QuantV2X: Quantized V2X Systems & Applications

• QuantV2X is a research domain unifying low-bit quantization in V2X systems, enabling cooperative perception, performance benchmarking, and quantum-safe security solutions.
• It uses deployment-oriented methods to reduce latency, memory, bandwidth, and energy while maintaining system-level accuracy in multi-agent environments.
• Applications span fully quantized cooperative perception modules, efficient feature compression via residual vector quantization, and quantum-inspired optimization for secure vehicular networks.

QuantV2X is used in several distinct but related ways in recent Vehicle-to-Everything research. In its most specific use, it denotes a fully quantized multi-agent intermediate-fusion system for cooperative perception that applies low-bit quantization to encoders, fusion modules, detection heads, and transmitted message representations, with deployment-oriented emphasis on latency, memory, and bandwidth (Zhao et al., 3 Sep 2025). The same label is also used for quantitative characterization of C‑V2X highway sidelink performance under varying propagation models (Wang et al., 2019), for practical feature-quantization modules built around residual vector quantization and index-only messaging in collaborative perception (Shenkut et al., 25 Sep 2025), and for broader post-quantum, quantum-inspired, or quantum-enhanced V2X designs that address cryptographic migration, runtime orchestration, or network optimization (Zeydan et al., 2021, Sengupta et al., 1 Feb 2026, Chen et al., 13 Jan 2025, Shakib et al., 2023, Awan et al., 26 Mar 2026, Yan et al., 2024).

1. Scope and research uses

The literature associates QuantV2X with three recurring technical themes: quantization of cooperative-perception stacks, quantitative benchmarking of vehicular communication performance, and quantum-era redesign of V2X security or control. This suggests that QuantV2X functions both as a specific system name and as a broader label for V2X research centered on quantization, quantitative analysis, or quantum-safe and quantum-inspired mechanisms.

Use of the term	Representative work	Core technical focus
Fully quantized cooperative perception	"QuantV2X: A Fully Quantized Multi-Agent System for Cooperative Perception" (Zhao et al., 3 Sep 2025)	End-to-end low-bit models and low-bit message representations
Quantitative C‑V2X benchmarking	"Performances of C-V2X Communication on Highway under Varying Channel Propagation Models" (Wang et al., 2019)	Highway sidelink PRR under Two-Ray, WINNER II, and 3GPP models
Communication-efficient V2X feature quantization	"Residual Vector Quantization For Communication-Efficient Multi-Agent Perception" (Shenkut et al., 25 Sep 2025)	Per-pixel code indices for BEV feature exchange
Post-quantum and hybrid-security V2X	(Zeydan et al., 2021, Sengupta et al., 1 Feb 2026, Chen et al., 13 Jan 2025, Shakib et al., 2023)	Runtime PQC selection, secure transitions, hybrid certificates, attack motivation
Quantum-inspired or quantum-enhanced V2X optimization	(Awan et al., 26 Mar 2026, Yan et al., 2024)	Superposition-style optimization or VQC-based MORL

2. Fully quantized cooperative perception system

In the primary named sense, QuantV2X is the first fully quantized intermediate-fusion V2X system that unifies model-side and communication-side quantization end-to-end. It replaces FP32 models with low-bit neural networks via deployment-oriented post-training quantization and replaces FP32 BEV features with compact codebook indices for transmission. The system is organized around intermediate fusion in BEV space, with LiDAR encoders PointPillars and SECOND, a camera encoder Lift-Splat-Shoot with ResNet-50, and Pyramid Fusion as the default fusion module, while also generalizing to F-Cooper, AttFuse, V2X-ViT, and Where2comm (Zhao et al., 3 Sep 2025).

