Sieving: Denoising-Robust Fine-Tuning for Semantic Structural Representation in Natural Language Inference

Natural Language Inference (NLI) models frequently suffer from "shortcut learning," over-relying on superficial lexical overlap rather than capturing deep semantic entailment. Drawing inspiration from diffusion-based language modelling and denoising autoencoders, we introduce Sieving, a dynamic token-level corruption strategy applied during fine-tuning. By stochastically injecting noise into the input sequence during training—specifically through a calibrated mixture of masking and random token replacement—Sieving forces the model to construct robust, globally aware semantic representations. Our method effectively filters (or "sieves") out surface-level heuristics, leading to superior generalisation on adversarial benchmarks (e.g., ANLI, PAWS) and noisy real-world text regimes. Within the Director Class AI architecture, Sieving serves as a critical stabilisation layer for the Coherence Engine.