Reliability-Aware Soft-Decision Beam Training for RIS-Assisted mmWave Systems

Millimeter-wave (mmWave) communications offer large bandwidth for high data rates, yet suffer from severe path loss and blockage, which makes highly directional beam alignment essential and training overhead a critical bottleneck. Reconfigurable intelligent surfaces (RISs), implemented as nearly passive programmable metasurfaces, can reshape the radio environment to create strong reflected paths and enhance mmWave coverage via controllable beamforming, but practical operation requires fast and reliable beam training over a large cascaded search space. To that end, coded beam training (CBT) was proposed to reduce training overhead by mapping hierarchical beam search to an error-control coding structure; however, existing RIS-CBT approaches mainly rely on hard decisions and fixed pilot allocation, which limits robustness under stringent overhead budgets and varying received signal-to-noise ratio (SNR) conditions. To deal with this challenge, this paper proposes a reliability-aware RIS beam training method that combines soft-information decoding with confidence-driven pilot reallocation. Specifically, we derive stage-wise soft metrics from the <tex-math notation="LaTeX">2× 2 </tex-math> pilot measurements and apply a soft maximum a posteriori (MAP) decoder to exploit measurement reliability rather than hard slicing. When decoding confidence is low, the remaining pilot budget is adaptively reallocated to repeat the most uncertain stages, improving overall decision reliability without increasing the nominal codebook depth. Simulation results for off-grid, RIS-assisted mmWave channels demonstrate that the proposed method achieves a higher success probability and improved achievable rate compared with binary hierarchical training and hard-decision CBT baselines across a wide range of SNRs and training overheads, and it approaches the oracle upper bound corresponding to perfect channel state information (CSI) with substantially reduced training overhead, while maintaining practical computational complexity.