In-the-wild facial expression recognition (FER) is inherently affected by ambiguous expressions and subjective annotations, making the observed labels only partially reliable. Conventional one-hot supervision nevertheless treats these labels as ground truth and may drive deep networks to memorize corrupted annotations. This paper proposes RDAS, a reliability-driven adaptive supervision framework for noisy-label FER. The central idea was to use label reliability not merely as a sample weight, but as a variable that determines the supervision policy of each training sample. RDAS enhances expression-sensitive representations by combining RGB appearance with structural high-frequency cues, estimates label reliability with an exponential moving average (EMA) teacher through the margin between the annotated class and its strongest competing class, and assigns reliability-specific objectives to high-, medium-, and low-reliability samples. Reliable samples retained hard-label learning, ambiguous samples received teacher-guided soft correction, and unreliable samples were prevented from dominating training with misleading gradients. Experiments on RAF-DB, AffectNet, and FERPlus datasets showed that RDAS improved both standard recognition performance and robustness under synthetic label noise. On RAF-DB, RDAS achieved an average accuracy of $89.70\mathrm{\%}$ on the original dataset and $86.15\mathrm{\%}$ under 30% label noise. Ablation studies provided quantitative support for reliability modeling and adaptive supervision, while the visual analyses provided complementary qualitative evidence consistent with the proposed mechanism.