Quantization is applied to the encoder, fusion, and detection head, and the communication channel carries only codebook indices instead of BEV tensors. The weight and activation mapping uses integer uniform quantization with zero-point:

$$q = \operatorname{clip}(\operatorname{round}(x/s)+z, q_{\min}, q_{\max}), \qquad \hat{x} = s \cdot (q-z).$$

The paper uses channel-wise linear quantization for weights, min–max activation ranges, block-wise reconstruction during calibration, and AdaRound for learnable rounding. PTQ only is used; no end-to-end fine-tuning with labels is required, and 0.5% unlabeled calibration data with 5,000 calibration steps is reported as the effective calibration regime. To offset quantization-induced distribution shifts, the fusion stage introduces a heterogeneity-alignment term,

$$L_{\text{hetero}} = D_{\mathrm{KL}}(H_i^{fp} \,\|\, H_i^{int}),$$

and a spatial-alignment term,

$L_{\text{spatial}} = \|B_i^{fp} - B_i^{int}\|^2.$

On the communication side, the message at each BEV location is an index into a learned dictionary rather than an FP32 feature vector. With multiple codes per location,

$\hat{F}[h,w] = \sum_r \alpha_r d_{\text{index}_r},$

and the transmitted size is approximated by

$$\text{sizebits} \approx H \cdot W \cdot n_R \cdot \lceil \log_2(n_L) \rceil.$$

The deployment-oriented latency model is

$$T_{\text{sys}} = T_{\text{local}} + T_{\text{comm}} + T_{\text{fus}}, \qquad T_{\text{comm}} = f_s / v + U(0,200\text{ ms}),$$

with $v = 27$ Mbps.

The reported results are system-level rather than purely model-level. On V2X-Real, QuantV2X reduces system-level latency by $3.2\times$ and improves mAP30 by $+9.5$ over full-precision baselines under deployment-aware evaluation. The system-level mAP30/50 values are 52.6/42.2 for QuantV2X at approximately 0.03 MB, compared with 43.1/34.8 for FP32 with no compression at 8.6 MB, 48.8/38.0 for FP32 with $16\times$ compression at 0.54 MB, 49.7/39.0 for Where2comm at 0.30 MB, and 51.4/40.8 for CodeFilling at approximately 0.03 MB. On an NVIDIA RTX 3090, Pyramid Fusion latency falls from 59.5 ms in FP32 to 27.1 ms in INT8; power shifts from 330 W at 47.6 QPS to 300 W at 124 QPS, corresponding to $$L_{\text{hetero}} = D_{\mathrm{KL}}(H_i^{fp} \,\|\, H_i^{int}),$$0 speedup, $$L_{\text{hetero}} = D_{\mathrm{KL}}(H_i^{fp} \,\|\, H_i^{int}),$$1 energy per query, and $$L_{\text{hetero}} = D_{\mathrm{KL}}(H_i^{fp} \,\|\, H_i^{int}),$$2 efficiency. Model-level accuracy retention is also high: for DAIR-V2X with Pyramid Fusion and $$L_{\text{hetero}} = D_{\mathrm{KL}}(H_i^{fp} \,\|\, H_i^{int}),$$3, FP 75.1/68.2 becomes 74.2/66.7 at INT4/INT8, and for $$L_{\text{hetero}} = D_{\mathrm{KL}}(H_i^{fp} \,\|\, H_i^{int}),$$4, FP 80.3/76.1 becomes 80.2/75.5.

The ablation results make the architectural limits explicit. Pyramid Fusion remains robust under low-bit inference, whereas deeper attention stacks are much more sensitive: V2X-ViT drops from 57.4/49.5 at 32/32 to 40.0/11.0 at 8/8 and 29.9/8.8 at 4/8. The paper therefore treats full-stack quantization not merely as model compression but as a deployment strategy that trades numerical precision for improvements in latency, energy, bandwidth, and memory footprint while keeping intermediate-fusion performance close to full precision.

3. Learned message quantization and residual vector quantization

A closely related line of work studies quantization at the feature-message level rather than across the entire stack. ReVQom is a learned feature codec for multi-agent collaborative perception that compresses intermediate BEV feature tensors while preserving per-pixel spatial identity. The paper states that it does not reference a specific method named QuantV2X, but that ReVQom operationalizes quantization for V2X collaborative perception as a QuantV2X module in practice. Its basic strategy is a $$L_{\text{hetero}} = D_{\mathrm{KL}}(H_i^{fp} \,\|\, H_i^{int}),$$5 bottleneck with GroupNorm followed by multi-stage residual vector quantization, with transmission restricted to per-pixel code indices rather than floating-point features (Shenkut et al., 25 Sep 2025).

