Segment-Wise Flow Matching for Vision-Aided mmWave V2I Beam Prediction

Can Zheng; Chongwen Huang; Chung G. Kang; Guofa Cai; Henk Wymeersch; Jiguang He

arxiv: 2511.20265 · v2 · submitted 2025-11-25 · 📡 eess.SP

Segment-Wise Flow Matching for Vision-Aided mmWave V2I Beam Prediction

Can Zheng , Jiguang He , Chung G. Kang , Guofa Cai , Chongwen Huang , Henk Wymeersch This is my paper

Pith reviewed 2026-05-17 04:48 UTC · model grok-4.3

classification 📡 eess.SP

keywords flow matchingbeam predictionmmWave V2Ivision-aidedcontinuous dynamicswireless communicationsmachine learninginference latency

0 comments

The pith

A vision-conditioned flow matching model learns continuous dynamics of beam receive power vectors to enable accurate low-latency prediction in mmWave V2I links.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper sets out a framework that conditions a flow matching process on vision data to capture how normalized beam receive power vectors evolve over time in millimeter-wave vehicle-to-infrastructure settings. Instead of classifying discrete beam indices, the method learns a continuous vector field defined by an ordinary differential equation so that future states can be obtained by integrating the field forward. The same model is trained to satisfy both the prediction task and the consistency of the learned flow, producing a single mechanism that generates smooth trajectories. Experiments indicate this yields higher prediction accuracy than standard baselines while coming close to the results of large language model approaches and cutting inference time substantially on both GPU and CPU hardware.

Core claim

The paper claims that imposing flow matching on the segment-wise transitions of normalized beam receive power vectors, when conditioned on vision inputs, produces a unified model whose learned continuous vector field can be integrated to forecast future beams, delivering improved prediction performance over discrete baselines, performance near that of large language model methods, and substantially lower predictor-side inference latency.

What carries the argument

Vision-conditioned flow matching that learns a continuous vector field governing the temporal evolution of normalized beam receive power vectors via an ordinary differential equation.

If this is right

Beam prediction accuracy rises markedly compared with conventional discrete-sequence baselines.
Prediction quality reaches levels comparable to those of large language model-based predictors.
Predictor-side inference latency drops by roughly 6.9 times on GPU hardware and by roughly 2800 times on CPU hardware.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The continuous formulation may permit sampling of intermediate beam states between the discrete prediction instants required by the link.
Because the flow is learned jointly with prediction, the same model could support variable-length prediction horizons without retraining separate heads.
If the underlying channel dynamics contain abrupt changes not captured by smooth flows, the method may need explicit segmentation or hybrid discrete-continuous extensions.

Load-bearing premise

The learned continuous vector field must accurately represent the real-world temporal changes in beam receive power vectors and vision data must supply sufficient conditioning without domain shift or alignment problems during actual deployment.

What would settle it

Time-series measurements of actual beam receive powers collected from a moving vehicle in a real mmWave V2I environment that fail to match the sequences obtained by integrating the model's learned vector field would show the central claim is incorrect.

Figures

Figures reproduced from arXiv: 2511.20265 by Can Zheng, Chongwen Huang, Chung G. Kang, Guofa Cai, Henk Wymeersch, Jiguang He.

**Figure 2.** Figure 2: Architecture of the proposed FM-based beam predicti [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Figure 3: Training loss curves over 100 epochs in Scenario 8 [16 [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Figure 4: ACCK performance of the proposed method compared to several baselines under configuration A (THist = 8 and TPred = 5). B. Results and Discussion 1) Training Dynamics and Convergence [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

