RFWeave: Learning Wireless Propagation Across Scenes and Transceiver Configurations

arXiv:2610.16656 · eess.SP, cs.LG · Submitted 2026-10-12 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: Today's paper: "RFWeave: Learning Wireless Propagation Across Scenes and Transceiver Configurations".

Jane: The gist: RFWeave introduces a learned wireless propagation model designed for joint generalization across scenes, transceiver locations, and transceiver configurations.

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we're looking at this paper 'RFWeave: Learning Wireless Propagation Across Scenes and Transceiver Configurations'. The title itself says what it does, which is learning wireless propagation across scenes and transceiver configurations. It sounds like it's tackling a big problem in communication systems where you need one model to handle many different environments and many different setups all at once.

Jane: Right, so when you hear 'across scenes and transceiver configurations', what does that mean for the person listening? They probably think about how hard it is to build a perfect prediction system that works everywhere, not just in one specific place or with one specific antenna setup.

Lu: Exactly. The core idea they are pushing is that we can get a single model to handle things like arbitrary TX and RX locations, varying array sizes, different receiver orientations, and even different transmit beams on scenes we haven't seen before. That’s the ambition here <ref:2610.16656#pg2>.

Meng: I mean, that sounds really cool in theory for cross-scene work. But practically speaking, what does a single model actually achieve without breaking at scale or with real-world data? That's where my skepticism kicks in <ref:2610.16656#pg3>.

Lalam: From the literature, I see that existing cross-scene methods usually assume a common transceiver configuration across different scenes, which means they don't really capture how the antenna setup itself affects the physics <ref:2610.16656#pg3>.

The paper's summary: Tom: Okay, so the paper says RFWeave models the receive-side response by looking at spatial covariance and received power instead of trying to reproduce a full instantaneous complex channel. That’s a big shift in how they approach the modeling problem <ref:2610.16656#pg1>.

Jane: Instead of trying to map every single tiny detail of the channel, they are learning a mapping f-theta(X, Q, C) that takes scene geometry X, query locations Q, and transceiver configuration C and outputs the normalized receive-side spatial covariance Rb and received power Pb <ref:2610.16656#pg1>.

Lu: The key mechanism they use to get this is tokenization. They break down the scene into material-annotated point clouds, and they also create tokens for transceiver geometry at the element, array, and link levels <ref:2610.16656#pg2>.

Meng: Tokenization sounds like a lot of work upfront. What’s the payoff? How does that tokenization actually help them generalize when things change, like moving from one scene to another?

Lalam: The scene tokens S are computed once per scene and then reused across all queries <ref:2610.16656#pg2>. Then, they use self-attention on those tokens, and that representation Sc gets adapted by attending to the transceiver tokens T through cross-attention to get a scene representation that fits the current link geometry <ref:2610.16656#pg3>.

The paper's improvements: Tom: Now they talk about what they improved. They introduced variable-length transceiver tokens to handle different array structures, and then they aggregate pairwise geometry between TX and RX elements into propagation tokens P <ref:2610.16656#pg2>.

Jane: So that means even if you query a system with a totally different number of antennas or an oddly arranged array, the model has the tokens to understand those physical arrangements <ref:2610.16656#pg2>. That addresses the issue of varying array sizes directly.

Lu: And they don't just look at one pair of elements; they compute a tensor E = eij Nt,Nr i=1,j=one which is the pairwise geometry between individual TX and RX elements <ref:2610.16656#pg3>. Then they aggregate those to form element-indexed propagation tokens P <ref:2610.16656#pg3>.

Meng: That sounds computationally heavy, generating all those pairwise features. What’s the trade-off? Is it worth the complexity just to get that better representation?

Lalam: The training objective has three parts: a loss for the spatial spectrum induced by the covariance, a loss for received power, and then losses for configuration-dependent variations along receiver orientation and transmit beam <ref:2610.16656#pg3>. That’s how they supervise the model to pay attention to those factors.

Conclusion: Tom: So, in summary, RFWeave is a learned wireless propagation model that aims for joint generalization across scenes, locations, and configurations by using these scene and transceiver tokens <ref:2610.16656#pg2>.

Jane: The main implication here is that we can predict the AoA spatial spectrum and received power using just one model for any query parameters you throw at it <ref:2610.16656#pg2>. It means less need to train separate models for every scene or every array size <ref:2610.16656#pg2>.

Lu: The work shows that the way they structure the model, particularly how they condition the scene tokens S through self-attention and then cross-attend to transceiver tokens T, is what allows it to adapt its scene representation Sc for every link and array geometry <ref:2610.16656#pg3>.

Meng: For deployment, I'm looking at those training losses—Lspec, Lpower, and the configuration variation losses—they are designed to keep things stable when you push the model to its limits with different beams or orientations <ref:2610.16656#pg3>.

Lalam: And what I've seen is that for the balanced setting where lambda rx and lambda beam are both one, they get the best results on five out of six reported metrics <ref:2610.16656#pg3>.