The communicated object is a BEV tensor $$L_{\text{hetero}} = D_{\mathrm{KL}}(H_i^{fp} \,\|\, H_i^{int}),$$6, typically with $$L_{\text{hetero}} = D_{\mathrm{KL}}(H_i^{fp} \,\|\, H_i^{int}),$$7 and $$L_{\text{hetero}} = D_{\mathrm{KL}}(H_i^{fp} \,\|\, H_i^{int}),$$8. Raw features at 32-bit precision imply 8192 bits per pixel, since $$L_{\text{hetero}} = D_{\mathrm{KL}}(H_i^{fp} \,\|\, H_i^{int}),$$9. ReVQom compresses channels to $L_{\text{spatial}} = \|B_i^{fp} - B_i^{int}\|^2.$0, keeps the BEV grid unchanged, and performs $L_{\text{spatial}} = \|B_i^{fp} - B_i^{int}\|^2.$1 stages of residual vector quantization. For each pixel and stage, the nearest codebook element is chosen under $L_{\text{spatial}} = \|B_i^{fp} - B_i^{int}\|^2.$2 distance, the residual is updated, and only the integer index is transmitted. The bits-per-pixel relation is

$L_{\text{spatial}} = \|B_i^{fp} - B_i^{int}\|^2.$3

With fixed $L_{\text{spatial}} = \|B_i^{fp} - B_i^{int}\|^2.$4 and $L_{\text{spatial}} = \|B_i^{fp} - B_i^{int}\|^2.$5, the paper defines 6 bpp for $L_{\text{spatial}} = \|B_i^{fp} - B_i^{int}\|^2.$6, 12 bpp for $L_{\text{spatial}} = \|B_i^{fp} - B_i^{int}\|^2.$7, 18 bpp for $L_{\text{spatial}} = \|B_i^{fp} - B_i^{int}\|^2.$8, 24 bpp for $L_{\text{spatial}} = \|B_i^{fp} - B_i^{int}\|^2.$9, and 30 bpp for $\hat{F}[h,w] = \sum_r \alpha_r d_{\text{index}_r},$0.

The bandwidth reduction is large in absolute terms. At $\hat{F}[h,w] = \sum_r \alpha_r d_{\text{index}_r},$1, 6 bpp corresponds to 98,304 bits per frame, approximately 12 KB, or approximately 0.983 Mbps at 10 Hz. By contrast, the raw-feature baseline is 134,217,728 bits per frame, approximately 16 MB, or approximately 1.342 Gbps at 10 Hz. The reported compression spans $\hat{F}[h,w] = \sum_r \alpha_r d_{\text{index}_r},$2 at 30 bpp to $\hat{F}[h,w] = \sum_r \alpha_r d_{\text{index}_r},$3 at 6 bpp.

On DAIR-V2X with a CoBEVT backbone, the results define a concrete accuracy-versus-bitrate frontier. ReVQom-S at 18 bpp achieves [email protected]/[email protected] of 0.747/0.651, compared with 0.728/0.657 for raw-feature CoBEVT at 8192 bpp. ReVQom-M at 24 bpp reaches the highest [email protected], 0.753, at $\hat{F}[h,w] = \sum_r \alpha_r d_{\text{index}_r},$4 compression. Even the ultra-low-bandwidth regimes remain competitive: ReVQom-μ at 6 bpp yields 0.690/0.558, and ReVQom-T at 12 bpp yields 0.699/0.609. The ablations identify $\hat{F}[h,w] = \sum_r \alpha_r d_{\text{index}_r},$5 as the preferred EMA decay rate over 0.99, $\hat{F}[h,w] = \sum_r \alpha_r d_{\text{index}_r},$6 as the optimal number of stages, $\hat{F}[h,w] = \sum_r \alpha_r d_{\text{index}_r},$7 as the preferred channel reduction ratio, and $\hat{F}[h,w] = \sum_r \alpha_r d_{\text{index}_r},$8 as the best compression-performance trade-off, reaching 99.2% of the $\hat{F}[h,w] = \sum_r \alpha_r d_{\text{index}_r},$9 [email protected] with 75% of its bandwidth cost.

