acceptodds
Under review as a conference paper at ICLR 2027

Scene Reconstruction Using Ubiquitous Wireless Signals via Diffusion Posterior Inference over Physics-Grounded Representations

Abstract

Knowing the layout of the physical world matters for many applications, yet the sensors we rely on for it (e.g., cameras and LiDAR) can fail in darkness, in adverse weather, or behind walls. Wireless signals do not face these limitations: every link between a base station and a user is shaped by the buildings it passes, reflects from, or is blocked by, leaving information about the surrounding scene embedded in the wireless channel measured at the endpoints. However, mapping channel measurements to the environment is a severely ill-posed problem: a measurement records only the structures that interact with the transmitted electromagnetic waves along its propagation paths, and even these observations can often be explained by multiple plausible scenes. We therefore formulate wireless scene reconstruction as posterior inference and develop a hierarchical framework with two complementary stages to approach this problem. A wireless foundation model, pretrained to respect propagation physics, transforms each channel measurement into a representation of the scene structure that shaped the observation. A diffusion model then jointly integrates representations from multiple measurements to model and sample the posterior over building occupancy, with consensus over samples producing the final reconstruction. We demonstrate that the framework can accurately recover building footprints from wireless channel measurements alone and show that its performance depends critically on both design components: physics-grounded representations expose scene-relevant information, while joint posterior inference coherently integrates complementary observations.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Diffusion Models, Representation Learning, Wireless Foundation Model, Wireless Sensing, Scene Reconstruction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_jSmaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents a framework for reconstructing 2D building-occupancy maps from wireless channel measurements by treating reconstruction as posterior inference. The framework comprises two key modules: a pretrained foundation model and a diffusion model. The foundation model maps each channel measurement and receiver position into a latent representation, while the diffusion model combines multiple representations to generate 2D occupancy samples, whose cellwise consensus produces the final reconstruction. The framework was evaluated on datasets generated from digital twins; it achieves a mean balanced accuracy of 0.758 and a mean IoU of 0.506.

Critical strengths
  • The posterior formulation is well motivated by the physical characteristics of radio propagation and its interactions with the environment. The distinction between propagation-relevant and propagation-irrelevant regions is interesting and well explained.

  • The hierarchical design is thoughtful. Propagation-aware masking, spatial-spectrum distillation, scene alignment, and consensus sampling are well connected to the underlying physical principles.

  • The foundation model is designed and pretrained using intense GPU resources. It provides a unified latent representation of the channel measurements and receiver positions, enabling the downstream diffusion model to combine information from multiple observations.

Critical weaknesses
  • My primary concern is the dataset. It is insufficient to evaluate the proposed framework and demonstrate its performance on simulated channel datasets generated by digital twin. There is no evidence that the performance from simulated dataset can be translated to the real channel measurements.

  • The methodological novelty of the individual learning components is limited. Majority of content in Section 3.2 might not be needed as it is the standard diffusion model.

  • The evaluation lacks baseline comparison, making it hard to understand how the proposed framework advances the state of the art.

Questions for authors
  • My primary concern is the dataset. The evaluation data are generated using Sionna RT, a ray-tracing radio-channel simulator operating on virtual environments. Although Sionna RT produces considerably higher-fidelity channel measurements than simplified statistical models such as Rayleigh fading, simulated channels are far from fully capturing real-world channel measurements. The paper provides no evidence that the performance gains observed on digital-twin-generated data will transfer to measured channels. I am highly skeptical about this transfer because prior work has demonstrated substantial gap between simulated and real wireless channels.

  • The methodological novelty of the individual learning components is limited. Both the transformer-based foundation model and the diffusion model build on established techniques. Although the paper makes a valuable effort to adapt and investigate these techniques for physics-guided radio-based reconstruction, their practical significance should be demonstrated using measured channels rather than only simulated data. We understand that acquiring and learning from representative real-world channel measurements is difficult, but addressing this challenge is essential to establishing the value of the proposed framework.

  • Several novelty claims appear overstated. For example, the paper states, “To our knowledge, this is the first demonstration that communication-infrastructure channel measurements can support reconstruction...” However, prior studies have investigated closely related problems using different formulations, including "Environment Reconstruction Based on Multi-User Selection and Multi-Modal Fusion in ISAC" and "Diffraction and Scattering Aware Radio Map and Environment Reconstruction". Although the paper includes internal baselines and ablations, it does not compare against representative methods from this prior literature (It's fine with different methodologies). Therefore, it is difficult to determine how much the proposed framework advances the state of the art.

  • Radio propagation interacts with the environment in three-dimensional space rather than on a two-dimensional plane. Building heights, trees, terrain, and other objects can fundamentally affect channel observations. Nevertheless, the framework reconstructs only a 2D occupancy map. Although the authors explain that their uniform linear array primarily resolves azimuth, the paper does not sufficiently establish when a 2D representation is physically adequate or how variations in object height affect reconstruction.

4 comments
AuthorsOpenReview ↗

Response to Reviewer jSma (1/2): Summary and real-world evaluation

Response to Reviewer jSma (1/2)

We thank the reviewer for the thoughtful feedback and for recognizing the physical motivation of our posterior formulation, the distinction between propagation-relevant and -irrelevant regions, and the hierarchical design. We addressed every concern with new experiments and analyses.