**Figure 5.** Figure 5: ACCK performance of the proposed method compared to baselines in under configuration B (THist = 3 and TPred = 10) . TABLE III: Average ACCK performance of FM under different environmental scenarios. Scenarios Description # Samples Top-1 Top-3 Scenario 1 Day-time, McAllister Ave 1 2411 0.62 0.91 Scenario 2 Night-time, McAllister Ave 1 2974 0.53 0.84 Scenario 5 Night-time, rainy, Tyler St. 2300 0.58 0.95 Sc… view at source ↗

read the original abstract

This paper proposes a vision-conditioned flow matching (FM) framework for beam prediction in millimeter-wave vehicle-to-infrastructure links. Instead of modeling discrete beam-index sequences, the proposed method learns the temporal evolution of normalized beam receive power vectors through a continuous vector field governed by an ordinary differential equation, enabling smooth dynamics and efficient sampling. By imposing FM over beam-state transitions and jointly optimizing beam prediction and flow consistency, the proposed framework provides a unified model for future beam prediction. Experimental results show that the proposed FM-based model significantly improves beam prediction performance over baselines, approaches the performance of large language model-based methods, and reduces predictor-side inference latency by about $6.9\times$ on GPU and $2.8\times10^3\times$ on CPU, respectively.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper shifts beam prediction to a continuous flow-matching vector field on normalized power vectors conditioned on vision, which is a clear step past discrete index methods, but the gains and latency claims rest on abstract-level evidence without visible ablations or trajectory checks.

read the letter

The main point is that this work moves away from classifying discrete beam indices and instead learns a continuous vector field for the normalized receive-power vector using segment-wise flow matching, conditioned on vision data in mmWave V2I settings. The ODE formulation lets the model generate smooth trajectories and sample efficiently, and the joint loss on prediction plus flow consistency is a straightforward way to keep everything tied together. That framing is genuinely different from the usual discrete or simple regression baselines cited in the abstract, and the reported latency drops (roughly 7x on GPU, thousands on CPU) would matter if they hold for real predictor-side use in vehicular links. The idea of treating beam-state evolution as a learned dynamical system is worth looking at for anyone working on continuous channel modeling. The soft spots are mostly in the experimental grounding. The abstract states clear performance lifts and latency wins, yet gives no dataset size or source, no description of the exact baselines, no ablation that isolates the flow-matching component from the vision encoder, and no check that integrated trajectories actually match held-out measurement sequences or respect physical Doppler effects. The stress-test concern lands: if the edge comes mainly from better vision conditioning or joint optimization rather than the verified continuous dynamics, then a simpler continuous regressor could deliver similar results without the ODE overhead. From the abstract alone it is hard to tell how load-bearing the flow-matching part really is. This paper is for wireless researchers who already use ML for beam management and are willing to try generative or dynamical models. A reader focused on practical 5G/6G beam prediction would get value from the framework and the latency angle, provided the full experiments and code are solid. It deserves a serious referee because the core modeling choice is distinct and the application is timely; the review would mainly push for the missing ablations, dataset details, and trajectory validation. I would send it to review rather than desk-reject.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a vision-conditioned segment-wise flow matching framework for mmWave V2I beam prediction. Rather than treating beam indices as discrete sequences, it learns a continuous vector field over normalized receive-power vectors governed by an ODE, trained end-to-end with flow matching on temporal transitions while jointly optimizing prediction accuracy and flow consistency. The central experimental claim is that the resulting model outperforms conventional baselines, approaches the accuracy of large-language-model methods, and delivers substantial predictor-side latency reductions (approximately 6.9× on GPU and 2.8×10³× on CPU).

Significance. If the performance and latency claims are substantiated with rigorous controls, the work could meaningfully advance real-time beam management in high-mobility mmWave vehicular links by replacing discrete classification with continuous dynamics and multi-modal conditioning. The reported CPU latency improvement would be especially relevant for edge deployment; however, the significance hinges on demonstrating that the flow-matching component itself, rather than vision conditioning alone, drives the gains.

major comments (2)

[Abstract] Abstract: the headline performance and latency claims are presented without dataset statistics, baseline specifications, ablation studies isolating the flow-matching ODE from the vision encoder, or statistical significance tests. These omissions make it impossible to determine whether the continuous vector field is load-bearing for the reported improvements or whether gains could be replicated by a simpler continuous regressor.
[Proposed method / experimental results] Proposed method / experimental results: the central modeling assumption—that the learned vector field accurately represents real-world temporal evolution of beam receive-power vectors—requires direct validation. The manuscript should report trajectory-matching metrics on held-out measurement sequences or consistency checks against known mmWave dynamics (e.g., Doppler-induced abrupt changes) to confirm that the ODE integration captures physical channel behavior rather than merely benefiting from joint vision optimization.

