CPRS: Channel Prediction-Based RS Allocation

• CPRS is a prediction-driven reference signal control paradigm that maps uplink CSI to optimal DM-RS configurations under strict standards.
• It utilizes deep learning models, including spatial-temporal transformers and 3D CNNs, to process CSI as image-like tensors and videos for end-to-end decision making.
• Experimental results demonstrate near-optimal throughput gains and significant BLER reductions by effectively trading off between pilot overhead and demodulation reliability.

Channel Prediction-Based Reference Signal Allocation (CPRS) denotes a prediction-driven reference-signal control paradigm in which channel evolution is used to choose a reference-signal configuration rather than treating channel prediction and pilot selection as separate stages. In its explicit formulation, CPRS is the learned mapping

$$\hat{p}^{*}=\mathcal{F}(\mathbf{H}_{\textnormal{UL}};\Theta)$$

from current-slot uplink CSI to the standards-compliant PDSCH DM-RS pattern for the next downlink slot in FDD massive MIMO, with the objective of maximizing throughput under the tradeoff between pilot overhead and demodulation reliability (Ryu et al., 15 Jul 2025). In a broader but partly inferential sense, CPRS also encompasses architectures that reuse existing uplink reference-signal observations, geometry-aware CSI prediction, or future reference-signal-strength forecasts to suppress, defer, or condition additional reference-signal procedures when prediction confidence is high (Gowgi et al., 2021).

1. Formal definition and optimization target

The canonical CPRS formulation is introduced for a massive MIMO-OFDM downlink with UE mobility and NR slot structure. The received signal at OFDM symbol $t$ and subcarrier $k$ is

$$y_{t,k}=\mathbf{h}_{\textnormal{DL},t,k}\,\mathbf{v}_{t,k}\,x_{t,k}+\mathbf{w}_{t,k},$$

where $$\mathbf{h}_{\textnormal{DL},t,k}\in\mathbb{C}^{N_{\textnormal{R}}\times N_{\textnormal{T}}}$$ is the downlink channel matrix, $$\mathbf{v}_{t,k}\in\mathbb{C}^{N_{\textnormal{T}}\times 1}$$ is the precoding vector, $x_{t,k}\in\mathbb{C}$ is the transmit symbol, and $$\mathbf{w}_{t,k}\in\mathbb{C}^{N_{\textnormal{R}}\times 1}$$ is noise. Symbol-level downlink CSI is collected as

$$\mathbf{H}_{\textnormal{DL},t} = \left[ \mathbf{h}_{\textnormal{DL},t,1}\; \mathbf{h}_{\textnormal{DL},t,2}\; \cdots\; \mathbf{h}_{\textnormal{DL},t,N_{\textnormal{C}}} \right] \in \mathbb{C}^{N_{\textnormal{R}}\times N_{\textnormal{T}}\times N_{\textnormal{C}}},$$

and slot-level CSI is

$$\mathbf{H}_{\textnormal{DL}}=\left\{\mathbf{H}_{\textnormal{DL},t}\right\}_{t=1}^{14}.$$

The BS observes only uplink CSI at uplink DM-RS positions

$t$0

and does not use explicit downlink CSI feedback when making the next-slot DM-RS decision (Ryu et al., 15 Jul 2025).

The baseline prediction problem is

$t$1

but CPRS replaces the disjoint sequence “predict CSI, then optimize DM-RS” with a direct classifier over standards-allowed DM-RS choices. The optimization target is the throughput-maximizing allocation

$t$2

where $t$3 is the code rate, $t$4 is the total number of bits transmitted in one slot, $t$5 is the number of DM-RS symbols in allocation $t$6, and $t$7. CPRS is then defined as the end-to-end mapping

$t$8

trained with cross-entropy to predict the ground-truth optimal DM-RS class (Ryu et al., 15 Jul 2025).

This formulation makes the central design tradeoff explicit. More DM-RS symbols improve channel tracking and demodulation reliability under mobility and channel aging, but consume data-plane resources that would otherwise carry payload. Fewer DM-RS symbols increase payload symbols, but can raise BLER when temporal channel variation between pilots becomes large. CPRS is therefore not merely a channel prediction problem; it is a throughput-maximizing control problem in which prediction is instrumental.

