High-frequency wireless links (millimeter-wave, roughly 30–300 GHz) reach high data rates only when the transmitter and receiver point narrow "pencil" beams at each other. Finding the right pair of beams usually means trying hundreds of candidate directions one after another. That search, called beam sweeping, can take tens of milliseconds. On a busy road, a moving truck can block the chosen direction before the search even finishes. Sensor-aided methods cut this cost by predicting a short list of likely beams from cameras, LiDAR, radar or GPS, so the radio only has to test that list. Today's predictors, however, read the scene as a picture: a pixel grid, a voxel grid, or a sequence of sensor tokens. They never ask the question that decides which beam works: which specific vehicle is sitting in this beam's path right now, and is it blocking the beam or reflecting it?
We propose a beam predictor that asks exactly that question. We turn the scene into a small set of detected physical entities: connected cars, passive trucks, buses and pedestrians. Each candidate beam in the codebook then acts as a query that computes its own weights over those entities. The weights are conditioned on physical geometry: how far the entity sits from the beam's pointing direction, and how deeply it cuts into the beam's Fresnel zone (the ellipsoid around a radio path that must stay mostly clear for the signal to arrive at full strength). Each beam gets a score, and only the top few are swept. The expected payoff is fewer sweeps for the same throughput in the dense, blockage-heavy scenes where current predictors lose the most. A useful side effect is that the per-beam weights can be read as "this entity is the blocker" or "this entity is the reflector."
Core Claim
Dynamic neighbor weighting is most useful for beam alignment when it is organized around beams rather than around users. The weight any scene entity receives should depend on its geometric relation to a particular candidate beam. With that structure, a small model can rank a 256-entry codebook well enough to reach high throughput with a shorter sweep list than grid-based LiDAR networks. The gain should concentrate in samples where a moving entity intrudes on the line-of-sight path. The cost stays linear in the number of entities for each beam, so it scales as $O(M n)$ for $M$ beams and $n$ entities, not the cubic cost reported for multi-head graph attention over vehicle links.
Concretely, for beam $b$ with learned embedding $q_b$, and entity $j$ with features $h_j$ (position, extent, velocity, detection confidence, class), we compute a geometric vector $g_{bj}$. This vector holds the angular offset of $j$ from beam $b$'s direction, the normalized Fresnel intrusion of $j$ along the transmitter–receiver path implied by $b$, and the ranges. The per-beam weight is
$$ \alpha_{bj} = \operatorname{softmax}_j\Big(\mathbf{a}^\top \mathrm{LeakyReLU}\big(\mathbf{W}[\,q_b \,\|\, h_j \,\|\, g_{bj}\,]\big)\Big), $$
which is the dynamic attention form of GATv2 with the geometric vector appended to the input. Beam $b$ is then scored from $q_b$ and $\sum_j \alpha_{bj} \mathbf{V} h_j$. Here $\mathbf{W}$, $\mathbf{V}$ and $\mathbf{a}$ are shared learned parameters. Before the softmax, entities whose Fresnel intrusion for beam $b$ is negligible are masked out, so each beam attends to only a handful of relevant entities.
How This Differs From What Exists
Three papers already put graphs next to mmWave problems, and none does this.
GBLinks selects beams and activates links in dense device-to-device networks. It uses a graph network with symmetric MAX/MEAN pooling over an interference graph built from channel state, and it reaches about 90% of exhaustive-search throughput. Its nodes are communicating pairs, its aggregation is isotropic, and its scenes are static. We use entity nodes that may never transmit, and anisotropic, beam-specific weights.
GAT-A2C feeds an 8-head graph attention network into a reinforcement-learning agent for sub-6 GHz power and subchannel allocation. It does not touch directional beams, and its authors report cubic scaling in vehicle count. That scaling is the constraint our per-beam masked design is built to avoid.
The edge-conditioned GATv2 IAB planner supplies the method idea we borrow, namely putting physical link quantities into the attention input. It applies that idea to infrastructure placement on a timescale of months, not to beam selection for moving users.