minor comments (2)

Clarify the precise definition of 'segments' used in the segment-wise flow matching procedure and how segment boundaries are chosen or aligned with vision frames.
Ensure all figures include error bars or confidence intervals when reporting prediction accuracy or latency, and expand the related-work discussion to include recent flow-matching applications in wireless signal processing.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and insightful comments on our manuscript. We address each major comment point by point below, indicating the revisions made to strengthen the work while maintaining scientific rigor.

read point-by-point responses

Referee: [Abstract] Abstract: the headline performance and latency claims are presented without dataset statistics, baseline specifications, ablation studies isolating the flow-matching ODE from the vision encoder, or statistical significance tests. These omissions make it impossible to determine whether the continuous vector field is load-bearing for the reported improvements or whether gains could be replicated by a simpler continuous regressor.

Authors: We agree that the abstract, due to length constraints, does not include supporting details. The full manuscript provides dataset statistics and collection methodology in Section III, baseline specifications and implementation details in Section IV, and ablation studies in Section V that compare the proposed segment-wise flow matching against variants without the ODE component. To directly address the concern about whether the flow-matching ODE is load-bearing, we have added a new ablation in the revised manuscript that isolates the continuous vector field from the vision encoder by comparing against a vision-conditioned MLP regressor trained on the same normalized power vectors. This ablation shows a consistent accuracy advantage for the flow-matching approach. We have also incorporated statistical significance testing (paired t-tests with reported p-values) into the main results table. These changes clarify the contribution of the continuous dynamics. revision: yes
Referee: [Proposed method / experimental results] Proposed method / experimental results: the central modeling assumption—that the learned vector field accurately represents real-world temporal evolution of beam receive-power vectors—requires direct validation. The manuscript should report trajectory-matching metrics on held-out measurement sequences or consistency checks against known mmWave dynamics (e.g., Doppler-induced abrupt changes) to confirm that the ODE integration captures physical channel behavior rather than merely benefiting from joint vision optimization.

Authors: We acknowledge the value of direct validation for the learned dynamics. The current evaluation already uses held-out temporal sequences to measure multi-step prediction accuracy, providing indirect evidence that the vector field captures relevant evolution. However, we agree that explicit trajectory-matching metrics would strengthen the claim. In the revised manuscript, we have added quantitative trajectory-matching results on held-out sequences, reporting average L2 error between ODE-integrated paths and ground-truth normalized power vectors over varying horizons. We also include qualitative and quantitative checks showing the model's response to abrupt power changes consistent with Doppler shifts in high-mobility mmWave scenarios. These additions demonstrate that the performance gains arise from modeling the continuous temporal evolution rather than vision conditioning in isolation. revision: yes

Circularity Check

0 steps flagged

No circularity: standard end-to-end trained neural flow-matching model with independent experimental validation

full rationale

The paper presents a data-driven vision-conditioned flow-matching framework that learns a continuous vector field via an ODE for beam-power evolution. No equations reduce predictions to fitted parameters by construction, no self-citation chains justify core premises, and no ansatz or uniqueness result is smuggled in. The derivation chain consists of standard FM training objectives and joint optimization, which remain independent of the target beam-prediction outputs. Experimental claims rest on held-out test performance rather than self-referential fits.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard neural network training assumptions plus the domain-specific premise that beam power vectors admit a smooth ODE-governed evolution; no new physical entities are introduced.

free parameters (1)

neural network weights and flow matching hyperparameters
Learned during training to define the vector field and ODE integration; values not specified in abstract.

axioms (1)

domain assumption Normalized beam receive power vectors evolve according to a continuous vector field governed by an ODE
Invoked to justify replacing discrete beam-index sequences with flow matching over state transitions.

pith-pipeline@v0.9.0 · 5442 in / 1214 out tokens · 37623 ms · 2026-05-17T04:48:37.057746+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

learns the temporal evolution of normalized beam receive power vectors through a continuous vector field governed by an ordinary differential equation
IndisputableMonolith/Foundation/AlphaCoordinateFixation.lean alpha_pin_under_high_calibration unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

terminal flow constraint enforces global consistency under finite-step integration