Summary:

  • Real-world validation (W1, Q1). As suggested, we evaluated the framework on over-the-air channels collected with our custom fully digital MIMO testbed at five outdoor sites. On held-out sites it reaches 0.776 balanced accuracy and 0.488 IoU, comparable to its digital-twin performance despite real hardware nonidealities.
  • Methodological contribution (W2, Q2). New controlled baselines show that each level of our design is essential: with the same decoder, a conventionally pretrained foundation model and conditioning directly on raw channels both severely underperform.
  • Comparison with prior work (W3, Q3). We added a baseline adapted from Chen & Chen (2024) as suggested, which significantly underperforms in our setting. Unlike prior work, our framework needs neither dense non-stationary aerial access point measurements nor additional sensors such as cameras.
  • Building heights (Q4). All simulated and measured channels arise from full 3D scenes with varying building heights, so the reported scene reconstruction results already include this effect.

We have uploaded a revised PDF (changes in blue): over-the-air evaluation in Sec. 4.5 (p. 10, Table 3); new baselines in Sec. 4 (p. 7); Table 1 and Sec. 4.1 (p. 8); Table 2 (p. 9).

W1 / Q1: evaluation on measured channels

As the reviewer suggested, we evaluated our framework on real over-the-air measurements. Since no public dataset provides multi-antenna wideband channels together with accurate user positions and ground-truth building layouts in a configuration compatible with our setting, we collected them with a custom large-scale, fully digital MIMO testbed at 5.8 GHz with 100 MHz bandwidth (Sec. 4.5), matching the carrier, bandwidth, and base-station array size of our digital-twin corpus. A fixed base station with a fully digital 32-element array receives OFDM pilots from a mobile 6-element user array and estimates the channel on every subcarrier, while RTK-GNSS provides user positions. We measured five outdoor sites with distinct building layouts, with about 200 user locations each, covering a 40 m × 40 m area in front of the base station (a 20 × 20 grid at 2 m resolution). These measurements contain what the digital twin may abstract away, including real hardware nonidealities, antenna radiation pattern and coupling, and the true RF materials properties of each site.

Protocol. The pretrained foundation model is kept frozen and maps each measured channel and position to a [PHYSC] token, exactly as in the digital-twin evaluation. The diffusion decoder is retrained on the same simulated environments to match the output configuration. To bridge the simulation-to-reality gap, both encoder and decoder stay frozen and only the linear context adapter network through which [PHYSC] tokens enter the decoder is fine-tuned. We use a leave-one-site-out protocol: the adapter is trained on four sites and evaluated on the fifth, whose map is never used within adaptation, scene matching, or model selection, ensuring no leakage during evaluation.

Results (Table 3, each site held out):

SiteBal. Acc.F1IoU
10.6900.5410.371
20.9070.8130.685
30.6420.5650.394
40.8100.5000.333
50.8300.7950.659
Mean0.7760.6430.488

On held-out sites, the framework recovers the main geometric structure of each scene, with results comparable to those on digital-twin data (Table 1) despite real hardware nonidealities. Additionally, since only the linear adapter is tuned during evaluation, this shows that the physics-grounded representations learned in simulation transfer well to real hardware and real propagation condition. It also directly addresses the reviewer's point in Q2 that practical significance should be demonstrated on measured channels. Extending the campaign to larger and more complex environments is left for future work.

Our response continues in Part 2/2.

AuthorsOpenReview ↗

Response to Reviewer jSma (2/2): Contribution, prior work, and building heights

Response to Reviewer jSma (2/2)

W2 / Q2 and W3 / Q3: methodological contribution and comparison with prior work

Our contribution lies in how we formulate and solve wireless scene reconstruction. We formulate it as posterior inference over occupancy and make this posterior tractable through a hierarchical design: a foundation model with physics-grounded representations maps each measurement to the scene structure that shaped it, a diffusion decoder jointly integrates these representations to model the posterior, and a consensus readout over posterior samples yields a per-cell Bayes-optimal reconstruction (Props. 3–4). We added controlled baselines showing that conventional techniques fail at each level of this hierarchy, with their results included in the table below:

  • Physics-guided pretraining. The new baseline Diffusion + CLS (results in the table below) conditions an identical decoder on the [CLS] token of a foundation model with the same architecture, capacity, and data as ours, pretrained as a conventional masked autoencoder. Its performance stays near chance-level (0.537 balanced accuracy vs. 0.758 for ours), demonstrating the importance of our design in representation learning; Masked autoencoding retains what is needed to recover the channel, but nothing ties the representation to the scene structure that shaped the link, whereas [PHYSC] is anchored to the underlying physical properties and explicitly aligned with the propagation-relevant scene structure (Sec. 3.1).
  • Hierarchical design. Conditioning the same diffusion decoder directly on raw channels (Diffusion + Raw Meas.) collapses to chance (0.505). This is a failure of learnability rather than information: the decoder must simultaneously discover the propagation-to-geometry mapping and perform posterior generation, the two burdens our hierarchy separates.
  • Comparison with prior work. We thank the reviewer for pointing to these works. We added Geometry-Assisted DL + RSS (results in the table below), adapted from Chen & Chen (2024), which fits a virtual obstacle map to each link's received power; we use a strengthened variant whose propagation model is pretrained on our data. We demonstrate that such a simplified obstructed-vs.-free-space propagation model, like the generic embedding baseline, cannot expose the scene structure each link interacted with, and it severely underperforms our framework on all metrics.
Model (mean over 671 test scenes)Bal. Acc.F1IoU
Geometry-Assisted DL + RSS0.5340.4490.293
Diffusion + Raw Meas.0.5050.0540.030
Marginal Prob. + Raw Meas.0.6210.3970.251
Diffusion + CLS0.5370.2250.138
Marginal Prob. + FM (ours)0.7380.6330.470
Diffusion + FM (ours)0.7580.6630.506

The two cited works also operate under different requirements. Chen & Chen rely on hundreds of thousands of received-power measurements per area (about 600k) from aerial transmitters across many different viewpoint, and Lin et al. (2024) reconstruct sparse multipath point clouds that are fused with camera images at the users. To our knowledge, our framework is the first to provide meaningful reconstruction of arbitrary outdoor building layouts from the channels that conventional communication infrastructure already measures, without dense aerial links or additional sensors.