2. Standards-compliant action space and learning architecture

The standards-compliant aspect of CPRS is central. The optimization space is restricted to 3GPP NR TS 38.211 PDSCH DM-RS symbol-placement options within a 14-symbol slot. The formulation considers Type-A PDSCH DM-RS only; Type-B is excluded because it applies to a special mini-slot scheduling case with fewer patterns. Only symbol-time placement is optimized, while frequency-domain flexibility is effectively fixed through Kronecker placement because Type 1 and Type 2 patterns provide little flexibility in this setting. Under the simplifying assumption $t$9, the admissible action space contains $k$0 candidate allocations, with $k$1 and $k$2 in the underlying standard description (Ryu et al., 15 Jul 2025).

The proposed models are two end-to-end classifiers. In CPRS (ViViT), each uplink CSI tensor

$k$3

is converted into a real-valued image-like tensor

$k$4

by flattening the antenna dimensions and separating real and imaginary parts. The sequence over $k$5 is then treated as a short video. The ViViT pipeline partitions each $k$6 into non-overlapping patches,

$k$7

uses a spatial transformer to capture spatial/frequency structure, uses a temporal transformer to capture channel evolution, and maps the pooled representation to one of the 13 DM-RS classes. The attention operator is

$k$8

The reported ViViT hyperparameters are tubelet embedding patch size $k$9, 8 encoder layers, and 16 attention heads. CPRS (CNN) uses the same input-output task with three Conv3D layers having filter counts 16, 32, and 64, each followed by batch normalization and 3D max pooling (Ryu et al., 15 Jul 2025).

Offline training labels are produced by exhaustive evaluation of every candidate $$y_{t,k}=\mathbf{h}_{\textnormal{DL},t,k}\,\mathbf{v}_{t,k}\,x_{t,k}+\mathbf{w}_{t,k},$$0 using Sionna link-level simulation, and the online BS-side decision is made from current-slot uplink CSI snapshots to configure the next downlink slot. The reported experimental environment is the NVIDIA Sionna ray-tracing scene simple_street_canyon_with_cars, with an $$y_{t,k}=\mathbf{h}_{\textnormal{DL},t,k}\,\mathbf{v}_{t,k}\,x_{t,k}+\mathbf{w}_{t,k},$$1 BS antenna array, a $$y_{t,k}=\mathbf{h}_{\textnormal{DL},t,k}\,\mathbf{v}_{t,k}\,x_{t,k}+\mathbf{w}_{t,k},$$2 UE antenna array, FDD in NR band n1, uplink carrier 1.95 GHz, downlink carrier 2.14 GHz, 100 MHz bandwidth, 64 subcarriers, and 14 symbols per slot. The dataset contains 6,000 slots with training/validation/test split 8:1:1. The reported ViViT complexity is $$y_{t,k}=\mathbf{h}_{\textnormal{DL},t,k}\,\mathbf{v}_{t,k}\,x_{t,k}+\mathbf{w}_{t,k},$$3; with $$y_{t,k}=\mathbf{h}_{\textnormal{DL},t,k}\,\mathbf{v}_{t,k}\,x_{t,k}+\mathbf{w}_{t,k},$$4, $$y_{t,k}=\mathbf{h}_{\textnormal{DL},t,k}\,\mathbf{v}_{t,k}\,x_{t,k}+\mathbf{w}_{t,k},$$5, and $$y_{t,k}=\mathbf{h}_{\textnormal{DL},t,k}\,\mathbf{v}_{t,k}\,x_{t,k}+\mathbf{w}_{t,k},$$6, the paper estimates about $$y_{t,k}=\mathbf{h}_{\textnormal{DL},t,k}\,\mathbf{v}_{t,k}\,x_{t,k}+\mathbf{w}_{t,k},$$7 FLOPs and a theoretical inference time around $$y_{t,k}=\mathbf{h}_{\textnormal{DL},t,k}\,\mathbf{v}_{t,k}\,x_{t,k}+\mathbf{w}_{t,k},$$8 under INT8 quantization on NVIDIA A100 hardware (Ryu et al., 15 Jul 2025).