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages

[1]

Study on channel model for frequ ency spectrum above 6 GHz,

3GPP TR 38.900 V15.0.0, “Study on channel model for frequ ency spectrum above 6 GHz,” Tech. Rep., Jul. 2018

work page 2018
[2]

Deep learni ng for mmwave beam-management: State-of-the-art, opportunities and chal- lenges,

K. Ma, Z. Wang, W. Tian, S. Chen, and L. Hanzo, “Deep learni ng for mmwave beam-management: State-of-the-art, opportunities and chal- lenges,” IEEE Wireless Commun. , vol. 30, no. 4, pp. 108–114, 2023

work page 2023
[3]

Millimeter wave base stations with cameras: Vision-aided beam and blockage pred iction,

M. Alrabeiah, A. Hredzak, and A. Alkhateeb, “Millimeter wave base stations with cameras: Vision-aided beam and blockage pred iction,” in Proc. IEEE V ehicular Technology Conference (VTC2020-Spring), 2020

work page 2020
[4]

Computer vision aided beam tr acking in A real-world millimeter wave deployment,

S. Jiang and A. Alkhateeb, “Computer vision aided beam tr acking in A real-world millimeter wave deployment,” in Proc. IEEE Globecom W orkshops (GC Wkshps), 2022, pp. 142–147

work page 2022
[5]

Multimodal transformers for wireless communications: A c ase study in beam prediction,

Y . Tian, Q. Zhao, Z. e. a. Kherroubi, F. Boukhalfa, K. Wu, a nd F. Bader, “Multimodal transformers for wireless communications: A c ase study in beam prediction,” ITU Journal on Future and Evolving Technologies , vol. 4, no. 3, pp. 461–471, 2023

work page 2023
[6]

BeamLLM: Vis ion- empowered mmwave beam prediction with large language model s,

C. Zheng, J. He, G. Cai, Z. Y u, and C. G. Kang, “BeamLLM: Vis ion- empowered mmwave beam prediction with large language model s,” arXiv preprint arXiv:2503.10432 , 2025

work page arXiv 2025
[7]

Large la nguage models empower multimodal integrated sensing and communic ation,

L. Cheng, H. Zhang, B. Di, D. Niyato, and L. Song, “Large la nguage models empower multimodal integrated sensing and communic ation,” vol. 63, no. 5, pp. 190–197, 2025

work page 2025
[8]

Flow matching for generative modeling,

Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” in Proc. International Conference on Learning Representations (ICLR) , May 2023

work page 2023
[9]

Pyramidal ﬂow matching for efﬁcie nt video generative modeling,

Y . Jin, Z. Sun, N. Li, K. Xu, K. Xu, H. Jiang, N. Zhuang, Q. Hu ang, Y . Song, Y . MU, and Z. Lin, “Pyramidal ﬂow matching for efﬁcie nt video generative modeling,” in Proc. International Conference on Learn- ing Representations (ICLR) , May 2025

work page 2025
[10]

Gener- ative pre-training for speech with ﬂow matching,

A. H. Liu, M. Le, A. Vyas, B. Shi, A. Tjandra, and W.-N. Hsu , “Gener- ative pre-training for speech with ﬂow matching,” in Proc. International Conference on Learning Representations (ICLR) , May 2024

work page 2024
[11]

Flow matching-based autonomous driving planni ng with advanced interactive behavior modeling,

T. Tan, Y . Zheng, R. Liang, Z. Wang, K. Zheng, J. Zheng, J. Li, X. Zhan, and J. Liu, “Flow matching-based autonomous driving planni ng with advanced interactive behavior modeling,” in Proc. Annual Conference on Neural Information Processing Systems (NeurIPS) , Dec. 2025

work page 2025
[12]

Technical speciﬁcation group radio access network; Study on artiﬁcial intelligence (AI)/machine lea rning (ML) for NR air interface,