Q4: building heights and the 2D representation

The 2D occupancy grid in our framework is a choice of output, not an assumption about propagation. All channels in our digital-twin evaluation arise from full 3D scenes: buildings are extracted from OpenStreetMap with their individual heights, and ray tracing computes propagation in 3D between a base station at 15 m and users at 1.5 m (App. G). The reported results therefore already include the effect of varying building heights. For example, whether a building blocks a link depends on whether it is tall enough to intercept the ray, which our formulation treats as part of the residual ambiguity of the problem (App. B.3). The over-the-air experiments (Sec. 4.5) further confirm this in real environments, where every measured channel is shaped by the full 3D geometry of the site.

We reconstruct footprints rather than heights because, with a uniform linear array at the base station, the channel resolves paths only in azimuth, so building height is severely underspecified by the measurements. A 3D extension using a uniform rectangular array that resolves paths in elevation is a natural next step that we leave for future work (Sec. 4.1).

We hope these additions address the reviewer's concerns and would be glad to clarify further.

Reviewer_jSmaOpenReview ↗

Experimental settings

Thank you for your response and adding results on the over-the-air experimental datasets. Could you please provide images to show your base station equipment and the tested sites, so that we have an idea about the system setup and environmental layout. Also, could you please confirm that the coverage of 32-antenna 5.8 GHz base station is only 40m x 40m?

We thank the reviewer for acknowledging the new experimental demonstration with over-the-air transmissions and for the opportunity to clarify the system setup and measurement environment.

Details and Images of the equipment and deployment sites. As the reviewer suggested, we have added Appendix N, which describes the measurement campaign in detail. Figure 9 shows the testbed: (a) the base station with its fully digital 32-element receive array and (b) the mobile transmitter with its 6-element array and RTK-GNSS receiver. Figure 10 shows the layouts of the five measurement sites, with the base-station location and the measured region at each site.

System setup summary.

  • Base station: a 32-element uniform linear array of 4-patch subarrays. It is fully digital: each element has its own RF chain and ADC, and all 32 channels are digitized coherently and processed on a CASPER ROACH-2 FPGA platform, so the channel is captured at all elements simultaneously without analog beamforming.
  • User: a 6-element uniform linear array driven by a Pi-Radio software-defined radio front end with a Xilinx ZCU111 RFSoC back end.
  • Waveform and channel estimation: 128 subcarriers spanning 100 MHz around 5.8 GHz, matching the carrier, bandwidth, subcarrier count, and base-station array size of the digital-twin corpus. The six transmit antennas send Zadoff–Chu OFDM pilots in turn; after CFO correction and least-squares estimation, each measurement yields a 32 × 6 × 128 channel tensor. Per-chain hardware phase offsets are calibrated with a far-field horn antenna for accurate over-the-air channel estimation.
  • Positioning: an RTK-GNSS receiver provides centimeter-level user positions.
  • Sites: five outdoor sites in different parts of a campus, each with different buildings and materials, with 150–241 measurement points per site (about 1,000 in total) and occupied fractions from 16.0% to 47.5%.

Coverage. The extent of the measurement region is set by the transmit power of our testbed: to stay within the unlicensed 5.725–5.850 GHz band and the FCC limits, and to keep the amplifier stage below saturation, we transmit at relatively low radiated power, at which the user signal cannot be received reliably outside the dedicated measurement campaign area and at larger distances from the base station. We therefore confirm that 40 m × 40 m is the region measured and reconstructed in this campaign, but we emphasize that this is not a limitation of the method or of a 32-antenna base station. The street-level mounting of the base station (1.5 m) further restricts the coverage area. The region is thus limited by the testbed’s transmit power and mounting height rather than by the method. In the digital-twin evaluation, the same framework reconstructs 200 m × 200 m areas from a base station mounted at 15 m (Section 4, Appendix G), and a deployed base station with higher transmit power and elevated mounting would be expected to observe a correspondingly larger area.

We thank the reviewer again and hope this addresses the questions.

Reviewer_bPRgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how channel measurements from existing communication links, together with user positions, can be used to reconstruct two-dimensional occupancy maps of outdoor buildings. The proposed method first uses PHOENIX, a multimodal model pretrained with physics-guided objectives, to extract a scene representation from each observation. It then performs diffusion sampling conditioned on representations from multiple observations and applies per-cell majority voting to the generated samples to obtain the final reconstruction. The authors train on large-scale ray-tracing data and evaluate on 671 held-out scenes, examining the effects of scene density, the number of observations, and the quality of position information on reconstruction performance.

Critical strengths
  1. The problem has practical relevance. Reconstructing environments from channel observations already available over communication links offers a meaningful way to extend wireless infrastructure with sensing capabilities. To address the fact that each observation constrains only part of the environment, the authors combine physics-guided representation learning with multi-observation fusion. This design is well suited to the problem.

