SynthSoM-V2I: Multi-Modal V2I Benchmark

• SynthSoM-V2I is a multi-modal sensing–communication platform that aligns mmWave radar, RGB-D images, and LiDAR data to generate fine-grained V2I channel multipath labels for 6G scenarios.
• It employs synchronized dual-view data from transmitter and receiver perspectives using digital twin consistency across AirSim, WaveFarer, and Sionna RT.
• Leveraging LLM4MG with a fine-tuned LLaMA 3.2 and LoRA, the approach achieves high LoS/NLoS classification accuracy and low NMSEs in multipath power and delay regression.

Searching arXiv for papers on “SynthSoM-V2I” and closely related SoM/V2I work to ground the article in current literature. SynthSoM-V2I denotes a Synesthesia of Machines (SoM)-oriented vehicle-to-infrastructure research construct centered on tightly aligned sensing and communication data for V2I learning problems. In its most specific and technically concrete usage, it is a multi-modal sensing–communication dataset introduced for fine-grained channel multipath generation in a 6G V2I setting, containing aligned mmWave radar point clouds, RGB-D images, LiDAR point clouds, and channel multipath ground truth. In the broader SoM literature, the term also refers to a V2I perspective on the earlier SynthSoM platform and to an FM-empowered V2I system abstraction that unifies RF and non-RF sensing with wireless communications (Huang et al., 18 Sep 2025, Cheng et al., 13 Jan 2025, Cheng et al., 9 Jun 2025).

1. Terminological scope and lineage

The literature uses “SynthSoM-V2I” in closely related but non-identical ways. The narrowest usage names a dataset purpose-built for multipath generation. A broader usage treats it as a V2I instantiation of the SoM paradigm, in which roadside infrastructure and vehicles are jointly modeled through aligned sensing and channel data. A still earlier precursor is the original SynthSoM dataset, whose paper states that the term “SynthSoM‑V2I” is not explicitly used, although vehicle-to-infrastructure interaction is intrinsic through roadside facilities acting as infrastructure nodes (Huang et al., 18 Sep 2025, Cheng et al., 13 Jan 2025, Cheng et al., 9 Jun 2025).

Usage	Source	Scope
V2I perspective on SynthSoM	(Cheng et al., 13 Jan 2025)	RSF-to-vehicle RF, radar, RGB, depth, LiDAR
FM-empowered SoM V2I system	(Cheng et al., 9 Jun 2025)	6G V2I architecture with LLMs and wireless FMs
Dedicated multipath-generation dataset	(Huang et al., 18 Sep 2025)	211,395 aligned V2I snapshots

This terminological variation matters because “SynthSoM-V2I” is sometimes treated as a single fixed benchmark, whereas the cited literature actually spans a lineage: a general synthetic sensing–communication platform, a V2I-oriented SoM systems perspective, and a later dataset specialized for channel multipath generation. A common misconception is therefore to equate the entire lineage with only the LLM4MG dataset release. The more precise view is that the dataset is the most concrete instantiation, while the surrounding SoM literature supplies the conceptual and methodological frame.

2. Dataset construction and sensing–communication composition

In the dedicated dataset formulation, SynthSoM-V2I is a multi-modal sensing–communication dataset for a 6G V2I scenario, built to learn mappings from environmental sensing to per-path channel structure. Its modalities are mmWave radar sensory data, RGB-D images, LiDAR point clouds, and channel multipath labels. The sensing stack is explicitly dual-view: data are generated for both the vehicle-side transmitter view and the base-station-side receiver view, and the resulting multi-view features are concatenated after modality-specific encoding (Huang et al., 18 Sep 2025).

The sensing and communications configuration is tightly specified. The mmWave radar operates in the 77–78 GHz band with chirp length $60\,\mu s$ and echo sampling interval $0.2\,\mu s$, simulated with Remcom WaveFarer. RGB-D imagery is rendered in AirSim at $960\times540$ resolution for both Tx and Rx viewpoints. LiDAR is also generated in AirSim, using 16 channels at 20 Hz, followed by denoising to retain stable structures and moving objects. Channel multipath ground truth is produced by Sionna RT for a SISO V2I link in which the transmitter is the vehicle and the receiver is the base station; the vehicle uses one phased-array element configured for omnidirectional transmission, and the BS uses one element. Two communications settings are reported: sub-6 GHz at 5.9 GHz with 20 MHz bandwidth, and mmWave at 60 GHz with 2 GHz bandwidth (Huang et al., 18 Sep 2025).

Temporal consistency is a defining property. All modalities are synchronized at 33.33 ms intervals, and sensor-to-channel alignment is attributed to digital-twin consistency across AirSim, WaveFarer, and Sionna RT. This suggests that the dataset is designed not merely for late fusion of heterogeneous modalities, but for learning physically grounded cross-modal correspondences at the snapshot level.