3. Broader predictive lineage and adjacent formulations

CPRS in the narrow DM-RS sense has important antecedents in prediction-driven mobility, signal-strength forecasting, geometry-based CSI inference, and uncertainty-aware scheduling. These works do not all allocate reference signals directly, but they clarify which predictive quantities can replace or condition explicit measurement procedures.

A first line of work uses existing uplink reference-signal-derived channel observations to infer whether a secondary carrier is usable. In an inter-frequency handover setting, the serving cell records the uplink CIR at carrier $$y_{t,k}=\mathbf{h}_{\textnormal{DL},t,k}\,\mathbf{v}_{t,k}\,x_{t,k}+\mathbf{w}_{t,k},$$9, extracts a feature vector $$\mathbf{h}_{\textnormal{DL},t,k}\in\mathbb{C}^{N_{\textnormal{R}}\times N_{\textnormal{T}}}$$0, and predicts the binary target

$$\mathbf{h}_{\textnormal{DL},t,k}\in\mathbb{C}^{N_{\textnormal{R}}\times N_{\textnormal{T}}}$$1

so that the learned mapping is essentially $$\mathbf{h}_{\textnormal{DL},t,k}\in\mathbb{C}^{N_{\textnormal{R}}\times N_{\textnormal{T}}}$$2. The paper frames this as secondary-carrier existence prediction from uplink reference signals already available for other BS functions, especially beamforming decisions. It uses Random Forest as the main classifier, compares against a Gaussian-likelihood MAP benchmark, applies hybrid sampling to balance the class ratio to one, and evaluates with six-fold cross validation and AUROC. Although it does not optimize reference-signal allocation itself, it demonstrates that serving-carrier uplink channel observations can substitute for explicit UE-side inter-frequency measurements in a decision-relevant thresholded task (Gowgi et al., 2021).

A second line of work predicts reference-signal strength sequences rather than CSI matrices. Channel-Diff formulates a UE trajectory $$\mathbf{h}_{\textnormal{DL},t,k}\in\mathbb{C}^{N_{\textnormal{R}}\times N_{\textnormal{T}}}$$3 and predicts an aligned RSRP series

$$\mathbf{h}_{\textnormal{DL},t,k}\in\mathbb{C}^{N_{\textnormal{R}}\times N_{\textnormal{T}}}$$4

using a physics-informed conditional diffusion model. Its physical priors separate large-scale and small-scale attenuation through large-scale propagation modelling, shadow-occlusion modelling, multipath propagation modelling, and a microenvironment feature extraction network. The framework uses a teacher-student two-stage training scheme with noise prior guidance and conditions on a 10-dimensional sequence of network parameters together with multi-attribute urban maps consisting of ground elevation and building-height distribution. This target is scalar RSRP rather than full CSI, but it is directly relevant wherever reference-signal planning is driven by predicted coverage evolution and blockage likelihood (Qi et al., 25 Dec 2025).

A third line of work infers geometry from noisy uplink CSI and uses it for downlink CSI prediction and fusion. The geometry-based formulation models the channel as LoS plus specular single- and double-bounce reflections, jointly estimates user position, velocity, and an environment map encoded by master virtual anchors, and predicts CSI as

$$\mathbf{h}_{\textnormal{DL},t,k}\in\mathbb{C}^{N_{\textnormal{R}}\times N_{\textnormal{T}}}$$5

If measured CSI is also available, it forms the fused estimate

$$\mathbf{h}_{\textnormal{DL},t,k}\in\mathbb{C}^{N_{\textnormal{R}}\times N_{\textnormal{T}}}$$6

which makes the relative trust in prediction and measurement explicit through covariance structure. This work is not a pilot scheduler, but it provides uncertainty quantities such as prediction covariance, measurement covariance, and particle-filter state covariance that are natural control inputs for any adaptive reference-signal policy (Deutschmann et al., 22 Mar 2025).

A fourth line of work uses predicted channels and prediction uncertainty directly in radio-resource decisions under imperfect CSI. In federated learning over wireless, future channels are predicted by Gaussian process regression,

$$\mathbf{h}_{\textnormal{DL},t,k}\in\mathbb{C}^{N_{\textnormal{R}}\times N_{\textnormal{T}}}$$7

with posterior variance

$$\mathbf{h}_{\textnormal{DL},t,k}\in\mathbb{C}^{N_{\textnormal{R}}\times N_{\textnormal{T}}}$$8

and RB allocations both exploit predicted channel quality and value uncertain links because scheduling an RB also acquires a new CSI sample. This is not reference-signal allocation in the literal sense, but it shows how prediction uncertainty can be monetized as an allocation variable rather than treated merely as a nuisance parameter (Wadu et al., 2020).