Relative to the full-stack QuantV2X system, ReVQom isolates a single but critical subsystem: communication-efficient exchange of intermediate features. This suggests a modular interpretation of QuantV2X in which message quantization alone can deliver large bandwidth savings while leaving the backbone architecture largely unchanged.

4. Quantitative C‑V2X highway characterization

A distinct use of QuantV2X concerns quantitative characterization of C‑V2X highway sidelink performance under varying propagation models. In this setting, the topic is not neural-network quantization but the measurement of PRR, BLER, coverage, and bandwidth effects for PC5 sidelink communication in a controlled highway scenario. The evaluated system uses direct device-to-device communication over PC5, operates exclusively in 3GPP sidelink Mode 3, and studies a single-direction highway with 3 lanes and 20 km length, BS inter-site distance $$\text{sizebits} \approx H \cdot W \cdot n_R \cdot \lceil \log_2(n_L) \rceil.$$0 km, vehicle antenna height $$\text{sizebits} \approx H \cdot W \cdot n_R \cdot \lceil \log_2(n_L) \rceil.$$1 m, BS antenna height $$\text{sizebits} \approx H \cdot W \cdot n_R \cdot \lceil \log_2(n_L) \rceil.$$2 m, isotropic antennas, $$\text{sizebits} \approx H \cdot W \cdot n_R \cdot \lceil \log_2(n_L) \rceil.$$3 receiver diversity, center frequency $$\text{sizebits} \approx H \cdot W \cdot n_R \cdot \lceil \log_2(n_L) \rceil.$$4 GHz, and bandwidths of 5, 6, 8, and 10 MHz (Wang et al., 2019).

The message model is fixed to periodic cooperative awareness messages of 212 bytes at 10 Hz, yielding 2120 B/s, approximately 16.96 kbps per UE. The highway cooperative-driving target is up to 1000 m. The paper defines the MCS-selection condition

$$\text{sizebits} \approx H \cdot W \cdot n_R \cdot \lceil \log_2(n_L) \rceil.$$5

with lower spectral efficiency corresponding to more robust MCS selection. Reception success is defined by BLER $$\text{sizebits} \approx H \cdot W \cdot n_R \cdot \lceil \log_2(n_L) \rceil.$$6, and the latency decomposition is

$$\text{sizebits} \approx H \cdot W \cdot n_R \cdot \lceil \log_2(n_L) \rceil.$$7

Only large-scale fading is modeled: path loss and log-normal shadowing. Small-scale multipath fading, Doppler, and detailed temporal channel evolution are not included.

Three propagation models are compared: Two-Ray Interference, WINNER II D1 rural macro with LOS/NLOS and shadowing, and a 3GPP Release 15 LOS/NLOS model for C‑V2X sidelink. The quantitative results are driven by PRR. At inter-vehicle distance 10 m, PRR at 5 MHz is 68.32% for Two-Ray, 39.83% for WINNER II, and 73.60% for 3GPP; at 10 MHz it becomes 90.10%, 54.85%, and 100.00%, respectively. At inter-vehicle distance 15 m, PRR at 5 MHz is 86.14%, 55.00%, and 90.35%; at 10 MHz it reaches 100.00%, 65.65%, and 100.00%. The trends are consistent across the study: larger bandwidth improves PRR by allowing lower-SE, more robust MCS choices; lighter traffic density improves PRR by reducing system load; the 3GPP model is the most optimistic; WINNER II D1 is the most conservative; and Two-Ray lies between them.