  2. The dataset is substantial, and the evaluation has good coverage. The method is evaluated on 671 held-out scenes and achieves the best mean reconstruction performance among the compared methods when user positions are accurate. Analyses across building densities, observation counts, and position availability also help clarify the conditions under which the proposed method is effective.

Critical weaknesses
  1. The comparisons do not sufficiently isolate the contributions of the individual design choices. The raw-channel diffusion baseline uses a smaller encoder trained from scratch, whereas PHOENIX undergoes large-scale pretraining. The performance gap therefore reflects differences in model capacity, pretraining, and physics-guided objectives. In addition, the marginal-probability approach uses a fixed fusion rule, and the comparisons do not include a deterministic decoder that jointly processes multiple observations using the same PHOENIX representations. The current results demonstrate the effectiveness of the overall pipeline, but do not yet establish whether physics guidance and diffusion-based posterior modeling are each necessary.

  2. The advantage of the diffusion approach is not consistently maintained under position errors. Table 3 in Appendix M shows that adding position errors with a standard deviation of 5 m along each coordinate axis reduces its mean IoU from 0.506 to 0.397, below the 0.453 achieved by the marginal-probability approach. The authors attribute this to a mismatch between training and test conditions, but do not experimentally verify whether training with perturbed positions can restore the advantage. Since perfectly accurate user positions are difficult to obtain in practice, the authors should clarify the range of position errors over which the method retains its advantage.

Questions for authors
  1. Could the authors provide better-controlled ablation studies to distinguish the gains from physics-guided pretraining and diffusion decoding? In particular, a deterministic decoder that jointly processes multiple observations using the same PHOENIX representations would help assess the additional benefit of diffusion modeling itself.

  2. To address the performance degradation caused by position errors in Table 3, could the authors report results after training with perturbed positions, or compare performance across different levels of position error? These results would help establish the positioning conditions under which the diffusion approach retains an advantage over the marginal-probability approach.

2 comments
AuthorsOpenReview ↗

Response to Reviewer bPRg (1/2): Summary and evaluation of isolated design components

Response to Reviewer bPRg (1/2)

We thank the reviewer for the constructive feedback and for recognizing the practical relevance of the problem, our design, and the breadth of our evaluation. We addressed both concerns with new experiments, which we believe fully resolve them.

Summary:

  • Importance of physics-guided pretraining (W1, Q1). Established with a new controlled baseline: Diffusion + CLS (same encoder size and data as ours, pretrained as a masked autoencoder) which significantly underperforms (0.537 vs. ours 0.758 balanced accuracy), showing the gain comes from our physics-guided objectives, not pretraining alone.
  • Importance of joint posterior inference (W1, Q1). On the same representations, joint diffusion inference outperforms the per-measurement marginal decoder on all metrics, by up to +16.4% P90 IoU and +42.7% IoU in dense scenes.
  • Position errors (W2, Q2). Resolved by training with perturbed positions: performance stays within 1.5% / 5.2% of exact positions at σ = 5/10 m, and the advantage over the marginal approach is restored at every error level.

We have uploaded a revised PDF (changes in blue): Sec. 4 (p. 7); Table 1 and Sec. 4.1 (p. 8); Table 2 (p. 9); App. M (pp. 24–25). It also adds over-the-air experiments at five outdoor sites (Sec. 4.5).

W1 / Q1: isolating each design choice

(a) Physics-guided vs. generic pretraining. We implemented Diffusion + CLS (Table 1): an encoder with the same architecture, capacity, and data as ours, pretrained as a masked autoencoder, whose [CLS] token conditions an identical decoder. Only the pretraining objective differs, yet its balanced accuracy (0.537) is far below ours (0.758), barely above raw measurements (0.505), and near chance. The benefit thus stems not from pretraining alone but from how the representation is formed: masked autoencoding retains what is needed to recover the channel, but nothing ties it to the scene structure that shaped the link, whereas [PHYSC] is anchored to the underlying physical properties and explicitly aligned with the propagation-relevant scene structure (Sec. 3.1). This scene-aligned representation substantially simplifies posterior inference: the decoder need not discover the propagation-to-geometry mapping while generating scenes. We also compared against Geometry-Assisted DL + RSS, adapted from the prior work (Chen & Chen, 2024), which fits a virtual obstacle map to each link’s received power. We demonstrate such a simplified obstructed-vs.-free-space propagation model, like a generic embedding, cannot expose the scene structure each link interacted with, and hence would not be viable to the proposed severely ill-posed inverse problem. It severely underperforms our framework on all metrics (Table 1).

(b) Joint posterior vs. fused marginals. Marginal Prob. + FM is a deterministic decoder on the same representations that estimates a separate marginal per measurement and then fuses them; Diffusion + FM processes all observations jointly to estimate the joint distribution over the grid. Joint inference improves every metric, most in the upper tail (P90 IoU +16.4%), and its advantage grows with scene density (Table 2). Fusing separately estimated marginals discards evidence that is weak in each observation but informative in combination; the joint posterior combines it and completes unobserved regions through the learned prior.

With these results, we show that both components are necessary.

Table 1 (revised), 671 test scenes, mean / median / P90.

