RC-GAUSS: Receiver-Conditioned RF Synthesis

• RC-GAUSS is a receiver-conditioned 3D Gaussian splatting framework that decouples shared scene geometry from receiver-specific directional radiance for RF data synthesis.
• It employs a two-stage design that first learns geometry from a reference receiver and then applies global and local conditioning to modulate radiance for multi-receiver inference.
• The method significantly reduces training, inference, and storage costs while matching or improving performance compared to per-receiver 3DGS baselines across various RF modalities.

RC-GAUSS, presented in the RF synthesis literature as RxGS, is a receiver-conditioned 3D Gaussian splatting framework for radio-frequency data synthesis in which the target is a single model that predicts the received signal from both transmitter and receiver positions, $$\hat{S}(\mathbf{t}, \mathbf{r}) = f(\mathbf{t}, \mathbf{r}; \Theta)$$, rather than a separate model per receiver. Its defining premise is that scene geometry is receiver-independent whereas directional radiance is receiver-dependent. On that basis, RC-GAUSS decomposes learning into a geometry stage and a receiver-conditioned radiance stage, combines global and local conditioning mechanisms, and introduces a multi-receiver CUDA rasterizer that amortizes geometry-dependent work across all receivers in a scene. In the reported evaluations, the method matches or improves per-receiver 3DGS baselines for RSSI, generalizes to receivers unseen during training within the same scene, and reduces training, inference, and storage costs by large factors within a unified model (Yang et al., 22 May 2026).

1. Problem setting and signal model

RC-GAUSS addresses RF data synthesis as a mapping from transmitter and receiver positions to a received signal,

$$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$

The signal $S(\mathbf{t}, \mathbf{r})$ may be a scalar RSSI in dBm, a spatial spectrum image such as a $90 \times 360$ azimuth–elevation power map, or complex-valued CSI such as $H \in \mathbb{C}^{N_{sc}}$, modeled via real and imaginary parts. This formulation is explicitly broader than earlier single-receiver 3DGS RF models, which learn a separate parameter set $\Theta_j$ for each fixed receiver $\mathbf{r}_j$,

$$\hat{S}(\mathbf{t}, \mathbf{r}_j) = f(\mathbf{t}; \Theta_j).$$

That prior formulation enables prediction at arbitrary transmitters but only for one receiver at a time, so supporting $N$ receivers requires $N$ independent models (Yang et al., 22 May 2026).

The receiver-generalizable formulation instead seeks a single shared model,

$$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$0

The central modeling claim is that geometry should be shared across receivers, while the directional radiance attached to the Gaussian primitives must vary with receiver position. This separation is what allows RC-GAUSS to replace $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$1 receiver-specific models by one model and to interpolate to receivers unseen during training, provided they remain within the same scene.

A recurrent misconception is to treat receiver variation as merely another latent nuisance variable that can be absorbed into a single undifferentiated neural field. RC-GAUSS does not do this. Its formulation exposes receiver dependence explicitly in the radiance term, while keeping the underlying scene geometry fixed after the first stage. This suggests that the method is not simply parameter sharing across receivers; it is a structural factorization of the RF synthesis problem.

2. Gaussian scene representation and RF rendering

The scene is represented by a set of $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$2 3D Gaussian primitives,

$$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$3

where $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$4 is the Gaussian mean, $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$5 is the covariance, $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$6 denotes directional radiance coefficients on the sphere, and $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$7 is a direction-agnostic transmittance. For a transmitter at $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$8, the transmitter-to-Gaussian direction is

$$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$9

The rendering equation is written per ray in direction $S(\mathbf{t}, \mathbf{r})$0 from $S(\mathbf{t}, \mathbf{r})$1: $S(\mathbf{t}, \mathbf{r})$2 where $S(\mathbf{t}, \mathbf{r})$3 is the Gaussian overlap with the ray, $S(\mathbf{t}, \mathbf{r})$4 is directional radiance, and $S(\mathbf{t}, \mathbf{r})$5 is cumulative transmittance under front-to-back compositing. Aggregation across rays yields the synthesized signal $S(\mathbf{t}, \mathbf{r})$6 (Yang et al., 22 May 2026).

