Autoregressive generators often rely on fixed conditioning representations, so uncertainty in noisy inputs can be amplified during rollout under distribution shift. We address this by resolving uncertainty in latent space before decoding: Latent-Level Uncertainty Resolution (LURS) encodes inputs as stochastic latent sequences and learns a denoising refinement operator. At inference, a single deterministic refinement step, without reference tokens or iterative sampling, feeds a refined latent memory to the decoder via cross-attention. In teacher-forced and teacher-forcing-free decoding on open-vocabulary EEG-to-text, LURS improves generation quality and substantially reduces cross-subject variability, while remaining competitive in shared-subject settings. Ablations and stability analysis attribute gains to latent refinement rather than decoder changes or hyperparameter effects.