acceptodds
Withdrawn from ICLR 2027

Flow-RI: Residual Inference with Flow Matching for Wireless Spatial Field Reconstruction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Inverse problems recover an unknown signal or physical field from incomplete and noisy observations. In wireless spatial field reconstruction, measurements are often available at only a small fraction of locations, making the problem particularly difficult at low sampling rates. A pretrained generative model can provide a strong prior for the complete field, while the local spatial correlation in wireless fields provides an additional structural prior that can be exploited through spatial regularization. A straightforward approach is to apply this regularization to the complete reconstructed field. However, doing so may alter structures that are already captured by the learned prior. To address this, we introduce Flow-RI, a residual inference framework built on a pretrained Flow Matching prior. Instead of regularizing the complete field, Flow-RI applies spatial regularization only to the prediction residual. At each flow step, the mismatch between the observed measurements and the predicted values at the sampled locations provides sparse observations of this residual. A spatial regularizer estimates the residual over the complete grid from these sparse observations, and the resulting residual estimate is used to correct the Flow Matching prediction before forming the next intermediate state. We further analyze the effect of applying spatial regularization to the prediction residual rather than to the complete field. At a fixed flow step, under the assumption that the reference prediction equals the conditional mean and that the measurement noise is conditionally zero-mean, residual regularization preserves conditional unbiasedness, whereas regularizing the complete field generally introduces an additional smoothing bias. Experiments on ray-traced wireless fields show that Flow-RI achieves the lowest mean absolute error (MAE) across all evaluated wireless reconstruction settings. The method also generalizes to different transmitter deployments and previously unseen urban scenes without retraining.

Primary area:
generative models
Keywords:
Flow Matching, residual inference, wireless spatial field, inverse problem

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

This paper presents Flow-RI for wireless field reconstruction from sparse and noisy measurements. It first uses a pretrained Flow Matching model to predict the complete field, then treats the measurement mismatch as sparse observations of the prediction error. A graph-Laplacian prior estimates this error at unobserved locations, and the corrected prediction is used in the next inference step. The paper also provides a fixed-step bias comparison with complete-field smoothing. Experiments on ray-traced wireless fields show consistent gains at low sampling rates.

Critical strengths
  1. The overall method is simple and well motivated. The Flow Matching model predicts the main field structure, while the spatial model only corrects the remaining error. This design helps preserve details that have already been reconstructed correctly.

  2. Applying a Flow Matching solver for inverse problems to wireless spatial field reconstruction is an interesting and practically relevant direction. The paper also shows how a pretrained generative prior can be combined with spatial information from the wireless domain.

  3. The evaluation covers several useful settings, including different noise levels, sampling patterns, transmitter and receiver settings, and unseen cities. This broad evaluation provides useful evidence of robustness.

Critical weaknesses
  1. Algorithm 1 of Flow-RI follows almost the same sequence of updates as Algorithm 1 of FLOWER. FLOWER defines three steps: flow-consistent destination estimation, measurement-aware destination refinement, and movement in time. These steps can be written as
\begin{aligned} &\widehat{x}_{1,t} =x_t+(1-t)v_\theta(x_t,t) \\\\ &\widehat{r}_t =\arg\min_r \frac{1}{2\sigma_n^2}\lVert Hr-e_t\rVert_2^2 +\frac{1}{2\nu_t^2}\lVert r\rVert_2^2 \\\\ &x_{t+\Delta t} =(1-t-\Delta t)\epsilon +(t+\Delta t)\widetilde{x}_{1,t}, \qquad \epsilon\sim p_0 \end{aligned}$$ Here, $e_t=y-H\widehat{x}_{1,t}$ and $\widetilde{x}_{1,t}=\widehat{x}_{1,t}+\widehat{r}_t$. Flow-RI follows the same sequence $$\qquad \begin{aligned} &\widehat{X}_{1,t_k} =X_{t_k}+(1-t_k)v_\theta(X_{t_k},t_k) \\\\ &\widehat{R}_{t_k} =\arg\min_R \frac{1}{2}\lVert A_\Omega\operatorname{vec}(R)-e_{t_k}\rVert_2^2 +\frac{\lambda}{2}\operatorname{vec}(R)^\top L\operatorname{vec}(R) \\\\ &X_{t_{k+1}} =(1-t_{k+1})Z_{k+1} +t_{k+1}\widetilde{X}_{1,t_k}, \qquad Z_{k+1}\sim p_0 \end{aligned}$$ Similarly, $e_{t_k}=y-A_\Omega\operatorname{vec}(\widehat{X}_{1,t_k})$ and $\widetilde{X}_{1,t_k}=\widehat{X}_{1,t_k}+\widehat{R}_{t_k}$. The first and last steps of the two algorithms have the same form. The main algorithmic difference seems to be that FLOWER uses an $\ell_2$ residual prior, while Flow-RI uses a graph-Laplacian residual prior. Although the authors cite FLOWER in the Related Work section, it is presented only as a related method, without sufficient attribution of the core algorithmic components inherited from it. Several highly overlapping components are re-derived consecutively in the Method section as part of Flow-RI's own procedure. The authors should properly cite FLOWER in the method section and explain more clearly what is new in the proposed algorithm. 2. All experiments are conducted on synthetic data generated by Sionna RT, and Flow-RI is not evaluated using measurements collected in a real environment. This limits the evidence for its effectiveness under real-world propagation. [1] Mehrsa Pourya, Bassam El Rawas, and Michael Unser. Flower: A flow-matching solver for inverse problems. ICLR, 2026.
Questions for authors
  1. Please properly cite FLOWER in the method section and clarify what is new in Flow-RI.