3. Scenario design, labels, and benchmark formulation

The dataset contains 211,395 multi-modal snapshots. Four representative cases are reported: urban, low vehicular traffic density, sub-6 GHz; urban, high vehicular traffic density, mmWave; urban, low vehicular traffic density, mmWave; and suburban, low vehicular traffic density, mmWave. The train/validation/test split is $3{:}1{:}1$. The paper also states that, for each modality, 42,279 snapshots are reported in aggregate across the included cases (Huang et al., 18 Sep 2025).

The channel labels comprise LoS/NLoS flags together with per-path powers $P_i$ and delays $\tau_i$ at each snapshot. Phases are not provided as labels; instead, when impulse or frequency responses are reconstructed, phases are assigned randomly within stationary intervals. This is an important technical boundary: SynthSoM-V2I is a power–delay benchmark rather than a full phase-resolved channel benchmark. Another common misconception is therefore to assume that coherent per-path phase supervision is included; it is not (Huang et al., 18 Sep 2025).

The multipath-generation task is posed over a fixed budget of $N=6$ dominant paths, which the paper states captures approximately 90% of received power. Benchmarking uses both classification and regression metrics. LoS/NLoS classification accuracy is defined as

$$A_{\mathrm{cls}}=\frac{N_{\mathrm{accurate}}}{N_{\mathrm{all}}}.$$

For multipath power and delay generation, the reported normalized mean square errors are

$$\mathrm{NMSE}_{\mathrm{power}}=\frac{\sum_{n=1}^{N}\mu_n(\omega_n-\hat{\omega}_n)^2}{\sum_{n=1}^{N}\omega_n^2}, \qquad \mathrm{NMSE}_{\mathrm{delay}}=\frac{\sum_{n=1}^{N}(\epsilon_n-\hat{\epsilon}_n)^2}{\sum_{n=1}^{N}\epsilon_n^2},$$

where $\mu_n=3$ for the strongest path and $0.2\,\mu s$0 otherwise. The benchmark also reports NMAE, time-varying power delay profile, mean delay, RMS delay spread, and frequency correlation function. The PDP is written as

$0.2\,\mu s$1

with the corresponding mean delay and RMS delay spread derived from $0.2\,\mu s$2 (Huang et al., 18 Sep 2025).

These choices reflect a benchmark emphasis on system-relevant channel structure rather than purely token-level generation fidelity. Because the outputs remain continuous and non-quantized, the dataset is particularly aligned with regression-heavy physical-layer tasks.

4. LLM4MG as the principal modeling framework

SynthSoM-V2I is introduced as the data foundation for LLM4MG, described as the first system adapting a LLM, specifically LLaMA 3.2, to multipath generation via SoM. The pipeline uses modality-specific front ends: ViT for RGB-D, PointNet++ for LiDAR, and RadarBEVNet (RCBEVDet) for mmWave radar point clouds. For each viewpoint, the modality features are concatenated and fused by ECA-Net, whose kernel size is given by

$0.2\,\mu s$3

with $0.2\,\mu s$4 and $0.2\,\mu s$5. The Tx-side and Rx-side fused features are then concatenated into a 2048-dimensional descriptor matched to the input embedding dimension of LLaMA 3.2 (Huang et al., 18 Sep 2025).

The model is conditioned not only on fused sensor features but also on propagation-aware prompts. These prompts encode carrier frequency, bandwidth, Tx–Rx distance, antenna azimuth and elevation, scenario tag, and vehicular traffic density. The paper gives example prompt structures such as an “urban-low-VTD” 60 GHz case and a “sub-6 GHz” suburban case. This prompt design is meant to inject domain priors alongside sensory evidence, rather than relying exclusively on latent scene encoding (Huang et al., 18 Sep 2025).

Parameter-efficient fine-tuning is implemented through LoRA in the feed-forward linear layers of LLaMA 3.2. For a pretrained matrix $0.2\,\mu s$6, the adapted weight is

$0.2\,\mu s$7

where $0.2\,\mu s$8, $0.2\,\mu s$9, $960\times540$0, and $960\times540$1. The decoder depth is set to 2. Instead of generating from a large vocabulary, LLM4MG attaches lightweight task heads after the LLaMA decoder: a classification head for LoS/NLoS and regression heads for per-path power and delay. Training uses a multi-task loss combining cross-entropy for classification and the reported NMSE objectives for regression. The schedule includes a 3-epoch warm-up, cosine annealing, total training length of 100 epochs, batch size 24, AdamW, and LoRA activation after epoch 10 (Huang et al., 18 Sep 2025).

A plausible implication is that SynthSoM-V2I is designed as much for studying cross-modal semantic transfer into wireless structure as for benchmarking a specific regression model. The use of LLaMA 3.2, prompt conditioning, and LoRA explicitly positions the dataset within the FM-oriented branch of SoM research.