ModelBal. Acc.F1IoU
Geometry-Assisted DL + RSS0.534 / 0.530 / 0.5920.449 / 0.445 / 0.5480.293 / 0.286 / 0.378
Diffusion + Raw Meas.0.505 / 0.500 / 0.5190.054 / 0.016 / 0.1620.030 / 0.008 / 0.088
Marginal Prob. + Raw Meas.0.621 / 0.618 / 0.6680.397 / 0.398 / 0.5030.251 / 0.248 / 0.336
Diffusion + CLS0.537 / 0.520 / 0.5990.225 / 0.194 / 0.4520.138 / 0.107 / 0.292
Marginal Prob. + FM (ours)0.738 / 0.739 / 0.8030.633 / 0.642 / 0.7370.470 / 0.473 / 0.584
Diffusion + FM (ours)0.758 / 0.746 / 0.8680.663 / 0.655 / 0.8100.506 / 0.487 / 0.680

Capacity of the Diffusion + Raw Meas. Encoder. This encoder is trained end-to-end with the decoder over the joint conditioning set, running forward and backward for all K = 100–500 observations of a scene at each step, while Diffusion + FM encodes each observation once with the frozen model. Cost grows linearly with K on top of quadratic attention per observation; at our capacity this is ~8×10¹² encoder tokens, ~9× our entire pretraining, which is impractical. Its training resources therefore match our pretraining at a lower parameter count, which is instead controlled by Diffusion + CLS and Marginal Prob. + Raw Meas. (from scratch), whose encoders match ours in size; both significantly underperform our framework.

Our response to W2/Q2 continues in Part 2/2

AuthorsOpenReview ↗

Response to Reviewer bPRg (2/2): Robustness to different levels of position errors

Response to Reviewer bPRg (2/2)

W2 / Q2: robustness to position errors

We thank the reviewer for this suggestion. Robustness to positioning error is a prerequisite for using the framework with commodity devices, whose consumer GNSS errors in outdoor urban settings commonly reach several meters. As the reviewer suggested, we implemented both evaluations, (i) comparing all models across different position-error levels and (ii) training the diffusion decoder with perturbed positions, and added them to Appendix M (pp. 24–25, Tables 4–5; formerly Table 3). The results show that our framework retains both its performance and its advantage even under large positioning errors.

Setup. We evaluate position errors with standard deviations of 5 m and 10 m per horizontal axis, applied independently to each user position.

Distribution shift and Robust Diffusion + FM. Position error is expected to induce a distribution shift for the diffusion decoder. Diffusion + FM is trained on embeddings computed at exact positions, so it learns to treat each embedding as precise evidence about a specific location. At test time, perturbed positions place this evidence in the wrong cells, shifted in different directions for different observations; these observations conflict, posterior samples disagree on the affected cells, and the consensus readout suppresses footprints that are otherwise recovered. This mismatch is a property of training rather than a limitation of the posterior formulation, and it can be compensated for directly by treating position error as part of the observation model. We therefore propose Robust Diffusion + FM: the frozen foundation model encodes each training observation at a position perturbed with σ ∈ {0, 5, 10} m, balanced within each scene, and the decoder learns pθ(s | Z̃) with the clean occupancy as target. It thus learns to reconstruct the scene from both exact and mislocalized evidence, without being told the noise level at test time. Architecture, objective, and hyperparameters are unchanged.

Results (mean bal. acc. / F1 / IoU over 671 test scenes; full statistics in Tables 4–5):

Modelσ = 0 m (exact)σ = 5 mσ = 10 m
Marginal Prob. + FM0.738 / 0.633 / 0.4700.728 / 0.617 / 0.4530.690 / 0.564 / 0.398
Diffusion + FM (exact-position training)0.758 / 0.663 / 0.5060.696 / 0.553 / 0.3970.604 / 0.353 / 0.231
Robust Diffusion + FM0.743 / 0.641 / 0.4820.739 / 0.634 / 0.4750.728 / 0.616 / 0.457
  • Performance is retained under large errors. Robust Diffusion + FM stays within 1.5% of its exact-position performance on all metrics at σ = 5 m, and even at σ = 10 m, where typical errors exceed a grid cell, it loses only 2.0% balanced accuracy and 5.2% IoU.
  • The advantage over fused marginals is restored and widens with error. It is the best model at both noise levels. Relative to Marginal Prob. + FM, IoU improves by 4.9% at σ = 5 m and 14.8% at σ = 10 m, and at σ = 10 m it reconstructs scenes nearly as accurately as Marginal Prob. + FM does with exact positions.
  • Robustness comes at only a small cost when positions are exact, while still outperforming Marginal Prob. + FM.

To directly answer the reviewer's question on the operating range: with training on perturbed positions, the diffusion approach retains its advantage over the marginal-probability approach at every error level we evaluated, up to σ = 10 m per axis. Positioning error is therefore a deployment condition that can be anticipated and absorbed during training rather than a fundamental obstacle, making the framework compatible with positions reported by commodity devices.

We hope these additions address the reviewer's concerns and would be glad to clarify further.

Reviewer_e34MOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper addresses the problem of environment reconstruction using electromagnetic waves as encountered in wireless devices.

The problem is formulated as posterior inference consisting of two stages, the first one is a wireless foundation model and maps CSI to the structured representation of the scene, and the second stage, a diffusion step, that combines measurement to sample and construct final constructions. The key ideas are physics grounded representation and joint posterior inference, the later defines a distribution over the reconstructed scene, addressing the challenge that multiple scenes can correspond to the same measurement. The value of these two stages are shown in the paper, mainly direct scene reconstruction underperforms.

