{"id":"95965433-edde-4ad9-886a-dad0afdab7b0","arxiv_id":"2507.13676","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"CARTS fuses DMRS and SRS channel estimates, with adaptive aperiodic SRS triggering and asynchronous channel stitching, achieving NMSE 0.167 and 85 cm tracking at 10 UEs in emulation.","lead":"A 5G base station scheme called CARTS combines two existing uplink reference signals, DMRS and SRS, to get more frequent channel measurements for sensing and communication. In trace-driven emulation, it supports about twice as many users as a periodic SRS-only baseline with similar accuracy, without extra radio resources.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stitching's single-boundary-subcarrier alignment and the missing re-add of the reference phase slope are load-bearing, but only validated at 3 km/h indoor; a higher-speed or richer-multipath test is needed before the twice-users claim generalizes.","rationale":"The reader's weakest assumption is that the stitching model's shared-phase-reference and time-stable-offset assumptions are load-bearing and under-tested; I agree with that identification. I add two more specific technical details that make the concern sharper: (i) the single-boundary-subcarrier estimate in Eqs. 12-13 is high-variance in frequency-selective channels and its errors propagate through the iterative stitching in §4.2.4, and (ii) the de-sloping step in Eq. 8 is never paired with an explicit re-add of α_b_ref/φ_b_ref,0, yet the final stitched phase in Fig. 11b retains a strong slope, so the CIR-peak/TA validation depends on an undocumented step. Neither issue by itself forces rejection: the trace-driven evidence is real, the algorithm is standard-compliant, and the reported numbers are plausible in the slow indoor regime. But both issues sit directly on the central claim that stitching produces a full-band CSI nearly as good as full-band SRS, and neither is resolved by the current evaluation. A synthetic channel stress test with controlled delay spread, SNR, and time gaps would settle whether the concern lands; until then the CONDITIONAL verdict is the right one, so I recommend no change to the reader's verdict.","tokens_in":20469,"tokens_out":10566,"duration_ms":142293,"concrete_test":"Implement §4.2 exactly as written on a synthetic OFDM channel with 2-3 Rayleigh taps, delay spread 0.5-2 μs, SNR 20 dB, and DMRS-to-SRS gaps from 1 to 20 ms at speeds from 3 to 30 km/h; compute NMSE and CIR peak error against the full-band true channel. In the same test, replace the single-boundary γ_b estimate (Eqs. 12-13) with an estimate averaged over 5 boundary subcarriers. If either change moves NMSE by more than ~0.05 or CIR peak error by more than 1 sample, the §5 conclusions are regime-dependent and the operating envelope must be stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on the stitching in §4.2: after removing each sub-band's linear phase slope (Eq. 8), all sub-bands are assumed to share a common phase reference, so a single complex scalar γ_b estimated from overlapping or boundary subcarriers (Eqs. 10-13) can align each sub-band to b_ref. This is the weakest load-bearing step. In a frequency-selective channel, H_b_ref(n)/H_b(n) is not constant in n, and when there is no overlap Eqs. 12-13 estimate γ_b from one boundary subcarrier pair, making the alignment noise-dominated. Since §4.2.4 stitches iteratively outward from b_ref, a bias in one γ_b propagates to every subsequently stitched sub-band. The same section removes the phase slope from each sub-band but never shows where the retained slope α_b_ref and intercept φ_b_ref,0 are re-added; because Fig. 11b shows a final stitched CSI with a large nonzero phase slope, either an undocumented re-add step exists or the equations are incomplete. If Eq. 8 is applied to b_ref exactly as written, the absolute delay information is destroyed and the CIR-peak/TA result in §5.1.2 cannot be produced as described. The evaluation uses a single UE at about 3 km/h in three indoor settings, so DMRS-to-SRS gaps are short and the channel is near-stationary; this does not stress the frequency-selectivity of the boundary estimate or the mobility-induced phase drift. If either effect is significant, the reported NMSE 0.167, 85 cm tracking error, and twice-user scalability comparison become artifacts of a benign test regime.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"CARTS is a framework for 5G uplink ISAC that fuses DMRS and SRS CSI streams to increase the frequency of channel estimates and extend sensing opportunities to more users. The paper proposes (i) a priority-based aperiodic SRS triggering algorithm (Algorithm 1) that complements the uncontrollable DMRS schedule, and (ii) a channel stitching and compensation method (§4.2) that aligns asynchronous, partially overlapping CSI sub-bands by spatial smoothing, slope removal, and complex scaling. The authors evaluate CARTS in a trace-driven emulation using CSI collected from an OAI/USRP testbed and PUSCH allocation traces sniffed from commercial 4G networks, reporting an NMSE of 0.167, a CIR peak error of about one sample, a UE tracking error of 85 cm, and support for roughly