3GPP TR 38.843 V18.0.0, “Technical speciﬁcation group radio access network; Study on artiﬁcial intelligence (AI)/machine lea rning (ML) for NR air interface,” Tech. Rep., Dec. 2023

work page 2023
[13]

Sched uled sam- pling for sequence prediction with recurrent neural networ ks,

S. Bengio, O. Vinyals, N. Jaitly, and N. Shazeer, “Sched uled sam- pling for sequence prediction with recurrent neural networ ks,” in Proc. Advances Neural Information Processing Systems (NIPS) , 2015, p. 1171–1179

work page 2015
[14]

Beam-based mo bility management in 5g millimetre wave v2x communications: A surv ey and outlook,

A. Kose, H. Lee, C. H. Foh, and M. Dianati, “Beam-based mo bility management in 5g millimetre wave v2x communications: A surv ey and outlook,” IEEE Open J. Intell. Transp. Syst. , vol. 2, pp. 347–363, 2021

work page 2021
[15]

Flow straight and fast: Lear ning to generate and transfer data with rectiﬁed ﬂow,

X. Liu, C. Gong, and Q. Liu, “Flow straight and fast: Lear ning to generate and transfer data with rectiﬁed ﬂow,” in Proc. International Conference on Learning Representations (ICLR) , May 2023

work page 2023
[16]

DeepSense 6G: a large-scale r eal-world multi-modal sensing and communication dataset,

A. Alkhateeb, G. Charan, T. Osman, A. Hredzak, J. Morais , U. Demirhan, and N. Srinivas, “DeepSense 6G: a large-scale r eal-world multi-modal sensing and communication dataset,” IEEE Commun. Mag. , vol. 61, no. 9, pp. 122–128, Sept. 2023

work page 2023
[17]

AI/ML for b eam management in 5G-Advanced: A standardization perspective,

Q. Xue, J. Guo, B. Zhou, Y . Xu, Z. Li, and S. Ma, “AI/ML for b eam management in 5G-Advanced: A standardization perspective,” IEEE V eh. Technol. Mag., vol. 19, no. 4, pp. 64–72, 2024

work page 2024

[1] [1]

Study on channel model for frequ ency spectrum above 6 GHz,

3GPP TR 38.900 V15.0.0, “Study on channel model for frequ ency spectrum above 6 GHz,” Tech. Rep., Jul. 2018

work page 2018

[2] [2]

Deep learni ng for mmwave beam-management: State-of-the-art, opportunities and chal- lenges,

K. Ma, Z. Wang, W. Tian, S. Chen, and L. Hanzo, “Deep learni ng for mmwave beam-management: State-of-the-art, opportunities and chal- lenges,” IEEE Wireless Commun. , vol. 30, no. 4, pp. 108–114, 2023

work page 2023

[3] [3]

Millimeter wave base stations with cameras: Vision-aided beam and blockage pred iction,

M. Alrabeiah, A. Hredzak, and A. Alkhateeb, “Millimeter wave base stations with cameras: Vision-aided beam and blockage pred iction,” in Proc. IEEE V ehicular Technology Conference (VTC2020-Spring), 2020

work page 2020

[4] [4]

Computer vision aided beam tr acking in A real-world millimeter wave deployment,

S. Jiang and A. Alkhateeb, “Computer vision aided beam tr acking in A real-world millimeter wave deployment,” in Proc. IEEE Globecom W orkshops (GC Wkshps), 2022, pp. 142–147

work page 2022

[5] [5]

Multimodal transformers for wireless communications: A c ase study in beam prediction,

Y . Tian, Q. Zhao, Z. e. a. Kherroubi, F. Boukhalfa, K. Wu, a nd F. Bader, “Multimodal transformers for wireless communications: A c ase study in beam prediction,” ITU Journal on Future and Evolving Technologies , vol. 4, no. 3, pp. 461–471, 2023

work page 2023

[6] [6]

BeamLLM: Vis ion- empowered mmwave beam prediction with large language model s,

C. Zheng, J. He, G. Cai, Z. Y u, and C. G. Kang, “BeamLLM: Vis ion- empowered mmwave beam prediction with large language model s,” arXiv preprint arXiv:2503.10432 , 2025

work page arXiv 2025

[7] [7]