Section 2 of the paper introduces the problem formulation, particularly equation 1 giving the MIMO CSI equation as a function of scene structure. Section 2.2 formulates the scene as occupancy grid, and 2.3 arguing for the two phase inference in equation (2), where the role of prior is to go from a scene representation to various reconstructions.

Section 3.1 introduces PHOENIX, which is representation learning framework to learn useful embedding for the diffusion space, and Section 3.2 is the diffusion posterior over the scene. For the representation learning phase, two methods are used: bidirectional reconstruction (reconstruct one modality from another) and physics-guided cross-modal distillation (predicting common physical quantities). Section 3 represents DDPM style diffusion model training, while the inverse problem uses the conditional diffusion training.

Experiments are presented in Section 4 mainly using simulated data.

Critical strengths
  • The system design is meaningful. Self-supervised learning based on cross-model prediction instead of contrastive loss is a good choice, as it forces to extract useful information relevant to the common information across modalities.
  • The idea of using posterior inference to address under-determined nature of the problem is a sound choice as well as choice of the diffusion models.
Critical weaknesses

Overall, I do not feel that the paper has strong technical novelty (many ideas already discussed in the literature) nor does a good job comparing with similar methods and show experimental adequacy.

  • The self-supervised learning ideas (bidirectional reconstruction ) is not new and is discussed in other works for instance:

Sadjad Alikhani, Gouranga Charan, Ahmed Alkhateeb, Large Wireless Model (LWM): A Foundation Model for Wireless Channels, https://arxiv.org/abs/2411.08872

Vahid Yazdnian, Yasaman Ghasempour, A Multi-Modal Foundational Model for Wireless Communication and Sensing https://arxiv.org/abs/2602.04016

Also the idea of using diffusion model for such wireless inverse problems is discussed in Hehn et al, ICLR 2025 (cited by the paper).

  • The experiments are all based on ray tracing (11,000 outdoor environments and 30 million ray-traced channel observations). In my opinion, ray tracing represents an ideal situation, from the way propagation effects are modeled, the scene geometry and transceiver chain pipelines. The problems in this space should be easy to solve as evidenced by many similar works in the literature.
  • There are good real datasets that can be useful for instance from DeepSense MIMO (ASU lab), or from Bell Labs - Stuttgart University setup, and some other cases. The authors should try to see if those datasets can be used.
  • The baselines are too simple. They are generally obtained by removing one of the components of the proposed system. There are many works that can be used as baseline, for instance using Sionna itself (assuming that scene is parametrized through occupancy and backproping), or using Nerf2 paper, or Hehn et al paper. These paper are highly relevant and should be part of the baselines.
  • The authors claim that "this is the first treatment of wireless scene reconstruction that assumes nothing about the number, shape, convexity, or location of scatterers". And differentiable ray tracer technically satisfies this assumption, ranging from Sionna simulator to neural wireless simulator as cited by the paper itself.
  • The authors use known antenna patterns in style of DFT beams as shown in equation (6). This is a highly ideal case. The idea that array responses ara_r and ata_t are known is not valid in practice. In general, base station and user sides are not aware of their antenna patterns, and these patterns can vary drastically across devices and depending on their pose in user's hand.
  • I recommend the authors to have two separate process for training and for inference. Currently, too much focus is on training in the main papers.
  • Conditional diffusion approach to inverse problems will need retraining anytime the type of conditioning signals change. Diffusion inverse problem solvers like DPS are more competitive and good to be included.
Questions for authors
  • Do authors have controlled experiment that show that the posterior is well calibrated? In other words, it contains the scenes that correspond to the measurements?
  • Can authors clarify the assumption on material properties of the scenes and how they vary in data, training and test? Figuring out material properties is an important challenge in itself.
  • What is the impact of antenna patterns on the outcome?
  • How does the proposed method compare with other baselines I mentioned above?
5 comments
AuthorsOpenReview ↗

Response to Reviewer e34M (1/3): Summary and real-world evaluation

Response to Reviewer e34M (1/3)

We thank the reviewer for the detailed feedback and for recognizing that our system design is meaningful, that cross-modal prediction is a good choice for self-supervised learning, and that posterior inference with diffusion models is a sound way to address the under-determined nature of the problem. We have addressed every concern with new experiments and analyses.

Summary:

  • Real-world evaluation. As suggested, we evaluated the framework on over-the-air channels from our custom fully digital MIMO testbed at five outdoor sites. On held-out sites it reaches 0.776 balanced accuracy and 0.488 IoU, comparable to its digital-twin performance, showing that it transfers to real antenna patterns, coupling, hardware nonidealities, and building materials.
  • Design components and comparison with prior work (Q4). New controlled baselines show that each level of our design is essential: with the same decoder, a conventionally pretrained foundation model and conditioning on raw channels both significantly underperform. We explain why differentiable ray tracing cannot infer occupancy in our setting and instead added a baseline adapted from the closest prior work (Chen & Chen 2024), which also underperforms.
  • Antenna patterns and materials (Q2, Q3). Pretraining assumes omnidirectional radiation patterns and does not explicitly model materials, yet the over-the-air evaluation shows that the representation transfers to unknown real antenna patterns and real building materials with only a linear context adapter. Known patterns or more diverse materials can be incorporated directly into pretraining.