twice as many users as a periodic SRS-only baseline at similar performance. The paper claims that CARTS is standard-compliant and requires no additional radio resources.","tokens_in":20763,"tokens_out":9630,"duration_ms":97540,"significance":"If the quantitative claims hold, CARTS would be a practical and inexpensive way to improve uplink sensing scalability in 5G networks without new hardware or radio resources. The paper's strengths include a clearly specified algorithm, a reproducible emulation pipeline, and direct comparisons against ToneTrack and a periodic-SRS baseline, as well as the useful spatial-smoothing step. However, the current evaluation is undermined by a partially circular ground truth (the reference sub-band is drawn from the same full-band SRS used as H_true), an incomplete specification of how the reference phase slope is restored, an alignment step that relies on a single boundary subcarrier, and an arbitrary scaling of 4G traces to 5G bandwidth. These issues must be resolved before the reported numbers can be considered reliable.","major_comments":[{"comment":"The NMSE ground truth H_true is the full-band SRS measurement, and the stitching reference sub-band b_ref is itself a subset of that same measurement chosen via the SRS allocation mask in Fig. 9. Consequently, the estimate is exactly equal to the ground truth on the reference subcarriers, and the reported NMSE values (e.g., 0.167 at N=10) measure how well the other sub-bands can be aligned to a reference that is known by construction, not how well the full channel is estimated independently. The claim that CARTS achieves a given NMSE therefore needs a non-circular evaluation, e.g., a held-out full-band measurement from a different SRS occasion, or a ground truth derived from an independent full-band channel estimate.","section":"§5.1, Fig. 9"},{"comment":"Equations (7)-(8) remove the linear phase slope from every sub-band, and the text says the reference slope α_b_ref and intercept φ_b_ref,0 are 'retained,' but no equation in §4.2 shows how these quantities are re-added to the final stitched estimate. If Eq. (8) is applied to b_ref exactly as written, the absolute delay (timing advance) information is destroyed, which contradicts the CIR peak position results of §5.1.2; if a re-add step exists, it is undocumented. The large phase slope visible in the 'Stitched (Ours)' trace in Fig. 11b suggests some restoration occurs, so the manuscript needs an explicit equation completing the stitching operation.","section":"§4.2.2, Eqs. (7)-(8), Fig. 11b"},{"comment":"When there is no overlap between a sub-band and the reference, the alignment scalar γ_b is estimated from a single boundary subcarrier pair (Eqs. 12-13). In a frequency-selective channel, the ratio Ĥ_b_ref(n)/Ĥ_b(n) is not constant across n, so this estimate is noise-dominated and biased; the iterative outward stitching of §4.2.4 then propagates that bias to every subsequently stitched sub-band. The evaluation only covers 3 km/h indoor motion in three settings (office, open floor, NLOS), with sub-bands as narrow as 5 RBs (Fig. 7), so the single-subcarrier boundary estimate and the effects of larger DMRS-to-SRS time gaps or richer multipath are not exercised. A higher-mobility or stronger-multipath test, or an estimator using multiple boundary subcarriers, is needed to support the claimed generalizability.","section":"§4.2.3, Eqs. (12)-(13), §4.2.4, §5"},{"comment":"The PUSCH allocation traces are collected from a 4G network and scaled from 100 RBs to 272 RBs by an arbitrary factor of 2.72, rounded to a multiple of 4. This scaling changes the burst size distribution and the resulting SRS triggering decisions, yet no justification or sensitivity analysis is provided. Because the 'twice as many users' claim is derived from these traces, the emulation should either use a 5G-based mapping (e.g., from 5G NR schedulers or standardized traffic models) or demonstrate that the conclusions are invariant to the scaling factor.","section":"§5, data collection, Fig. 9"},{"comment":"CARTS is described in the abstract as a 'real-time SRS triggering algorithm,' but the evaluation is entirely offline because OAI lacks aperiodic SRS triggering. The SRS allocation decisions are computed from pre-recorded PUSCH traces, so the loop's latency, jitter, and its coupling with the actual PUSCH scheduler are never measured. This limitation is acknowledged in §5, but the phrasing of the abstract and §4.1 should be revised to distinguish the algorithm's design from its offline validation, or supplemented with a latency and signaling-overhead analysis.","section":"Abstract, §4.1, §5"}],"minor_comments":[{"comment":"The pseudocode uses 'arg max(value_matrix)' but the variable is named 'urgency_matrix' throughout; this should be corrected for reproducibility.","section":"Algorithm 1, line 12"},{"comment":"The text states 'median NMSE values well below 0.25 (i.e., 3 dB)', but 10 log10(0.25) is approximately -6 dB, not 3 dB; the conversion should be fixed.","section":"§5.1.1"},{"comment":"The phrase 'In spit of more complex fading processes' contains a typo and should read 'In spite of'.","section":"§5.1.3"},{"comment":"The