The study also reports that shadowing has negligible impact for WINNER II at distances below 1000 m but becomes more significant beyond approximately 6500 m, whereas the 3GPP shadowing impact is less pronounced. Because the paper focuses on the $$\text{sizebits} \approx H \cdot W \cdot n_R \cdot \lceil \log_2(n_L) \rceil.$$8 m range, the practical outcome is a bounded highway benchmark for cooperative-driving sidelink design under large-scale propagation assumptions. In that benchmark, Mode-3 PC5 sidelink with at least 8–10 MHz and robust MCS selection is the recommended operating region for high PRR, while the absence of small-scale fading and HARQ means the results are optimistic with respect to fully realistic highway channels.

5. Post-quantum security, hybrid certificates, and cryptographic orchestration

A security-oriented interpretation of QuantV2X treats V2X as a post-quantum system whose cryptographic overhead must be managed jointly with latency and reliability constraints. One strand uses a Service & Computation Orchestrator to select polynomial-multiplication methods for NTRU based on real-time load. In the reported experiments, Kubernetes v1.12 acts as the orchestrator, the hardware has 48 CPU cores and 64 GB RAM with tests restricted to five cores, and Karatsuba and Toom-Cook are the candidate methods. For degree 512, Karatsuba single-core requires 21.2 ms and Toom-Cook single-core 27.1 ms; Toom-Cook with five cores outperforms Karatsuba up to 25% load, but beyond 25% degrades relative to Karatsuba and beyond 45% becomes worse than Toom-Cook single-core. For higher degree, reported as 821 in text and 812 in the figure label, Karatsuba single-core is 50.2 ms, Toom-Cook single-core is 120 ms, and Toom-Cook five-core loses to Karatsuba after only 5% load. The proposed rule is therefore load-dependent method selection rather than a fixed QRSA implementation (Zeydan et al., 2021).

A more adaptive framework extends this logic from arithmetic kernels to complete PQC profiles. The context-aware 6G-ready vehicular design uses short-term prediction over 100–200 ms with updates every 20–50 ms, APMOEA for multi-objective selection, RL stabilization to suppress oscillatory switching, and a secure monotonic-upgrade protocol with version counters, context hashes, fresh nonces, and two-phase acknowledgment. The experiments report latency reduction up to 27% and communication-overhead reduction up to 65% relative to static PQC baselines. Static end-to-end latencies are 9.3 ms for Kyber-768, 10.8 ms for Dilithium-3, 8.7 ms for Classic McEliece-348864, and 17.4 ms for SPHINCS+-128s; adaptive PQC without RL yields 7.8 ms and 14.2 switches per 60 s, while APMOEA+RL yields 7.2 ms and 4.3 switches per 60 s. In the evaluated adversarial scenarios, the monotonic-upgrade protocol prevents downgrade, replay, and desynchronization attacks, and all downgrade attempts are rejected (Sengupta et al., 1 Feb 2026).

A certificate-centered line of work addresses the incompatibility between current PQC signatures and V2X packet-length constraints. The proposed SCMS architecture keeps pure PQC for infrastructure certificates and EE enrollment certificates, but uses a hybrid design for EE authorization or pseudonym certificates: the Verify Key Indicator carries a compressed ECC P-256 public key, the certificate issuer signature is PQC, and the on-air SPDU signature remains ECDSA P-256. Among the evaluated combinations, Falcon-512 plus ECDSA P-256 is the only hybrid profile that fits comfortably below the 1400-byte WSM MTU: certificate size $$\text{sizebits} \approx H \cdot W \cdot n_R \cdot \lceil \log_2(n_L) \rceil.$$9 bytes and SPDU size $$T_{\text{sys}} = T_{\text{local}} + T_{\text{comm}} + T_{\text{fus}}, \qquad T_{\text{comm}} = f_s / v + U(0,200\text{ ms}),$$0 bytes. By contrast, Dilithium-2 plus ECDSA P-256 produces $$T_{\text{sys}} = T_{\text{local}} + T_{\text{comm}} + T_{\text{fus}}, \qquad T_{\text{comm}} = f_s / v + U(0,200\text{ ms}),$$1 bytes, and SPHINCS+ SHA2-128f plus ECDSA P-256 produces $$T_{\text{sys}} = T_{\text{local}} + T_{\text{comm}} + T_{\text{fus}}, \qquad T_{\text{comm}} = f_s / v + U(0,200\text{ ms}),$$2 bytes. On the Clientron OBU, measured verification times are 26.51 ms for ECDSA P-256 and 0.87 ms for Falcon-512, giving an aggregate verification time of approximately 27.38 ms per message, while Falcon-512 signing is 46.88 ms and still meets the 100 ms BSM period (Chen et al., 13 Jan 2025).