We have uploaded a revised PDF (changes in blue): over-the-air evaluation in Sec. 4.5 (p. 10, Table 3); new baselines in Sec. 4 (p. 7); Table 1 and Sec. 4.1 (p. 8); Table 2 (p. 9).

Evaluation on measured channels

As reviewer suggested to test the framework beyond ray tracing, we evaluated it on real over-the-air measurements. Since none of the public datasets provides multi-antenna wideband channels together with accurate user positions and ground-truth building layouts, we collected measurements with a custom large-scale, fully digital MIMO testbed at 5.8 GHz with 100 MHz bandwidth (Sec. 4.5), matching the carrier, bandwidth, and base-station array size of our digital-twin corpus. A fixed base station with a fully digital 32-element array receives OFDM pilots from a mobile 6-element user array and estimates the channel on every subcarrier, while RTK-GNSS provides user positions. We measured five outdoor sites with distinct building layouts, with about 200 user locations each, covering a 40 m × 40 m area in front of the base station (a 20 × 20 grid at 2 m resolution).

Protocol. The pretrained foundation model is kept frozen and maps each measured channel and position to a [PHYSC] token, exactly as in the digital-twin evaluation. The diffusion decoder is retrained on the same simulated environments to match the output configuration. To bridge the simulation-to-reality gap, both encoder and decoder stay frozen and only the linear context adapter network through which [PHYSC] tokens enter the decoder is fine-tuned. We use a leave-one-site-out protocol: the adapter is trained on four sites and evaluated on the fifth, whose map is never used within adaptation, scene matching, or model selection, ensuring no leakage during evaluation.

Results (Table 3, each site held out):

SiteBal. Acc.F1IoU
10.6900.5410.371
20.9070.8130.685
30.6420.5650.394
40.8100.5000.333
50.8300.7950.659
Mean0.7760.6430.488

These measurements contain everything ray tracing may idealize: the real radiation patterns and mutual coupling of the arrays, hardware nonidealities across the RF chains, and the true geometry and RF interaction properties of the building materials at each site. On held-out sites, the framework recovers the main geometric structure of each scene, with results comparable to those on digital-twin data (Table 1). Additionally, since only the linear adapter is tuned during evaluation, this shows that the physics-grounded representations learned in simulation transfer well to real hardware and real propagation condition.

Our response continues in Part 2/3.

AuthorsOpenReview ↗

Response to Reviewer e34M (2/3): Design components and comparison with prior work

Response to Reviewer e34M (2/3)

Importance of design components and comparison with prior work (Q4)

Our contribution lies in how we formulate and solve wireless scene reconstruction. We formulate it as posterior inference over occupancy and make this posterior tractable through a hierarchical design: a foundation model with physics-grounded representations maps each measurement to the scene structure that shaped it, a diffusion decoder jointly integrates these representations to model the posterior, and a consensus readout over posterior samples yields a per-cell Bayes-optimal reconstruction (Props. 3–4). We do not claim self-supervised pretraining or diffusion for inverse problems as new in themselves. Instead, we added controlled baselines showing that conventional techniques fail at each level of this hierarchy, with their results included in the table below:

  • Physics-guided pretraining. The new baseline Diffusion + CLS (results in the table below) conditions an identical decoder on the [CLS] token of a foundation model with the same architecture, capacity, and data as ours, pretrained as a conventional masked autoencoder, as in channel foundation models such as LWM [1]. It stays near chance and significantly underperforms (0.537 balanced accuracy vs. 0.758 for ours). Masked autoencoding retains what is needed to recover the channel, but nothing ties the representation to the scene structure that shaped the link, whereas [PHYSC] is anchored to the underlying physical properties and explicitly aligned with the propagation-relevant scene structure (Sec. 3.1).
  • Hierarchical design. Conditioning the same diffusion decoder directly on raw channels (Diffusion + Raw Meas.), closest in spirit to Hehn et al. [2], collapses to chance (0.505). This is a failure of learnability rather than information: the decoder must simultaneously discover the propagation-to-geometry mapping and perform posterior generation, the two burdens our hierarchy separates. On the same representations, the joint diffusion posterior also outperforms fusing per-measurement marginals (Marginal Prob. + FM), most strongly in dense scenes (Table 2).

Discussion on differentiable ray tracing. Differentiable ray tracers such as Sionna RT [3] provide gradients with respect to continuous parameters of a given scene, such as materials, antenna patterns, and positions. Occupancy instead determines which propagation paths exist: adding or removing a structure changes the set of paths, so the channel is discontinuous in occupancy, and occlusion creates regions where the gradient is zero [4]. If a structure is absent from the current estimate, no traced path interacts with it, so no gradient points toward creating it. Accordingly, recent inverse rendering such as RFDT [5] refines an initialized mesh from dense radar measurements rather than deciding where spatial structures exist, and NeRF² [6] fits a radiance field to dense measurements of one scene to synthesize channels rather than recover geometry. With sparse links, the problem is moreover severely ill-posed, so even a converged fit returns only one of many consistent scenes. A differentiable ray tracer therefore imposes no parametric form on scatterers but takes the geometry as input, whereas our claim in the introduction concerns inferring it.

Making occupancy optimizable along conventional techniques requires relaxing hard path visibility into a simpler propagation model like soft obstructed-vs.-free-space model, which is precisely the approach of the prior work, Chen & Chen [7]. We therefore implemented it as the baseline Geometry-Assisted DL + RSS (results in the table below), which fits a virtual obstacle map to each link's received power; we use a strengthened variant whose propagation model is pretrained on our data. We demonstrate that such a simplified propagation model, like a generic embedding, cannot expose the scene structure each link interacted with: it provides low accuracy of 0.534 and severely underperforms our framework on all metrics.