4. Empirical evidence

The direct CPRS evidence comes from the standards-compliant DM-RS allocation study. CPRS (ViViT) achieves near-perfect classification accuracy across tested SNRs from $$\mathbf{h}_{\textnormal{DL},t,k}\in\mathbb{C}^{N_{\textnormal{R}}\times N_{\textnormal{T}}}$$9 dB to $$\mathbf{v}_{t,k}\in\mathbb{C}^{N_{\textnormal{T}}\times 1}$$0 dB, typically $$\mathbf{v}_{t,k}\in\mathbb{C}^{N_{\textnormal{T}}\times 1}$$1–$$\mathbf{v}_{t,k}\in\mathbb{C}^{N_{\textnormal{T}}\times 1}$$2. In throughput terms, its error relative to the oracle optimum is between $$\mathbf{v}_{t,k}\in\mathbb{C}^{N_{\textnormal{T}}\times 1}$$3 and $$\mathbf{v}_{t,k}\in\mathbb{C}^{N_{\textnormal{T}}\times 1}$$4, with up to $$\mathbf{v}_{t,k}\in\mathbb{C}^{N_{\textnormal{T}}\times 1}$$5 throughput gain over the second-best benchmark and up to $$\mathbf{v}_{t,k}\in\mathbb{C}^{N_{\textnormal{T}}\times 1}$$6 gain in the saturation region under 64-QAM. By contrast, the disjoint predictor-plus-optimizer baseline degrades sharply because errors accumulate across stages, and under 256-QAM it yields $$\mathbf{v}_{t,k}\in\mathbb{C}^{N_{\textnormal{T}}\times 1}$$7, causing zero throughput (Ryu et al., 15 Jul 2025).

The inter-frequency secondary-carrier study provides evidence that uplink reference-signal-derived features can be highly predictive even when the target is only a coarse binary usability event. On the real-world-derived dataset D1, mean AUROC is 0.93, while on the simulated D2 and D3 scenarios mean AUROC is reported as close to one across all scenarios, including FR1-to-FR2 prediction. This supports the narrower claim that already-available uplink channel estimates can separate usable from unusable secondary-carrier opportunities with high fidelity (Gowgi et al., 2021).

The RSRP trajectory-prediction results show that predictive reference-signal strength modeling can substantially outperform formula-based and purely data-driven baselines. On RSRP-CPGMCM, Channel-Diff achieves JSD $$\mathbf{v}_{t,k}\in\mathbb{C}^{N_{\textnormal{T}}\times 1}$$8, NRMSE $$\mathbf{v}_{t,k}\in\mathbb{C}^{N_{\textnormal{T}}\times 1}$$9, and MAE $x_{t,k}\in\mathbb{C}$0, while the second-best baseline GBDT+DNN has JSD $x_{t,k}\in\mathbb{C}$1, NRMSE $x_{t,k}\in\mathbb{C}$2, and MAE $x_{t,k}\in\mathbb{C}$3; the reported average improvement is $x_{t,k}\in\mathbb{C}$4. On RSRP-Image, Channel-Diff achieves JSD $x_{t,k}\in\mathbb{C}$5, NRMSE $x_{t,k}\in\mathbb{C}$6, and MAE $x_{t,k}\in\mathbb{C}$7, compared with EFEM at JSD $x_{t,k}\in\mathbb{C}$8, NRMSE $x_{t,k}\in\mathbb{C}$9, and MAE $$\mathbf{w}_{t,k}\in\mathbb{C}^{N_{\textnormal{R}}\times 1}$$0, for an average improvement of $$\mathbf{w}_{t,k}\in\mathbb{C}^{N_{\textnormal{R}}\times 1}$$1. Approximately $$\mathbf{w}_{t,k}\in\mathbb{C}^{N_{\textnormal{R}}\times 1}$$2 of predictions on RSRP-CPGMCM and $$\mathbf{w}_{t,k}\in\mathbb{C}^{N_{\textnormal{R}}\times 1}$$3 on RSRP-Image have absolute error below the 3GPP TS 38.133 measurement-accuracy threshold of $$\mathbf{w}_{t,k}\in\mathbb{C}^{N_{\textnormal{R}}\times 1}$$4 dB (Qi et al., 25 Dec 2025).