parameter 'maxNrofSRS-TriggerStates-1' is described as limiting the number of resource sets, but the 3GPP parameter name suggests it controls trigger states; the authors should verify the exact parameter and its meaning in TS 38.331.","section":"§4.1.1"},{"comment":"The legend entry '100% RR Traffic' is ambiguous and appears inconsistent with the caption's 'Baseline - Periodic SRS' label; the legend and caption should be aligned.","section":"Fig. 12a"}],"recommendation":"major_revision","confidential_remarks":"The main risks are the circular NMSE evaluation and the undocumented phase re-add step; both are fixable in revision. The paper would be stronger with a sensitivity analysis of the 2.72 RB scaling factor and a higher-mobility experiment. The offline-only evaluation should be framed as a limitation rather than having the abstract describe the triggering as real-time."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Plain-English take: this is a serious engineering paper. It fuses DMRS and SRS uplink CSI for sensing in a standard-compliant way, adds an adaptive aperiodic SRS triggering algorithm, and shows a credible path to roughly doubling the number of supported sensing users. The idea of combining the two reference signals is genuinely new in this form, and the authors have done a real trace-driven study rather than a toy simulation. The algorithm is clearly specified, the comparison against periodic SRS and ToneTrack is useful, and the reporting of NMSE and CIR peak errors is thorough.\n\nThe soft spots are real but mostly addressable. The stitching method in §4.2 removes the linear phase slope from every sub-band including the reference, but the equations never show where the reference slope and intercept are re-added. Since the final stitched CSI (Fig. 11b) shows a large nonzero phase slope, either there is an undocumented re-add step or the equations are incomplete. That is a load-bearing gap because the absolute delay information is needed for the CIR-peak and TA results. The single-boundary-subcarrier alignment (Eqs. 12-13) is indeed noise-dominated in rich multipath, and errors will propagate iteratively. The evaluation does not stress this: only 3 km/h indoor walks, mostly LOS, with short DMRS-to-SRS gaps. The 4G-to-5G scaling factor of 2.72 is heuristic, and the ground truth is the same full-band SRS used to build the reference sub-band, so the numbers are partly self-referential. No code or data are released.\n\nI don't think any of this sinks the central claim. The paper is careful to frame its results within the tested conditions, and the scalability gain from using DMRS as free sensing opportunities is robust to these concerns. The main fixes are: document the slope re-add step, test at higher UE speed and in a frequency-selective outdoor environment, and either justify the 2.72 scaling or treat it as a sensitivity parameter. Releasing the stitching code would also help.\n\nThis paper deserves a serious referee. It is a net contribution to the 5G/6G ISAC subfield. I'd send it to review with major-revision expectations rather than desk reject.","headline":"A useful, well-scoped 5G ISAC paper whose main claims likely hold in the tested regime, but the stitching equations have a real gap and the evaluation is too gentle to fully stress them.","tokens_in":21414,"tokens_out":2485,"would_cite":true,"duration_ms":26233,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CARTS fuses the two 5G uplink reference signals—traffic-borne DMRS and on-demand SRS—into a single stitched channel estimate, letting a base station serve twice as many sensing users at the same accuracy.","keywords":["5G ISAC","uplink sensing","channel state information","DMRS","SRS","channel stitching","aperiodic SRS triggering","CSI fusion"],"falsifier":"Measure the stitched channel's NMSE against full-band SRS ground truth while systematically widening the time gap between a DMRS capture and the SRS reference capture, first at the paper's 3 km/h indoor setting and then at vehicular speeds: if NMSE stays well below the 0.25 threshold as the gap grows, the single-gain alignment is robust, while a sharp rise with gap size pinpoints temporal staleness as the weak link. A second decisive check is error growth as a function of stitched sub-band count, comparing two-sub-band stitching against six-sub-band stitching, where superlinear growth would confirm that errors accumulate across iterative stitching and would quantify how much of the reported accuracy depends on the number of SRS resources per user.","tokens_in":2075,"feed_emoji":"📡","tokens_out":3779,"duration_ms":146896,"temperature":0.7,"pith_summary":"CARTS sets out to prove that a 5G base station can get faster, fresher channel information by fusing the two uplink reference signals it already receives, instead of treating them as separate streams. One of those signals, DMRS, rides inside every data transmission but appears only when users have traffic to send; the other, SRS, can be triggered on demand but is a scarce resource shared among users. The paper's answer is to stitch the partial frequency-band measurements of the two signals into one full-band channel estimate, and