The urgency of quantum-safe migration is underscored by a proof-of-concept quantum cyber-attack on blockchain-based VANET. In that study, RSA signatures protect blockchain-backed trust transactions; an attacker factors a small RSA modulus $$T_{\text{sys}} = T_{\text{local}} + T_{\text{comm}} + T_{\text{fus}}, \qquad T_{\text{comm}} = f_s / v + U(0,200\text{ ms}),$$3 with Shor’s algorithm implemented in Qiskit, recovers the private key, and forges crash alerts under the identity of a trusted vehicle. The reported factorization time per run is 5–17 s with mean approximately 10.4 s, and the end-to-end impersonation time is approximately 115 s. The attack therefore fits within the 10-minute PoW-style block window referenced by the paper and leads to trust-chain failure, miner selection distortion, and eventual blocking of the impersonated vehicle (Shakib et al., 2023).

Taken together, these works define QuantV2X in security terms as a migration problem: cryptography must become quantum-resistant, but the migration must also preserve packet budgets, verification budgets, interoperability, and runtime stability.

6. Quantum-inspired and quantum-enhanced V2X optimization

A further usage of QuantV2X appears in work that imports quantum-inspired optimization or quantum-enhanced learning into vehicular networking and control. QIVNOM is a quantum-inspired framework that jointly optimizes V2V and V2I communication together with urban traffic control on classical edge–cloud hardware, without requiring a quantum processor. Candidate routing–signal plans are represented as normalized probabilistic superpositions on the unit sphere, updated by sphere-projected gradients, coupled by an entanglement-style regularizer, and selected through annealed sampling under Tchebycheff multi-objective scalarization with feasibility projection. In a METR-LA–calibrated SUMO–OMNeT++/Veins simulation over a $$T_{\text{sys}} = T_{\text{local}} + T_{\text{comm}} + T_{\text{fus}}, \qquad T_{\text{comm}} = f_s / v + U(0,200\text{ ms}),$$4 km urban map with IEEE 802.11p and 5G NR sidelink, QIVNOM reaches mean end-to-end latency 57.3 ms, approximately 20% lower than the best baseline, incident latency 62 ms versus 79 ms, RSU-outage latency 67 ms versus 86 ms, packet delivery 96.7%, reliability 96.7% overall and 96.8% under RSU outages versus 94.1% for the baseline, and corridor-closure travel time 12.8 min versus 14.5 min with congestion 33% versus 37% (Awan et al., 26 Mar 2026).

A more explicitly quantum-enhanced formulation replaces the Q-function in multi-objective reinforcement learning with a variational quantum circuit. The vehicular-network setting is a four-lane highway with autonomous vehicles choosing both driving maneuvers and network-selection actions across RF base stations and THz base stations. The VQC uses 5 qubits, 3 layers, $$T_{\text{sys}} = T_{\text{local}} + T_{\text{comm}} + T_{\text{fus}}, \qquad T_{\text{comm}} = f_s / v + U(0,200\text{ ms}),$$5 gates, and multi-observable readout for 15 joint actions. Compared with a DDQN baseline, the reported gains are 31.32% faster convergence and 18.64% average reward improvement. The paper’s claim is not that current V2X systems require a quantum processor, but that compact VQC-MORL policies can improve sample efficiency and attained policies for handover-aware, safety-aware joint control of mobility and connectivity (Yan et al., 2024).

These works shift QuantV2X away from compression or cryptography and toward decision-making formalisms. This suggests a broader conceptual boundary for the term: V2X systems may be “quant” either because they are quantized, because they are quantitatively benchmarked, because they are quantum-safe, or because they adopt quantum-inspired or quantum-enhanced optimization primitives.