Tom: So, to wrap up, RFWeave is a model that generalizes across scenes, locations, and configurations by learning how to predict the spatial spectrum and received power using a single model <ref:2610.16656#pg2>.

Jane: It’s a solid piece of work because it adapts to measured data from the DICHASUS dataset, showing the highest spatial spectrum SSIM at every budget and performing best in low-data regimes <ref:2610.16656#pg3>.

Lu: And that adaptation to real measurements is important because it shows the representation isn't just theoretical; it actually works on data <ref:2610.16656#pg3>.

Meng: I just hope this complexity doesn't make the inference time too slow for production use, because running all those pairwise geometry calculations has to be efficient <ref:2610.16656#pg3>.

Lalam: We need to keep an eye on how they handle those configuration changes in a real deployment scenario, because that’s the trickiest part of deploying any generalized model <ref:2610.16656#pg3>.

Shurui Xu, Xiao Ye Ye, Yijie Bian, Junjie Ai, Yuxuan Chen, Weicong Chen, Wankai Tang

Southeast University · The Hong Kong University of Science and Technology

eess.SP, cs.LG

Submitted: 2026-10-12

Updated: 2026-10-12

Comments: 20 pages, 12 figures, 12 tables, including appendices

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

The gist: The gist: RFWeave introduces a learned wireless propagation model designed for joint generalization across scenes, transceiver locations, and transceiver configurations.

Key concepts

Scene Tokenizer: This component represents a scene as a material-annotated point cloud. Each point includes its position, surface normal, and material properties like permittivity. It uses a hierarchical structure to downsample the points, resulting in reusable 'scene tokens' that capture the physical characteristics of the environment.
Transceiver Tokenizer: This tokenizer captures transceiver geometry at element, array, and link levels. It treats each antenna element as a token based on its position normalized by wavelength. It also includes tokens for the TX and RX array centers and their displacement to fully describe the system's physical setup.
Wireless Propagation Modeling: This involves computing pairwise geometry between individual TX and RX elements to create an element-indexed propagation tensor. These features are aggregated along both dimensions to form propagation tokens, which are then refined by attending to scene and transceiver context for prediction.
Training Objectives: The model is trained using a multi-task loss function that supervises the spatial spectrum induced by the covariance rather than just the factor. It also includes losses for configuration-dependent variations along receiver orientation and transmit beam axes, ensuring robust generalization.

Summary

The gist: RFWeave introduces a learned wireless propagation model designed for joint generalization across scenes, transceiver locations, and transceiver configurations.

Model Overview

RFWeave is a learned wireless propagation model designed to generalize across scenes, transceiver locations, and transceiver configurations It models the receive-side response through spatial covariance and received power rather than reproducing a full instantaneous complex channel The goal is to learn a mapping fθ(X, Q, C) 7→ Rb, Pb, (1) where Rb denotes the normalized receive-side spatial covariance and Pb denotes the received power

Tokenization Strategies

The model employs two main types of tokens to capture scene and transceiver geometry

** Scene tokenizer represents the scene as a material-annotated point cloud, with each point described by its position and surface normal in the global frame and by material properties including relative permittivity and conductivity This involves a PointNet++-style set abstraction hierarchy that progressively downsamples the point set using farthest point sampling to yield material-aware scene tokens S The scene tokens are computed once per scene and reused across all queried transceiver links 3.2 SCENE AND TRANSCEIVER TOKENIZATION This involves a PointNet++-style set abstraction hierarchy that progressively downsamples the point set by factors of 4, 4, and 2, respectively, using farthest point sampling to yield material-aware scene tokens S The resulting Ns local features are projected to the common token dimension d, yielding S = [s1,..., sNs] ∈ R Ns×d **