Model (mean over 671 test scenes)Bal. Acc.F1IoU
Geometry-Assisted DL + RSS0.5340.4490.293
Diffusion + Raw Meas.0.5050.0540.030
Marginal Prob. + Raw Meas.0.6210.3970.251
Diffusion + CLS0.5370.2250.138
Marginal Prob. + FM (ours)0.7380.6330.470
Diffusion + FM (ours)0.7580.6630.506

Our response continues in Part 3/3; references are listed at its end.

AuthorsOpenReview ↗

Response to Reviewer e34M (3/3): Antenna patterns, materials, and inverse-problem solvers

Response to Reviewer e34M (3/3)

Antenna patterns (Q3)

The array steering vectors in Eq. 6 are part of the channel itself, meaning that for each propagation path, they describe the phase shift the path produces across the antenna elements, determined by its angle of arrival or departure and the array geometry. In our simulated pretraining corpus, both ends are modeled as ideal phased arrays with omnidirectional radiation patterns. Changing the radiation pattern does affect the channel, since it reweights each path according to its direction and thus changes what the encoder observes. We tested this transferability in the over-the-air experiments, where the testbed arrays have non-ideal radiation patterns that are unknown to the model and differ from the omnidirectional elements of pretraining, together with mutual coupling and per-chain hardware responses. Even under these conditions, the results show that the frozen representation can be transferred with only a linear adapter to match the practical data distribution (Sec. 4.5, Table 3). Moreover, when the radiation pattern of a deployment is known, it can be taken into account directly by including it in the simulated pretraining corpus, so that the representations are learned under the array that is actually deployed.

Material properties (Q2)

The model does not explicitly account for material properties: it neither receives nor estimates them, and the occupancy grid does not encode them, so they act as unobserved factors of the likelihood (App. F, Eq. 19). The adaptation of our framework to real materials is tested in the over-the-air evaluation, where the measured channels reflect the true RF interaction properties of real buildings and the framework still recovers the main geometric structure of held-out sites (Table 3). For more severe out-of-distribution material variations, robustness can be further improved by pretraining on a corpus with more diverse material properties, so that the representations and the diffusion posterior learn to account for this variability.

Diffusion inverse-problem solvers

We thank the reviewer for this suggestion. Our decoder cross-attends to a variable-size set of observations, so it needs no retraining as the number of observations changes (trained with 100–500 and evaluated up to 1,000; App. G). Solvers such as DPS [8] avoid retraining when the type of conditioning changes, but require a differentiable likelihood, which here is defined only through ray tracing of the physical scene, with the discontinuities discussed above. Combining such solvers with a likelihood learned on our representations is a promising direction that we will explore.

We hope these additions address the reviewer's concerns and would be glad to clarify further.

References

[1] S. Alikhani, G. Charan, A. Alkhateeb. Large Wireless Model (LWM): A foundation model for wireless channels. arXiv:2411.08872, 2024.

[2] T. Hehn, M. Peschl, T. Orekondy, A. Behboodi, J. Brehmer. Differentiable and learnable wireless simulation with geometric transformers. ICLR, 2025.

[3] J. Hoydis et al. Sionna RT: Differentiable ray tracing for radio propagation modeling. IEEE Globecom Workshops, 2023.

[4] J. Eertmans, L. Jacques, C. Oestges. Fully differentiable ray tracing via discontinuity smoothing for radio network optimization. EuCAP, 2024.

[5] X. Chen, X. Zhang, K. Zheng, X. Fang, T.-M. Li, C. X. Lu, Z. Li. Physically accurate differentiable inverse rendering for radio frequency digital twin (RFDT). ACM MobiCom, 2026.

[6] X. Zhao, Z. An, Q. Pan, L. Yang. NeRF²: Neural radio-frequency radiance fields. ACM MobiCom, 2023.

[7] W. Chen, J. Chen. Diffraction and scattering aware radio map and environment reconstruction using geometry model-assisted deep learning. IEEE Transactions on Wireless Communications, 2024.

[8] H. Chung, J. Kim, M. T. McCann, M. L. Klasky, J. C. Ye. Diffusion posterior sampling for general noisy inverse problems. ICLR, 2023.

Comments on the paper

1 comment

We thank all reviewers for their constructive feedback. We have uploaded a revised PDF (changes in blue) and responded to each reviewer individually. The main additions are:

  • Real-world evaluation (Sec. 4.5, Table 3). Over-the-air channels measured with our custom fully digital MIMO testbed at five outdoor sites. With the foundation model frozen and only a linear adapter tuned, the framework reaches 0.776 balanced accuracy and 0.488 IoU on held-out sites, comparable to its digital-twin performance.
  • New baselines (Sec. 4, Tables 1–2, App. I). Diffusion + CLS, a foundation model of the same size and data pretrained as a conventional masked autoencoder, and Geometry-Assisted DL + RSS, adapted from Chen & Chen (2024). Both significantly underperform our framework, showing that both the physics-guided representation and the joint posterior are essential.
  • Robustness to position errors (App. M, Tables 4–5). Training with perturbed positions keeps performance within 1.5% / 5.2% IoU of exact positions at σ = 5 / 10 m, while retaining the advantage over the marginal approach at every error level.

We would be glad to answer any further questions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.