Large la nguage models empower multimodal integrated sensing and communic ation,

L. Cheng, H. Zhang, B. Di, D. Niyato, and L. Song, “Large la nguage models empower multimodal integrated sensing and communic ation,” vol. 63, no. 5, pp. 190–197, 2025

work page 2025

[8] [8]

Flow matching for generative modeling,

Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” in Proc. International Conference on Learning Representations (ICLR) , May 2023

work page 2023

[9] [9]

Pyramidal ﬂow matching for efﬁcie nt video generative modeling,

Y . Jin, Z. Sun, N. Li, K. Xu, K. Xu, H. Jiang, N. Zhuang, Q. Hu ang, Y . Song, Y . MU, and Z. Lin, “Pyramidal ﬂow matching for efﬁcie nt video generative modeling,” in Proc. International Conference on Learn- ing Representations (ICLR) , May 2025

work page 2025

[10] [10]

Gener- ative pre-training for speech with ﬂow matching,

A. H. Liu, M. Le, A. Vyas, B. Shi, A. Tjandra, and W.-N. Hsu , “Gener- ative pre-training for speech with ﬂow matching,” in Proc. International Conference on Learning Representations (ICLR) , May 2024

work page 2024

[11] [11]

Flow matching-based autonomous driving planni ng with advanced interactive behavior modeling,

T. Tan, Y . Zheng, R. Liang, Z. Wang, K. Zheng, J. Zheng, J. Li, X. Zhan, and J. Liu, “Flow matching-based autonomous driving planni ng with advanced interactive behavior modeling,” in Proc. Annual Conference on Neural Information Processing Systems (NeurIPS) , Dec. 2025

work page 2025

[12] [12]

Technical speciﬁcation group radio access network; Study on artiﬁcial intelligence (AI)/machine lea rning (ML) for NR air interface,

3GPP TR 38.843 V18.0.0, “Technical speciﬁcation group radio access network; Study on artiﬁcial intelligence (AI)/machine lea rning (ML) for NR air interface,” Tech. Rep., Dec. 2023

work page 2023

[13] [13]

Sched uled sam- pling for sequence prediction with recurrent neural networ ks,

S. Bengio, O. Vinyals, N. Jaitly, and N. Shazeer, “Sched uled sam- pling for sequence prediction with recurrent neural networ ks,” in Proc. Advances Neural Information Processing Systems (NIPS) , 2015, p. 1171–1179

work page 2015

[14] [14]

Beam-based mo bility management in 5g millimetre wave v2x communications: A surv ey and outlook,

A. Kose, H. Lee, C. H. Foh, and M. Dianati, “Beam-based mo bility management in 5g millimetre wave v2x communications: A surv ey and outlook,” IEEE Open J. Intell. Transp. Syst. , vol. 2, pp. 347–363, 2021

work page 2021

[15] [15]

Flow straight and fast: Lear ning to generate and transfer data with rectiﬁed ﬂow,

X. Liu, C. Gong, and Q. Liu, “Flow straight and fast: Lear ning to generate and transfer data with rectiﬁed ﬂow,” in Proc. International Conference on Learning Representations (ICLR) , May 2023

work page 2023

[16] [16]

DeepSense 6G: a large-scale r eal-world multi-modal sensing and communication dataset,

A. Alkhateeb, G. Charan, T. Osman, A. Hredzak, J. Morais , U. Demirhan, and N. Srinivas, “DeepSense 6G: a large-scale r eal-world multi-modal sensing and communication dataset,” IEEE Commun. Mag. , vol. 61, no. 9, pp. 122–128, Sept. 2023

work page 2023

[17] [17]

AI/ML for b eam management in 5G-Advanced: A standardization perspective,

Q. Xue, J. Guo, B. Zhou, Y . Xu, Z. Li, and S. Ma, “AI/ML for b eam management in 5G-Advanced: A standardization perspective,” IEEE V eh. Technol. Mag., vol. 19, no. 4, pp. 64–72, 2024

work page 2024