to trigger SRS adaptively only where DMRS leaves gaps. In a trace-driven evaluation, CARTS reports a channel-estimation error (normalized mean squared error) of 0.167 and user-tracking accuracy of 85 cm, while serving twice as many users as a periodic SRS-only baseline at similar accuracy. If correct, this gives operators a standard-compliant path to 5G sensing that costs no extra radio resources.","feed_headline":"Stitching two 5G signals doubles sensing capacity","feed_subtitle":"By fusing DMRS and SRS measurements, CARTS keeps 85 cm tracking accuracy while serving twice as many users.","key_machinery":"The mechanism that carries the argument is the slope-removal stitching and compensation pipeline of Section 4.2. Each partial measurement is a tuple storing the complex CSI matrix, the subcarrier set, and the measurement time of one sub-band. The pipeline first projects all sub-bands onto the principal eigenvector of a time-weighted spatial covariance matrix so they share one spatial reference; then it fits and removes each sub-band's linear phase slope, deliberately keeping the reference band's slope and intercept; then it aligns each sub-band to the reference with a single complex scaling factor estimated from overlapping or boundary subcarriers; finally it stitches outward iteratively, preferring adjacent sub-bands and keeping the newest measurement in any overlap, with spline interpolation restoring uniform sampling. The load-bearing idea is that under mobility a moving user changes spatial signatures, path delays, and fading simultaneously, so all three corrections must work together for the fused estimate to behave like a real full-band measurement.","core_discovery":"The paper's central claim is that asynchronous DMRS and SRS measurements, each covering only part of the band at different times, can be stitched into a coherent full-band channel estimate nearly as accurate as full-band SRS sounding. The stitching pipeline works in three stages: spatial smoothing projects each sub-band's CSI onto the dominant spatial mode of a time-weighted covariance matrix; time alignment removes each sub-band's linear phase slope instead of shifting channel-impulse-response peaks, which the paper argues is unreliable for narrow sub-bands and accumulates errors when stitched repeatedly; and frequency compensation scales and re-phases each sub-band by a single complex factor estimated from overlapping or boundary subcarriers, applied iteratively outward from a reference SRS sub-band with the most recent measurement kept in overlaps. A second claim is that the unavoidable DMRS stream can be treated as free sensing opportunity: a greedy priority-based scheduler, working within the standard aperiodic-SRS trigger mechanism and per-user target estimation rates, fills only the gaps DMRS leaves. The evaluation, combining real uplink-scheduling traces from three traffic environments with full-band CSI collected at roughly 3 km/h movement, reports NMSE 0.167 and 85 cm tracking accuracy at 10 users, on par with the periodic SRS baseline's accuracy at 5 users.","pith_inferences":["A stress test the paper leaves open: raising UE speed from 3 km/h toward vehicular rates, or deliberately lengthening the DMRS-to-SRS delay, would find how large the temporal gap can be before the single-gain alignment goes stale.","The same slope-removal stitching could transfer to carrier aggregation or non-contiguous spectrum chunks, where partial CSI from separate bands must be combined under user motion.","Because every aperiodic SRS firing consumes a downlink control grant, dense-cell control-channel capacity could become the next bottleneck; the paper's evaluation does not model DCI channel loading.","The fixed DMRS/SRS power offset the paper measures suggests a cheap practical refinement: pre-calibrating the reference-signal power difference before stitching would relieve the amplitude-scaling factor and directly shrink the ranging error that dominates tracking error."],"forward_implications":["A standard-compliant base station can raise its CSI update rate without spending extra radio resources, because DMRS already accompanies data traffic and SRS is triggered through the standard aperiodic-SRS control field.","Sensing opportunity spreads to more users: CARTS reports NMSE 0.167 and tracking accuracy 0.85 m at 10 users, where the periodic SRS-only baseline reaches similar numbers at just 5 users.","Timing-advance integrity holds as load grows: average CIR peak error stays near one sample (about 16.3 ns at 30 kHz subcarrier spacing) even with 100 users, so uplink synchronization does not become the bottleneck.","The error decomposition (ranging error growing 47% versus angular error 24% as users rise from 5 to 100) indicates that improving CSI amplitude fidelity is the higher-leverage path to better positioning."],"supporting_citations":[{"why":"Supplies the direct stitching baseline: the prior Wi-Fi multi-channel alignment method whose peak-shifting approach accumulates errors on 5G sub-bands, which CARTS's slope-removal method is designed to beat.","marker":"[41]"},{"why":"The adaptive SRS scheduling scheme whose scalability limits, communication side