  2. I am not familiar with the wireless spatial field domain. Could the authors explain why Flow-RI was not evaluated using measurements collected in real environments? If such an evaluation is not feasible, please clarify the reasons.

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

This paper proposes a residual inference framework named Flow-RI based on a pre-trained fluid matching model for wireless spatial field reconstruction. The authors conducted a theoretical analysis of residual regularization and carried out experiments on simulated wireless spatial fields under different sampling rates, noise levels, and deployment scenarios. The results show that its reconstruction accuracy is improved compared to several existing methods based on fluid matching.

Critical strengths

The proposed method has achieved lower reconstruction errors than multiple baseline methods based on flow matching under different sampling rates, noise levels, and deployment conditions, fully demonstrating its potential effectiveness in sparse wireless spatial field reconstruction.

Critical weaknesses
  1. The experimental evaluation lacks sufficient generalization ability. In the core task of wireless spatial field reconstruction, this study relied solely on an automatically generated dataset without verifying it against established public benchmarks or corpora. Moreover, the natural image restoration experiment only used eight test images, which is far from sufficient for a reliable assessment.

  2. Although several Flow Matching-based methods and component ablations are evaluated, comparisons with alternative residual estimation techniques are missing. Consequently, it remains unclear whether the performance improvements stem from the proposed residual inference framework or simply from incorporating conventional spatial regularization.

  3. The proposed framework mainly combines the pre-trained flow matching prior with the traditional regularization method. Although this fusion is reasonable, its methodological innovation beyond the existing residual correction and spatial interpolation techniques has not been fully verified. The authors should more clearly distinguish the differences between their own contributions and the existing methods.

  4. This manuscript is difficult to understand as it lacks any overview diagram that explains the research motivation or the proposed methods. Moreover, the experimental section does not begin until page 8, resulting in limited space for presenting and discussing the main results.

Questions for authors
  1. Could the author further elaborate on the methodological innovation of Flow-RI compared to existing residual correction and spatial interpolation techniques? What specific advantages does the proposed framework have over simpler residual estimation methods?

  2. Could the author provide stronger evidence of generalization ability, especially by conducting evaluations in the existing public wireless reconstruction benchmark tests and using a much larger natural image test set?

Reviewer_f4yEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes Flow-RI for reconstructing wireless signal-strength maps from sparse measurements. It uses a pretrained flow-matching model to predict a complete map. It does so by computing prediction residuals at observed locations, and estimating the residual over the remaining grid using spatial regularization. The corrected prediction is fed back into subsequent inference steps. Evaluation was performed with simulated data, including changes in transmitter deployment and urban scene.

Critical strengths

Regularizing prediction errors rather than the complete predicted field is an effective way to preserve structure supplied by the learned prior. The paper compares these alternatives and includes ablations separating the contributions of the prior, spatial residual propagation, and iterative feedback.

Critical weaknesses

1. The technical contribution appears limited beyond spatial residual interpolation within flow matching.

The inverse problem considered here is essentially a completion or interpolation problem. The method does not infer the physical scene or propagation parameters that generated the measurements. Its central operation is to compute errors at observed locations, extend those errors across the grid using a quadratic spatial smoothness penalty, and add the resulting correction to the current prediction.

Combining a baseline prediction with spatial interpolation of its residuals is not new (e.g., regression kriging). Existing flow-based inverse solvers, such as FLOWER, also follow a predict–correct–update structure. The paper does not sufficiently distinguish its formulation from existing residual spatial estimation.

2. The improvement was marginal comparing with baselines.

As described in the paper, the MAE gain is less than 1 dB, and becomes smaller as the sampling rate increases.

3. The evaluation is conducted in simulation only without any real-world data.

The authors should test their method on real-world datasets and study the generalization of the learned prior. Right now all the training and evaluation use ray-traced data. The prior is trained on 2k maps from one urban scene with receivers always at 60m height. It's not clear how well this can generalize to a new scene or real-world environments.

Questions for authors
  1. How well does the method generalize to substantially different environments? What evidence supports transfer to real wireless measurements? Why not rest the proposed method on real-world datasets?

  2. Why are TX and RX heights fixed within each experimental setting, rather than varying across devices and locations? The receiver-height experiment changes one horizontal plane to another, but does not test receivers distributed across different heights. What application motivates this restriction, and how sensitive is the method to this assumption?

  3. KNN seems to be the strongest baseline on all three unseen cities, with Flow-RI reducing MAE by only approximately 0.3 dB. This suggests that spatial interpolation already captures much of the structure under the evaluated conditions. How much additional benefit comes from the learned flow prior rather than the spatial smoothness? What downstream improvement results from these sub-1 dB improvement, and does it justify the additional training and inference cost?

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

Related papers

Discussion (0)

Sign in to comment.