In implementation, RC-GAUSS rasterizes a spherical azimuth–elevation grid, performs per-tile depth sorting, and blends front-to-back with cumulative transmittance in an alpha-like form. The CUDA kernel accumulates

$S(\mathbf{t}, \mathbf{r})$7

where $S(\mathbf{t}, \mathbf{r})$8 is the tile-local Gaussian weight and $S(\mathbf{t}, \mathbf{r})$9 is the per-receiver radiance contribution. The geometry projection, tile bounds, depth sorting, and basis evaluation depend only on the transmitter. That observation underlies the later batching of multiple receivers in one rendering pass.

A plausible implication is that RC-GAUSS inherits the explicit-primitive advantages of 3DGS while adapting the compositing semantics to RF modalities that may be scalar, image-valued, or complex-valued. The framework is therefore described as radiance-representation-agnostic.

3. Geometry–radiance decomposition and physical factorization

The architecture is organized into two stages. In Stage I, RC-GAUSS learns geometry from a single reference receiver $90 \times 360$0 by jointly optimizing $90 \times 360$1: $90 \times 360$2 with $90 \times 360$3 chosen by modality: L1 for RSSI, L2/MSE for complex CSI, and a composite loss for spectrum. After this stage, geometry is frozen (Yang et al., 22 May 2026).

Stage II replaces the radiance with a receiver-conditioned form,

$90 \times 360$4

and optimizes shared base coefficients and conditioning parameters jointly across all receivers: $90 \times 360$5 The receiver-independent parameters $90 \times 360$6 remain fixed.

The physical interpretation is formalized as a last-segment factorization: $90 \times 360$7 with

$90 \times 360$8

Here $90 \times 360$9 and $H \in \mathbb{C}^{N_{sc}}$0 is the occlusion product over intervening scatterers. The factorization separates receiver-independent upstream radiance from receiver-dependent last-leg propagation.

This decomposition is the conceptual core of RC-GAUSS. It asserts that geometry can be identified from one receiver and then reused, whereas receiver-specific behavior should be injected only where propagation physics requires it. A common misunderstanding would be to read the first stage as learning a complete RF model for the reference receiver and the second stage as mere fine-tuning; the stated design instead treats the first stage as geometry acquisition and the second as receiver-conditioned radiance learning.

4. Receiver conditioning, affine modulation, and multi-receiver rendering

Receiver conditioning begins with a learned Fourier feature encoding,

$H \in \mathbb{C}^{N_{sc}}$1

where the $H \in \mathbb{C}^{N_{sc}}$2 are initialized with logarithmically spaced magnitudes and learned end-to-end. This encoding is intended to capture both room-scale path loss and wavelength-scale phase effects (Yang et al., 22 May 2026).

The global branch applies component-wise complex FiLM to the basis coefficients. For basis component $H \in \mathbb{C}^{N_{sc}}$3,

$H \in \mathbb{C}^{N_{sc}}$4

where $H \in \mathbb{C}^{N_{sc}}$5 encodes degree and order and $H \in \mathbb{C}^{N_{sc}}$6 is a learnable embedding. The modulation is

$H \in \mathbb{C}^{N_{sc}}$7

Because it acts on complex-valued coefficients, the branch can jointly scale amplitude and rotate phase. Its cost is $H \in \mathbb{C}^{N_{sc}}$8, independent of $H \in \mathbb{C}^{N_{sc}}$9.

The local branch models per-Gaussian residuals using receiver direction,

$\Theta_j$0

distance $\Theta_j$1, and occupancy-grid occlusion proxies

$\Theta_j$2

A shared MLP yields local complex FiLM parameters,

$\Theta_j$3

and the combined modulation is

$\Theta_j$4

Both branches are zero-initialized, so conditioning starts as the identity.

The reported ablation on BLE RSSI quantifies the role of each component. Full RxGS obtains MAE $\Theta_j$5 dBm; global-only gives $\Theta_j$6 dBm; local-only gives $\Theta_j$7 dBm; additive-only FiLM gives $\Theta_j$8 dBm; removing occlusion features gives $\Theta_j$9 dBm; and removing the two-stage design yields $\mathbf{r}_j$0 dBm. The paper states that both branches are required, affine FiLM is important, and occlusion helps (Yang et al., 22 May 2026).