The geometry-based CSI study shows that prediction can outperform or stabilize conventional reciprocity-based operation in regimes where pilot quality or timeliness is poor. Using the stochastic likelihood, the reported positioning accuracy after convergence is 6.37 cm horizontal RMSE and 4.92 cm vertical RMSE, whereas a LoS-only model degrades to 41 cm and 8.4 cm. For beamforming, at SNR $$\mathbf{w}_{t,k}\in\mathbb{C}^{N_{\textnormal{R}}\times 1}$$5 dB, reciprocity beamforming incurs a loss of at least $$\mathbf{w}_{t,k}\in\mathbb{C}^{N_{\textnormal{R}}\times 1}$$6 dB relative to perfect CSI, while the geometry-based beamformer given predicted CSI outperforms reciprocity beamforming by $$\mathbf{w}_{t,k}\in\mathbb{C}^{N_{\textnormal{R}}\times 1}$$7 dB on average. Under mobility, future predicted CSI incurs only a marginal efficiency loss compared with the severe loss from outdated CSI (Deutschmann et al., 22 Mar 2025).

The uncertainty-aware scheduling study shows a different but related resource-allocation gain from replacing dedicated CSI measurement with prediction. The abstract reports up to $$\mathbf{w}_{t,k}\in\mathbb{C}^{N_{\textnormal{R}}\times 1}$$8 reduction in loss of accuracy compared to state-of-the-art client scheduling and RB allocation methods; in the body, QAW-GPR outperforms QAW by reducing loss by $$\mathbf{w}_{t,k}\in\mathbb{C}^{N_{\textnormal{R}}\times 1}$$9 under imperfect CSI. The stated reason is that prediction allows utilization of all $$\mathbf{H}_{\textnormal{DL},t} = \left[ \mathbf{h}_{\textnormal{DL},t,1}\; \mathbf{h}_{\textnormal{DL},t,2}\; \cdots\; \mathbf{h}_{\textnormal{DL},t,N_{\textnormal{C}}} \right] \in \mathbb{C}^{N_{\textnormal{R}}\times N_{\textnormal{T}}\times N_{\textnormal{C}}},$$0 RBs for scheduling clients, whereas the perfect-CSI comparator allocates one RB for CSI measurements (Wadu et al., 2020).

5. Operational interpretation for reference-signal control

In the direct DM-RS formulation, CPRS is a BS-side configuration function that uses current-slot uplink CSI snapshots to select the next-slot downlink DM-RS pattern from a standards-compliant action space. The decision rule is therefore slot-local, feedback-light, and throughput-oriented. It does not invent new pilot structures; it chooses among existing NR Type-A symbol placements, so the method changes which allowed OFDM symbols in the next slot carry PDSCH DM-RS, not the overall slot structure or the general DM-RS design rules (Ryu et al., 15 Jul 2025).

The broader literature suggests several control patterns that fit naturally under a wider CPRS interpretation. The inter-frequency prediction results suggest that if a serving BS can infer secondary-carrier usability from uplink reference-signal-derived features, explicit UE inter-frequency measurement occasions may be suppressed for low-value candidates and triggered only when uncertainty is high. This suggestion is indirect, because the paper itself does not implement adaptive reference-signal allocation or a closed-loop overhead-accuracy policy (Gowgi et al., 2021).

The RSRP-forecasting framework suggests a trajectory-aware control loop in which predicted future weak-signal intervals, blockage likelihood, and stable unblocked intervals modulate how aggressively the network uses reference signals. A plausible implication is that falling predicted RSRP along a route could justify denser CSI-RS or SSB support, more conservative handover preparation, or additional sounding, whereas a stable and unobstructed forecast could justify reduced sounding or lower pilot density. Those allocation actions are not explicitly implemented in the reported experiments; the demonstrated result is sequence prediction of scalar RSRP under physics-informed conditioning (Qi et al., 25 Dec 2025).