effects, and RRC reconfiguration overhead motivate CARTS's switch to aperiodic DCI-triggered SRS.","marker":"[8]"},{"why":"The 5G RRC specification that fixes maxNrofSRS-TriggerStates-1, bounding the three resource sets that shape CARTS's six-resource SRS allocation.","marker":"[1]"},{"why":"The UE capability specification giving maxNumberSRS-ResourcePerSet=2, which fixes the two-resources-per-set structure of the SRS resource model.","marker":"[3]"},{"why":"The sniffer used to collect real PUSCH allocation traces from commercial networks, the traffic input that drives the multi-user SRS-triggering emulation.","marker":"[40]"},{"why":"Provides the angle-of-arrival estimation approach reused for the sensing half of the evaluation, against which tracking accuracy is computed.","marker":"[20]"},{"why":"Documents the tens-to-hundreds-of-millisecond latency of RRC reconfiguration, the basis for avoiding reconfiguration-based SRS adaptation.","marker":"[14]"},{"why":"The indoor-scenario guidelines used to set the default N=10 active users in the emulation.","marker":"[2]"},{"why":"The SNR-to-CQI mapping used to argue that the reported NMSE below 0.25 is too small to perturb modulation-and-coding selection.","marker":"[11]"}],"fun_headline_variants":["Fusing DMRS and SRS doubles 5G sensing users","Stitching 5G signals boosts sensing scalability","CARTS: cooperative CSI fusion for 5G ISAC","Two signals, one channel: 5G sensing improved","Adaptive stitching of 5G reference signals"],"cache_read_input_tokens":23296,"weakest_assumption_plain":"The scheme rests on the assumption that after each sub-band's linear phase slope is stripped off, the leftover phase differences between DMRS and SRS measurements can be absorbed by a single complex gain, an assumption that quietly goes stale if a user moves quickly between measurements or if rich multipath makes the phase profile much more than a straight line.","fun_headline_variants_meta":{"raw":{"variants":["Fusing DMRS and SRS doubles 5G sensing users","Stitching 5G signals boosts sensing scalability","CARTS: cooperative CSI fusion for 5G ISAC","Two signals, one channel: 5G sensing improved","Adaptive stitching of 5G reference signals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000517,"raw_usage":{"total_tokens":2574,"prompt_tokens":1076,"completion_tokens":1498,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":692,"completion_tokens_details":{"reasoning_tokens":1418}},"tokens_in":692,"tokens_out":1498,"duration_ms":13024,"temperature":1.0,"reasoning_tokens":1418,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:18:48.336543+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the stitched channel's NMSE against full-band SRS ground truth while systematically widening the time gap between a DMRS capture and the SRS reference capture, first at the paper's 3 km/h indoor setting and then at vehicular speeds: if NMSE stays well below the 0.25 threshold as the gap grows, the single-gain alignment is robust, while a sharp rise with gap size pinpoints temporal staleness as the weak link. A second decisive check is error growth as a function of stitched sub-band count, comparing two-sub-band stitching against six-sub-band stitching, where superlinear growth would confirm that errors accumulate across iterative stitching and would quantify how much of the reported accuracy depends on the number of SRS resources per user.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the direct stitching baseline: the prior Wi-Fi multi-channel alignment method whose peak-shifting approach accumulates errors on 5G sub-bands, which CARTS's slope-removal method is designed to beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The adaptive SRS scheduling scheme whose scalability limits, communication side effects, and RRC reconfiguration overhead motivate CARTS's switch to aperiodic DCI-triggered SRS."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The 5G RRC specification that fixes maxNrofSRS-TriggerStates-1, bounding the three resource sets that shape CARTS's six-resource SRS allocation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The UE capability specification giving maxNumberSRS-ResourcePerSet=2, which fixes the two-resources-per-set structure of the SRS resource model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The sniffer used to collect real PUSCH allocation traces from commercial networks, the traffic input that drives the multi-user SRS-triggering emulation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the angle-of-arrival estimation approach reused for the sensing half of the evaluation, against which tracking accuracy is computed."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the tens-to-hundreds-of-millisecond latency of RRC reconfiguration, the basis for avoiding reconfiguration-based SRS adaptation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The indoor-scenario guidelines used to set the default N=10 active users in the emulation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The SNR-to-CQI mapping used to argue that the reported NMSE below 0.25 is too small to perturb modulation-and-coding selection."}],"review_version":1}