The computational burden associated with transformer architectures like BERT presents obstacles for deployment in resource-constrained environments. Contemporary compression methodologies primarily employ uniform compression strategies across diverse input instances, neglecting the inherent variability in computational requirements among different examples. In this paper, we present a distinct example-aware adaptive layer pruning framework that dynamically orchestrates transformer layer selection contingent upon input complexity characteristics. Our methodology incorporates a compact policy network architecture that generates binary activation masks for individual layers, allowing for personalized computational resource allocation per input instance. Through the implementation of differentiable Gumbel-Sigmoid relaxation mechanisms, we enable end-to-end optimization protocols while preserving classification accuracy. Comprehensive empirical evaluation on the Stanford Sentiment Treebank (SST-2) corpus demonstrates our approach achieving 93.0% classification accuracy while utilizing just 3.81 layers on average from the complete 12-layer architecture, yielding a substantial 3.15 × compression ratio. This methodology outperforms established compression techniques including DistilBERT, TinyBERT, and conventional static pruning approaches, establishing a superior equilibrium between accuracy preservation and computational efficiency in BERT compression paradigms.