On the beam-prediction side, attention already appears in several forms, but it is never placed across physical neighbors. Zecchin et al. use non-local attention over pixels of a top-view LiDAR grid. The VLM contrastive model gates across sensor modalities. M2BeamLLM attends across time. These works motivate the proposal. The non-local block improves non-line-of-sight top-1 throughput by 25%, which is evidence that long-range relational reasoning matters, but pixel attention costs grow with grid resolution instead of with the number of objects. Graph Query Networks show that radar returns can be handled as dynamic graphs for detection. We reuse that idea as the front end that produces entities, including passive ones that communication-only graph networks cannot represent.
Research Program
We will build three components.
The first is an entity extractor. On simulated data, entities can come from ground-truth mobility traces (an oracle setting). In the realistic setting, they come from clustering of LiDAR or radar points plus the target's GPS fix. Each entity carries a detection confidence, so sensor degradation such as night, rain or fog shows up as uncertain or missing nodes rather than corrupted pixels.
The second is the geometric encoder. For each codebook beam, it computes the steering direction. It then computes each entity's angular offset and its Fresnel intrusion, using the first-zone radius $r=\sqrt{\lambda d_1 d_2/(d_1+d_2)}$, where $\lambda$ is the wavelength and $d_1, d_2$ are the distances from the entity's projection to each end of the path. We compute these for the direct path and for single-bounce paths off each entity treated as a candidate reflector.
The third is the beam-query attention module above, trained with a ranking loss against ray-traced per-beam received power. At inference, it hands a top-$k$ list to a standard sweep.
We will also study two design choices. The first is one versus two rounds of message passing, where entities first update from each other. This tests whether blocker–blocker interactions matter or whether depth only causes oversmoothing. The second is a learned stopping rule for choosing $k$ from how peaked the scores are.
Baselines
The naive reference is exhaustive sweeping over the full codebook: 256 beam pairs on Raymobtime, and the roughly 20 ms 5G NR sweep cited in the radar distillation paper. It defines 100% throughput before overhead, and it defines the overhead we are trying to remove. We also include a GPS-only position lookup (most frequent best beam per location bin) as the cheap non-relational floor.
The strongest existing method on our main benchmark is the non-local CNN with LiDAR and position of Zecchin et al.. On Raymobtime it reports 79.9% of maximum throughput at top-1, and 95% throughput when sweeping only the top 6 of 256 beam pairs. The lightweight federated LiDAR CNN of Mashhadi et al. is a second strong, very small competitor: 91.17% top-10 accuracy and 90% throughput with a top-3 list. We will retrain both on identical train/test splits with the same LiDAR and GPS inputs. Each gets the same hyperparameter search budget as our model, and we will report parameter counts and latency on the same hardware.
Beating them means one of two outcomes. Either we reach 95% throughput with a list shorter than 6, or we deliver higher throughput at equal $k$ with confidence intervals that do not overlap, measured over the full test set. The gain should be clearly larger on the non-line-of-sight and high-density slices.
On the CARLA co-simulation, the reference is the radar-only distilled student at 87.50% mean percentile rank in 2-lane scenes, against the multimodal teacher's 92.17% (cross-modal distillation). A radar-entity version of our model should beat the student at comparable latency, below about 1 ms.
The key-ingredient ablations are central. We will compare the full model against three variants. The first is the same architecture with $g_{bj}$ removed: plain learned attention between beam queries and entity tokens, essentially a small Transformer decoder. The second is a GBLinks-style mean/max pooling over the same entities. The third is the geometric features alone with no attention (below).
Why not just…?
Why not just use a small Transformer over detected objects, with no physics? Attention between beam tokens and object tokens can in principle learn angular relations from coordinates. This is the most serious threat. Our bet is that explicit Fresnel and angle terms matter most when data are limited and in transfer settings, such as 2-lane to 3-lane scenes or unseen densities, where memorized coordinate patterns break. The no-$g_{bj}$ ablation settles the question directly. If it matches the full model in both in-distribution and transfer tests, the contribution shrinks to "entity tokens beat grids," and we will report it that way.