** Transceiver tokenizer represents transceiver geometry at the element, array, and link levels To accommodate different array sizes and element arrangements, it represents each antenna element as an individual token using its local position normalized by the carrier wavelength λ Following the TX–RX link tokenization in Wi-GATr (Hehn et al., 2025), it additionally introduces tokens for the TX and RX array centers and their center-to-center displacement Together, these tokens capture transceiver geometry at the element, array, and link levels and form a variable-length sequence T ∈ R(Nt+Nr+3)×d 3.3 WIRELESS PROPAGATION MODELING This involves computing the pairwise geometry between individual TX and RX elements to produce a tensor E = [eij] Nt,Nr i=1,j=1, whose size scales with NtNr These pairwise features are aggregated along both the TX and RX dimensions to form element-indexed propagation tokens P ∈ R(Nt+Nr)×d Scene conditioning refines the scene tokens S through self-attention, after which they attend to the transceiver tokens T through cross-attention, yielding a scene representation Sc adapted to the current link and array geometry 3.4 SPATIAL COVARIANCE AND POWER PREDICTION The backbone representation is shared across configuration queries, while query-specific receiver orientations and the transmit beam are introduced only at the prediction stage Configuration conditioning introduces tokens z r j = h r j + ϕ r(Rpj/λ, GRR Rpj/λ, GfR, GuR, Re(wi), Im(wi), GRT T b) and z t i = h t i + ϕ t(T pi/λ, GRT T pi/λ, GfT, GuT, Re(wi), Im(wi), GRT T b) (2) 3.5 TRAINING OBJECTIVES The full training objective is L = Lspec + Lpower + λrxLvar,rx + λbeamLvar,beam (3) This loss supervises the spatial spectrum induced by the resulting covariance rather than the factor itself It also supervises configuration-dependent variation along the receiver orientation and transmit beam axes 4.3 ADDITIONAL EVALUATIONS Measured channel adaptation tests simulation-to-measurement adaptation on the single hallway DICHASUS dataset (Euchner et al., 2021) RFWeave achieves the highest spatial spectrum SSIM at every budget and performs best in the low-data regime 4.3 QUALITATIVE COMPARISONS Fig. 9–Fig. 12 provide additional qualitative comparisons for cross-scene prediction and single-factor configuration generalization RFWeave better preserves the narrow high-power structure and separated spectrum lobes than the feed-forward baselines 4.3 DOWNSTREAM BEAM MANAGEMENT By conditioning on transmit beams, RFWeave can be directly applied to downstream beam management RFWeave-B (FT) substantially outperforms location-based steering, increasing mean Top-1 SE from 7.24 to 9.74 bps/Hz (Table 6) The exhaustive sweep evaluates all beam pairs and selects the pair with the largest Pgt, providing an oracle reference 5 CONCLUSION We presented RFWeave, a learned wireless propagation model for joint generalization across scenes, transceiver locations, and transceiver configurations Across WiPTR and Wi3R, RFWeave consistently outperforms learned baselines under scene and configuration shifts It also adapts to DICHASUS measurements and supports downstream beam management 4.2 ABLATION STUDIES Removing configuration conditioning causes the largest degradation, reducing spatial spectrum SSIM by 5.5% Removing pairwise geometry encoding also consistently degrades performance, highlighting the importance of element-level pairwise geometry 4.3 CONFIGURATION VARIATION LOSS The balanced setting (λrx, λbeam) = (1, 1) achieves the best result on five of six reported metrics and is used in the main experiments 4.3 QUALITATIVE COMPARISONS Fig. 9–Fig. 12 provide additional qualitative comparisons for cross-scene prediction and single-factor configuration generalization The balanced setting (λrx, λbeam) = (1, 1) achieves the best performance on five of six reported metrics and provides the strongest overall results 4.3 DOWNSTREAM BEAM MANAGEMENT RFWeave-B (FT) substantially outperforms location-based steering, increasing mean Top-1 SE from 7.24 to 9.74 bps/Hz (Table 6) The exhaustive sweep evaluates all beam pairs and selects the pair with the largest Pgt, providing an oracle reference The best spectral efficiency among the retained pairs is then evaluated using ground-truth beam-pair power 4.3 CONFIGURATION VARIATION LOSS The balanced setting (λrx, λbeam) = (1, 1) achieves the best performance on five of six reported metrics and provides the strongest overall results The full sweep is reported in Appendix C.3 4.4 A IMPLEMENTATION DETAILS A.3 DETAILS OF BACKBONE The backbone uses two self-attention blocks for scene refinement Propagation tokens are refined through a sequence of stages where each stage applies self-attention to the propagation tokens, followed by cross-attention to the transceiver and scene context 4.4 B EXPERIMENTAL DETAILS B.

Improvements for AI systems

Bold header: Generalization across joint variations

RFWeave is designed to generalize across scenes, transceiver locations, and transceiver configurations, meaning it can accurately predict AoA spatial spectrum and received power using a single model even when the query involves unseen array sizes, receiver orientations, or transmit beams simultaneously.

Bold header: Robust representation of variable array structures

The system incorporates variable-length transceiver tokens that preserve antenna element structure across array sizes, allowing RFWeave to handle varying array structures by representing each antenna element as an individual token and aggregating pairwise geometries into element-indexed propagation tokens.

Bold header: Downstream beam management capability

RFWeave can be directly applied to downstream tasks because it enables beam decisions that achieve a 34.5% higher rates than location-based beam steering in simulated factory environments by conditioning on transmit beams.

Bold header: Adaptation to real-world measurements

The model demonstrates adaptation under limited measured data, achieving the highest spatial spectrum SSIM at every budget and performs best in the low-data regime when tested on measured DICHASUS channels with as few as 100 samples.

Bold header: Configuration sensitivity analysis

By supervising configuration-dependent variation along axes, RFWeave can be tuned using terms like Lvar,rx = SmoothL1(σc(Sb), σc(S)) to specifically improve generalization across receiver orientations or transmit beam directions.