5. Quantitative results, generalization, and system-level validation

The headline LLM4MG results on SynthSoM-V2I are a LoS/NLoS classification accuracy of 92.76%, multipath power NMSE of 0.099, and multipath delay NMSE of 0.032. In the detailed urban low-VTD mmWave evaluation, the paper reports power NMSE $960\times540$2, delay NMSE $960\times540$3, power NMAE $960\times540$4, and delay NMAE $960\times540$5. Against the reported baselines—MLP, ResNet-34, and Transformer—LLM4MG achieves the highest classification accuracy and the lowest power and delay errors. The same section states that the derived PDPs, RMS delay spreads, and FCFs align closely with ray-tracing ground truth, whereas the baselines show larger deviations and weaker spatial consistency across snapshots (Huang et al., 18 Sep 2025).

Generalization is a central empirical claim. In low-to-high vehicular traffic density transfer, the model matches full-shot baseline performance with less than 1.4% of training samples. In 60 GHz to 5.9 GHz transfer, it reaches comparable power generation with less than 1.3% of samples and comparable delay generation with about 1.4% of samples. In urban-to-suburban transfer, it attains comparable power generation with less than 1.6% and delay generation with less than 1.0% of samples. Efficiency figures are also reported: LLM4MG has 0.16M trainable parameters out of 121.81M total, training time per batch of 9.24 ms, and inference time per batch of 5.92 ms (Huang et al., 18 Sep 2025).

The ablation study attributes measurable performance losses to removing multi-modality, removing propagation prompts, or removing or freezing the LLM backbone. The real-world validation pairs the synthetic urban-low-VTD-mmWave case with DeepSense 6G Scenario 32 in downtown Tempe, using Real2Real, Sim2Real, and Mixed2Real protocols; Mixed2Real performs best, and Sim2Real improves over Real2Real. At the system level, the generated multipath is used for channel-capacity estimation. In the reported sub-6 GHz urban-low-VTD setting, LLM4MG reaches 96.20% capacity accuracy relative to ray tracing, more than 30% higher than the deep-learning baselines. This capacity result is used to argue that high-precision multipath generation is not only a modeling exercise but also consequential for downstream PHY design (Huang et al., 18 Sep 2025).

6. Position within the broader SoM and V2I research landscape

The dedicated SynthSoM-V2I dataset is best understood as a specialization of a broader SoM program. The original SynthSoM dataset integrates AirSim, WaveFarer, and Wireless InSite, covers five road layouts, spans sunny, rainy, and snowy conditions as well as morning and night, includes high, medium, and low agent densities, and provides RF channels, mmWave radar, RGB, depth, and LiDAR in a dataset of approximately 3 TB. In that earlier formulation, roadside facilities function as V2I infrastructure nodes, but the emphasis is general air–ground multi-link SoM rather than per-path V2I multipath generation (Cheng et al., 13 Jan 2025).

The FM-oriented SoM framework later categorizes two design roadmaps—LLM-based and wireless-foundation-model-based—and explicitly frames SynthSoM-V2I as a 6G V2I SoM system. Within that roadmap, LLM4PG and LLM4SG address path-loss and scatterer generation, WiCo addresses visual-to-radio mechanism exploration, WiFo targets CSI prediction, LLM4WM targets multi-task transceiver functions, and WiPo targets modality-agnostic cooperative perception. The dedicated multipath-generation dataset can therefore be read as one concrete substrate for this broader FM agenda (Cheng et al., 9 Jun 2025).

Adjacent V2I literature shows why such a dataset is useful. Robust predictive routing studies combine V2I and V2V links to preserve QoS under blockage and coverage loss (Chang et al., 31 Mar 2025). Vision-centric HD mapping work uses coordinated vehicle and roadside cameras to construct vectorized maps in a unified BEV (Fan et al., 31 Mar 2025). V2I work-zone geometry reconstruction uses pose-conditioned, permutation-equivariant UWB range denoising to infer cone layouts from multi-anchor ranging (Liu et al., 28 May 2026). SoM-empowered wireless image transmission explicitly conditions JSCC-like image transmission on SNR, CSI, Doppler, and channel-aging delay (Li et al., 14 Sep 2025). Taken together, these works suggest that SynthSoM-V2I occupies a specific but strategically important niche: it does not attempt to solve routing, HD mapping, geometry reconstruction, or dynamic image transmission directly, but it provides the kind of aligned sensing–propagation substrate that such tasks increasingly depend on.

The dataset also inherits several explicit limitations. The labels do not include phase, and AoA, AoD, and Doppler are not generation targets. The reported study focuses on SISO rather than MIMO. Synthetic-to-real domain gaps remain nontrivial, especially under missing vehicle-side real sensing or missing real depth information. Extending the benchmark toward richer multipath attributes, adverse-weather sensing degradation, and MIMO channel structure is therefore a natural direction. This suggests that SynthSoM-V2I is presently most valuable as a benchmark for structured channel generation and cross-modal SoM mechanism learning, rather than as a complete V2I systems benchmark (Huang et al., 18 Sep 2025).