Highly directional mmWave/THz links require rapid beam alignment, but exhaustive codebook sweeps impose prohibitive training overhead. This letter proposes a sensing-assisted adaptive probing policy that maps multimodal sensing (radar/LiDAR/camera) to a calibrated prior over beams, predicts per-beam reward with a deep Q-ensemble whose disagreement serves as a practical epistemic-uncertainty proxy, and selects a small probe set using a Prior-Q upper-confidence score. The probing budget is adapted from prior entropy, explicitly coupling sensing confidence to beam-training overhead, where E[Kt] denotes the average number of probed beams per sweep. A margin-based safety rule prevents low-SNR locks: outages are defined by an SNR threshold θ, and the shield uses a robustness margin Δ dB above this boundary. Experiments on DeepSense-6G (train: scenarios 42 and 44; test: 43) with a 21-beam DFT codebook achieve Top-1/Top-3 of 0.81/0.99 with E[Kt] ≈ 2 and zero observed outages at θ = 0 dB with Δ = 3 dB.
1 January 2026 in IEEE Open Journal of the Communications Society