Rendering across multiple receivers is accelerated by a multi-receiver CUDA rasterizer with

$\mathbf{r}_j$1

where the $\mathbf{r}_j$2-dimension indexes receivers. Projection, per-tile sorted Gaussian lists, and basis evaluation at $\mathbf{r}_j$3 are computed once and shared; per-receiver conditioned coefficients $\mathbf{r}_j$4 are then reduced in a single broadcast to obtain per-receiver signals $\mathbf{r}_j$5. The paper describes this as cutting rasterizer cost from $\mathbf{r}_j$6 to $\mathbf{r}_j$7. This suggests that the conditioning branch is intentionally lightweight relative to geometry-dependent rendering.

5. Optimization, datasets, and empirical behavior

Stage I uses task-specific losses, densification and pruning by gradient-based cloning and splitting, transmittance resets, and a progressive ramp of harmonic degree to stabilize geometry-first fitting. A typical schedule is 30K iterations with Adam. Stage II optimizes $\mathbf{r}_j$8 and the conditioning networks jointly across all receivers, uses an occupancy grid built once from frozen geometry, and typically runs for 60K–100K iterations depending on modality. The CSI evaluation uses the SNR metric

$\mathbf{r}_j$9

Typical hyperparameters include $$\hat{S}(\mathbf{t}, \mathbf{r}_j) = f(\mathbf{t}; \Theta_j).$$0, grid resolution $$\hat{S}(\mathbf{t}, \mathbf{r}_j) = f(\mathbf{t}; \Theta_j).$$1, basis degree up to 9 for spectrum and RSSI and up to 4 for CSI, Fourier bands $$\hat{S}(\mathbf{t}, \mathbf{r}_j) = f(\mathbf{t}; \Theta_j).$$2–6, MLP hidden dimension $$\hat{S}(\mathbf{t}, \mathbf{r}_j) = f(\mathbf{t}; \Theta_j).$$3–256, component embedding dimension $$\hat{S}(\mathbf{t}, \mathbf{r}_j) = f(\mathbf{t}; \Theta_j).$$4, and occupancy samples $$\hat{S}(\mathbf{t}, \mathbf{r}_j) = f(\mathbf{t}; \Theta_j).$$5 (Yang et al., 22 May 2026).

The evaluation covers three RF datasets:

Dataset	Modality	Setup
BLE RSSI	Real-world RSSI	$$\hat{S}(\mathbf{t}, \mathbf{r}_j) = f(\mathbf{t}; \Theta_j).$$6 receivers; 6,000 TX; indoor nursing home at 2.4 GHz
RFID spatial spectrum	Simulated spectrum	$$\hat{S}(\mathbf{t}, \mathbf{r}_j) = f(\mathbf{t}; \Theta_j).$$7 receivers; 5,089 TX; $$\hat{S}(\mathbf{t}, \mathbf{r}_j) = f(\mathbf{t}; \Theta_j).$$8 spectra at 2.4 GHz in an $$\hat{S}(\mathbf{t}, \mathbf{r}_j) = f(\mathbf{t}; \Theta_j).$$9 conference room
WiFi CSI	Complex CSI	$N$0 receivers; 6,000 TX; 26 subcarriers at 2.61 GHz

The baselines are NeRF$N$1, GRaF, RF-3DGS, GSRF, and WRF-GS+.

On BLE RSSI, the reported mean$N$2std MAE values are: RF-3DGS $N$3, Ours{RF-3DGS} $N$4; GSRF $N$5, Ours{GSRF} $N$6; WRF-GS+ $N$7, Ours{WRF-GS+} $N$8, where the receiver-conditioned version eliminates catastrophic outliers. On spectrum, PSNR is: NeRF$N$9 $N$0; GRaF $N$1; RF-3DGS $N$2, Ours{RF-3DGS} $N$3; GSRF $N$4, Ours{GSRF} $N$5; WRF-GS+ $N$6, Ours{WRF-GS+} $N$7. On CSI, SNR is: RF-3DGS $N$8, Ours{RF-3DGS} $N$9; GSRF $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$00, Ours{GSRF} $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$01; WRF-GS+ $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$02, Ours{WRF-GS+} $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$03. The stated conclusion is that RC-GAUSS matches or improves RSSI over per-receiver 3DGS baselines with a single shared model, and trails spectrum and CSI by $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$04–$$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$05 dB (Yang et al., 22 May 2026).