The geometry-based prediction-and-fusion formulation suggests a more explicit confidence-control strategy. The quantities $$\mathbf{H}_{\textnormal{DL},t} = \left[ \mathbf{h}_{\textnormal{DL},t,1}\; \mathbf{h}_{\textnormal{DL},t,2}\; \cdots\; \mathbf{h}_{\textnormal{DL},t,N_{\textnormal{C}}} \right] \in \mathbb{C}^{N_{\textnormal{R}}\times N_{\textnormal{T}}\times N_{\textnormal{C}}},$$1, $$\mathbf{H}_{\textnormal{DL},t} = \left[ \mathbf{h}_{\textnormal{DL},t,1}\; \mathbf{h}_{\textnormal{DL},t,2}\; \cdots\; \mathbf{h}_{\textnormal{DL},t,N_{\textnormal{C}}} \right] \in \mathbb{C}^{N_{\textnormal{R}}\times N_{\textnormal{T}}\times N_{\textnormal{C}}},$$2, and $$\mathbf{H}_{\textnormal{DL},t} = \left[ \mathbf{h}_{\textnormal{DL},t,1}\; \mathbf{h}_{\textnormal{DL},t,2}\; \cdots\; \mathbf{h}_{\textnormal{DL},t,N_{\textnormal{C}}} \right] \in \mathbb{C}^{N_{\textnormal{R}}\times N_{\textnormal{T}}\times N_{\textnormal{C}}},$$3 provide direct measures of prediction uncertainty, measurement uncertainty, and state uncertainty. This suggests a CPRS policy in which low prediction covariance and low state uncertainty support reduced reference signaling, while large uncertainty or persistent model mismatch triggers denser pilots or fresh measurements. The paper itself provides the uncertainty-aware predictor and fusion equations rather than a scheduler (Deutschmann et al., 22 Mar 2025).

The scheduling-under-uncertainty study suggests a complementary principle: prediction uncertainty need not merely penalize a link; it can be treated as information value. In that work, the posterior variance $$\mathbf{H}_{\textnormal{DL},t} = \left[ \mathbf{h}_{\textnormal{DL},t,1}\; \mathbf{h}_{\textnormal{DL},t,2}\; \cdots\; \mathbf{h}_{\textnormal{DL},t,N_{\textnormal{C}}} \right] \in \mathbb{C}^{N_{\textnormal{R}}\times N_{\textnormal{T}}\times N_{\textnormal{C}}},$$4 is explicitly rewarded in the resource-allocation objective, because transmission over an uncertain link both serves the current task and acquires a new CSI sample. A plausible CPRS implication is that scarce reference-signal resources should be concentrated where predictive uncertainty is largest and where an additional observation has high future utility (Wadu et al., 2020).

6. Limits, misconceptions, and open directions

A common misconception is that CPRS necessarily means full CSI prediction. The direct DM-RS paper indeed uses uplink CSI tensors as inputs, but several adjacent works use much coarser targets: binary secondary-carrier existence in inter-frequency handover and scalar RSRP sequences along user trajectories. These targets are decision-relevant, but they are not equivalent to predicting delay-angle CSI, subband selectivity, rank, or beam-domain coefficients (Gowgi et al., 2021).

A second misconception is that prediction-driven reference-signal control is equivalent to reciprocity. The inter-frequency handover predictor does not rely on strict physical reciprocity between uplink and downlink or across $$\mathbf{H}_{\textnormal{DL},t} = \left[ \mathbf{h}_{\textnormal{DL},t,1}\; \mathbf{h}_{\textnormal{DL},t,2}\; \cdots\; \mathbf{h}_{\textnormal{DL},t,N_{\textnormal{C}}} \right] \in \mathbb{C}^{N_{\textnormal{R}}\times N_{\textnormal{T}}\times N_{\textnormal{C}}},$$5 and $$\mathbf{H}_{\textnormal{DL},t} = \left[ \mathbf{h}_{\textnormal{DL},t,1}\; \mathbf{h}_{\textnormal{DL},t,2}\; \cdots\; \mathbf{h}_{\textnormal{DL},t,N_{\textnormal{C}}} \right] \in \mathbb{C}^{N_{\textnormal{R}}\times N_{\textnormal{T}}\times N_{\textnormal{C}}},$$6; it relies on learned statistical correlation mediated by UE location, environment, blockage structure, and deployment geometry. Similarly, the geometry-based predictor uses explicit physical structure and Bayesian fusion rather than direct reciprocity alone. CPRS therefore spans both reciprocity-adjacent and non-reciprocal predictive mechanisms (Deutschmann et al., 22 Mar 2025).

