null,"evidence_quote":"Gives the ConvLSTM spatiotemporal module used for channel attention."},{"cited_title":"Sensing-assisted sparse channel recovery for massive antenna systems,","cited_arxiv_id":null,"evidence_quote":"Is a prior sensing-assisted channel recovery work that motivates using environmental sensing for communication channel acquisition."}],"review_version":1}