The unseen-receiver experiment is a central test of the method’s stated objective. In a 3-fold setup that trains on 14 receivers and tests on 21, BLE RSSI per-receiver baselines degrade to $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$06–$$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$07 dBm MAE on unseen receivers, while RC-GAUSS achieves $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$08 dBm and preserves seen performance at $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$09 dBm; NeRF$$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$10 obtains $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$11 dBm unseen MAE. On spectrum, per-receiver baselines drop by 4–6 dB, whereas RC-GAUSS obtains unseen PSNR $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$12 dB, near NeRF$$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$13 at $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$14 dB. The practical guidance section further states that unseen performance improves sharply up to $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$15 diverse training receivers and then saturates.

6. Efficiency, limitations, and position within wireless 3DGS

The efficiency claims are unusually explicit. Collapsing $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$16 per-receiver models into a single two-stage training run yields up to $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$17 speedup, with WRF-GS+ reported at $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$18 h $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$19 h for $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$20, and $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$21–$$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$22 for RF-3DGS and GSRF, reported as $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$23 h $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$24 h and $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$25 h $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$26 h. Multi-receiver batching yields up to $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$27 faster inference: GSRF $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$28 ms $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$29 ms, RF-3DGS $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$30 ms $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$31 ms, and WRF-GS+ $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$32 ms $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$33 ms, with the last case limited by the absence of a batched rasterizer variant. Storage is reduced from $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$34 checkpoints to one unified checkpoint, with $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$35 compression examples including GSRF $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$36 MB $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$37 MB, RF-3DGS $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$38 MB $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$39 MB, and WRF-GS+ $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$40 MB $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$41 MB (Yang et al., 22 May 2026).

The method’s assumptions are equally explicit. It assumes a static scene with fixed scatterers and material properties; it generalizes across receivers within the same scene but not across scenes; its conditioning is frequency-specific; and its occlusion path transmittance $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$42 is approximated from a frozen occupancy grid rather than exact EM ray tracing on the last segment. The reported failure cases include extreme occlusion changes, highly sparse receiver coverage during training such as $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$43, and distributions with very fine-grained spectrum structure, all of which may reduce unseen-receiver fidelity by $$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$44–$$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$45 dB relative to per-receiver fits.

In relation to prior RF synthesis methods, RC-GAUSS is positioned against per-receiver 3DGS models such as RF-3DGS, GSRF, and WRF-GS+, which do not expose receiver position to the model and therefore learn receiver-specific radiance. It is also contrasted with NeRF-style RF synthesis such as NeRF$$S: (\mathbf{t}, \mathbf{r}) \mapsto S(\mathbf{t}, \mathbf{r}), \qquad \mathbf{t}, \mathbf{r} \in \mathbb{R}^3.$$46 and GRaF, which can encode receiver position implicitly but incur higher computational cost and exhibit smoothness bias at unseen receivers. A related wireless scene representation is GAI-GS, which the literature describes as equivalent in objectives to RC-GAUSS in the sense that it also couples 3D Gaussian splatting with physically grounded propagation modeling, but does so through geometric algebra–based attention and joint spatial–electromagnetic tokenization rather than the two-stage receiver-conditioned radiance formulation of RxGS (Shen et al., 18 May 2026).

The broader significance of RC-GAUSS lies in how it reframes receiver variation. Rather than treating each receiver as a separate inverse problem or collapsing all dependence into a monolithic latent representation, it isolates receiver-conditioned propagation as a structured modulation of shared Gaussian primitives. This suggests a general template for RF scene models in which geometry is learned once, propagation is conditioned explicitly, and rendering work is amortized across many receiver queries.