Why not just add a hand-crafted "is something in the line-of-sight" feature to an MLP or the CNN? This is cheap and plausibly strong for pure blockage. It cannot easily represent which entity opens a reflection path, or how several partial intruders combine. We will include this variant, with per-beam Fresnel intrusion features summed over entities and fed to an MLP. If it closes most of the gap, learned per-beam weighting has not earned its complexity. We expect the gap to remain on the non-line-of-sight slice, where the best beam points at a reflector rather than the target.
Why not just keep the non-local CNN and raise grid resolution? Pixel-level non-local attention grows quadratically with the number of grid cells. Coarse grids blur small or distant vehicles. Our cost grows with the handful of entities near each beam's path. We will sweep grid resolution for the CNN baseline and plot throughput against latency. The claim holds only if our curve dominates.
Why not just use hierarchical or bandit search? These reduce the number of probes without sensors, as in correlated bandits and physics-informed parametric bandits. They still need closed-loop probing rounds, and wide early beams lose gain. The approaches are complementary. We will run a correlated-bandit search both over the full codebook and seeded with our top-$k$ list, and count total probes to 95% throughput.
Proposed Experiments
The main study uses Raymobtime s008 for training and s009 for testing. Metrics are the throughput ratio at $k \in \{1,3,6,10\}$, the smallest $k$ reaching 90% and 95% throughput, and end-to-end latency. All results are stratified by line-of-sight versus non-line-of-sight and by the number of entities intruding on the Fresnel zone.
The first decisive check is where the gain lives. If physics-conditioned weighting is the reason for the improvement, the advantage over the non-local CNN should grow with intruding-entity count and be near zero in empty scenes. A uniform gain would instead point to a generic capacity effect.
The second check is interpretability against ground truth. From the ray tracer's path data, we label which entity blocks the direct path or produces the strongest reflection. We then measure how often the highest-weight entity for the winning beam matches that label.
The third check is robustness and transfer on the CARLA co-simulation, which has up to 40 vehicles, weather variation, and 2- and 3-lane layouts. We train on 2-lane noon scenes and test on 3-lane, night, rain and fog scenes. Entity confidences drop under poor sensing, and we test whether the model degrades gracefully relative to the radar student.
DeepSense 6G Scenarios 32–34 serve as a real-data sanity check. Entities there come from camera detections. Its scenes are mostly single-target, so we expect little relational gain. We will report against M2BeamLLM's 68.9% top-1 accuracy on Scenario 32 and the VLM model's 0.9016 distance-based accuracy score, framed as a check that entity abstraction does not hurt in easy scenes rather than as the main claim. A null result there is expected and informative.
Finally, we will measure latency as entity count rises from 5 to 40, to confirm the linear per-beam scaling holds in practice and stays under 1 ms.
References
- GBLinks: GNN-Based Beam Selection and Link Activation for Ultra-Dense D2D mmWave Networks
- Dynamic Allocation of C-V2X Communication Resources Based on Graph Attention Network and Deep Reinforcement Learning
- Digital Twin-Assisted Resilient Planning for mmWave IAB Networks via Graph Attention Networks
- LIDAR and Position-Aided mmWave Beam Selection With Non-Local CNNs and Curriculum Training
- When Vision-Language Model (VLM) Meets Beam Prediction: A Multimodal Contrastive Learning Framework
- M2BeamLLM: Multimodal Sensing-empowered mmWave Beam Prediction with Large Language Models
- Graph Query Networks for Object Detection with Automotive Radar
- Cross-Modal Knowledge Distillation for Efficient Radar-Only Beam Prediction in mmWave Communications
- Federated mmWave Beam Selection Utilizing LIDAR Data
- Fast mmwave Beam Alignment via Correlated Bandit Learning
- Physics-Informed Parametric Bandits for Beam Alignment in mmWave Communications