A low-outage anchor
An arXiv preprint reports a low measured outage rate for a system that predicts future radio beams and adjusts how widely it covers adjacent beam codes. In its full-hybrid anchor, the empirical outage probability was 0.0060 at a threshold of 40% of the oracle power, meaning the power of the best reference beam used for comparison. The normalized covered-power gain ratio was 0.895 and the beam-center switching rate was 0.134. The predictor took 76.37 milliseconds per window and the planner 1.49 milliseconds.
The prediction figures tell a different part of the story. The correct beam appeared first in 39.3% of anchor cases, in the top three in 77.8%, and in the top five in 89.7%. These figures came from a single reported checkpoint, with no uncertainty interval shown. Top-1, Top-3 and Top-5 describe whether the correct beam was ranked first, within the first three or within the first five.
What the system is trying to control
BeamGuard is a risk-aware multimodal framework for mmWave vehicle-to-infrastructure links. It combines information from several sensors with partial beam-power observations, forecasts future beams and uses a planner to select beam centers and adjacent-codebook virtual widths. A measurement-validity mask marks what was actually observed, so an unseen power entry is not treated as a measured low-power beam.
The standardized data set contained 6,600 synchronized rows. For the joint day-night protocol, 4,000 rows were used for training, 800 for validation, 800 for calibration and 1,000 for testing. After the study converted the sequences into history-and-horizon windows, those counts became 3,760, 752, 752 and 940.
Among the operating regimes tested, the full configuration was described as the most reliability-oriented point and had the lowest threshold-based empirical outage. It did not consistently lead the Top-K accuracy rankings, underscoring the difference between predicting a likely beam and choosing a control policy for the link.
The tradeoff behind the headline number
With identical forecast outputs, a fixed width of 5 reported an outage probability of 0.0043, a gain ratio of 0.9895 and a switching rate of 0.1234. A fixed width of 1 reported 0.0070, 0.9631 and 0.2346 on the same measures. The adaptive set {1, 3, 5} reported 0.0060 outage, 0.8948 gain and 0.1338 switching.
Risk-aware logic did not lead every metric in the same comparison. With the same forecaster outputs, the risk-aware and no-risk greedy controllers both reported an outage of 0.0060. Risk-aware control had lower switching, 0.1332 versus 0.1545, but also a lower gain ratio, 0.9056 versus 0.9420. The comparison therefore paired the lower switching rate with a lower gain ratio while leaving the reported outage unchanged.
In the G+Pwr comparison, the row with 8 observed beam entries had a Brier score of 0.7465, a DBA@3 value of 0.9046 and a power mean absolute error of 0.1163. With 64 observed entries, the corresponding values were 0.7320, 0.9149 and 0.1139. The higher-budget row thus reported the lower Brier score and power error and the higher DBA@3 score.
In a matched comparison using G+Pwr inputs and 64 observed beam entries, BeamGuard's Top-1, Top-3 and Top-5 values were 0.405, 0.791 and 0.910. A G+Pwr temporal CNN reported 0.373, 0.752 and 0.880. BeamGuard also had the lower negative log-likelihood, 1.784 versus 1.894, and lower expected calibration error, 0.046 versus 0.050. This was a predictor comparison under the same input condition, not a claim that one controller led every operating metric.
Transfer is the harder test
Transfer between day and night scenes was much less settled. In zero-shot S32 to S33 transfer, full-hybrid Top-1, Top-3 and Top-5 accuracy values were 0.1403, 0.3184 and 0.5096. With 20% target-domain supervision, the corresponding values were 0.2779, 0.5890 and 0.7376. In the reverse S33 to S32 direction, the values were 0.1933, 0.4252 and 0.5436 in zero-shot transfer, compared with 0.3108, 0.6654 and 0.8145 with the same target-domain supervision.
The held-out scenario test was harsher. On S31, G+Pwr had the lowest listed outage, 0.0177, with a gain ratio of 0.7740. On S34, full hybrid had the strongest listed ranked accuracy, but those values were only 0.0593 for Top-1, 0.1477 for Top-3 and 0.2166 for Top-5. Its outage was 0.2236 and its gain ratio was 0.6384. The results limit any claim that the main-scenario performance transfers universally across scenarios.
A benchmark, not a deployment claim
The runtime result also needs careful reading. On an Apple M2 Pro with 16 GB of unified memory, using PyTorch MPS and an inference batch size of one, the hybrid anchor had 39.56 million parameters and a 151.17 MB checkpoint. End-to-end latency was 77.86 milliseconds per window, including 76.37 milliseconds for prediction and 1.49 milliseconds for planning, with throughput of 12.84 windows per second. These are software measurements on the specified platform, not a hardware-independent deployment guarantee.
The work is a preprint listed as arXiv:2608.25433v1 and dated 26 Aug 2026. Its central contribution is an overhead-aware forecasting-and-control framework that reports reliability, gain, switching and observation-budget results together. The evidence supports software-benchmark comparisons, while the sharp held-out degradation leaves broader transfer as an open question.
Paper data and sources
Original title: BeamGuard: Risk-Aware Multimodal Beam Forecasting and Adaptive Virtual Beamwidth Control for 6G mmWave V2I Links
Authors: Abidemi Orimogunje, Dejan Vukobratovic, Sunwoo Kim et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text