The direct standards-compliant CPRS proposal also has narrow assumptions. It studies a single-user downlink style formulation, excludes HARQ, considers only Type-A PDSCH DM-RS, largely ignores frequency-domain DM-RS flexibility, reduces the action space to 13 allocations under $$\mathbf{H}_{\textnormal{DL},t} = \left[ \mathbf{h}_{\textnormal{DL},t,1}\; \mathbf{h}_{\textnormal{DL},t,2}\; \cdots\; \mathbf{h}_{\textnormal{DL},t,N_{\textnormal{C}}} \right] \in \mathbb{C}^{N_{\textnormal{R}}\times N_{\textnormal{T}}\times N_{\textnormal{C}}},$$7, and demonstrates performance in one Sionna ray-tracing environment over speeds $$\mathbf{H}_{\textnormal{DL},t} = \left[ \mathbf{h}_{\textnormal{DL},t,1}\; \mathbf{h}_{\textnormal{DL},t,2}\; \cdots\; \mathbf{h}_{\textnormal{DL},t,N_{\textnormal{C}}} \right] \in \mathbb{C}^{N_{\textnormal{R}}\times N_{\textnormal{T}}\times N_{\textnormal{C}}},$$8–$$\mathbf{H}_{\textnormal{DL},t} = \left[ \mathbf{h}_{\textnormal{DL},t,1}\; \mathbf{h}_{\textnormal{DL},t,2}\; \cdots\; \mathbf{h}_{\textnormal{DL},t,N_{\textnormal{C}}} \right] \in \mathbb{C}^{N_{\textnormal{R}}\times N_{\textnormal{T}}\times N_{\textnormal{C}}},$$9 km/h. The reported inference feasibility is theoretical rather than an end-to-end measured real-time result (Ryu et al., 15 Jul 2025).

The inter-frequency predictor is limited in a different way. It does not optimize reference-signal allocation itself, has no explicit cost function trading pilot load against prediction quality, and uses coarse binary existence labels. In D1, the secondary-carrier labels are artificial and location-derived rather than directly measured inter-frequency RSRP, RSRQ, or SINR; many input features also include deployment geometry and nearest-micro-cell information, which may affect generalization when topology changes (Gowgi et al., 2021).

The RSRP diffusion framework is substantial as a predictive front-end but not a turnkey CPRS scheduler. It predicts RSRP only, requires accurate trajectory or location information together with environment maps, and the reported model is large, with 66.5M parameters, 400 diffusion steps, 12 Transformer blocks, hidden size 256, and attention-based MFEN. The paper itself notes UE location accuracy as a future challenge and does not evaluate actual reference-signal allocation, pilot-density control, or online scheduling latency (Qi et al., 25 Dec 2025).

The geometry-based approach likewise remains a strong backend rather than a full RS-allocation solution. It depends on map learning through master virtual anchors, neglects diffuse multipath in the inference model, uses a fixed model order in the reported experiments, assumes synchronized and phase-calibrated anchors, and does not provide runtime or per-slot scheduling benchmarks. It is therefore best viewed as an uncertainty-aware prediction and fusion engine that could support CPRS, not as a complete CPRS algorithm (Deutschmann et al., 22 Mar 2025).

The clearest open direction stated in the direct CPRS work is extension of prediction-driven optimization to reference signals beyond DM-RS, practical extensions to beyond-5G scenarios, joint design with beamforming, and broader mobility and channel conditions. The adjacent literature suggests that such extensions may benefit from combining low-complexity uplink-reference-signal reuse, physics-informed signal-strength forecasting, geometry-based uncertainty quantification, and uncertainty-aware allocation objectives rather than relying on any single predictive modality alone (Ryu et al., 15 Jul 2025).