<scp>DSAC</scp> ‐Gen: Dynamic Semantic Alignment–Constrained Consistency Learning for Reliable Sequence Generation in  <scp>IIoT</scp> Systems

ABSTRACT Consistency learning improves the robustness and generalization of sequence generation models, yet existing methods typically constrain only outputs or static latent representations, neglecting the token‐wise and temporally evolving semantic states during generation and leading to progressive semantic deviation that accumulates across decoding steps. In this work, we explicitly define dynamic semantic drift as the gradual divergence between the intended global semantics of a generated sequence and its intermediate semantic representations along the decoding trajectory, which is not fully captured by classical exposure bias or local error propagation. We propose DSAC‐Gen, a model‐agnostic framework that models generation as a continuous semantic trajectory in representation space and imposes multi‐scale semantic alignment constraints on intermediate representations at token and segment levels, thereby explicitly constraining semantic evolution rather than only output consistency, mitigating semantic deviation and improving representation stability. DSAC‐Gen integrates with existing architectures without altering decoding strategies. Experiments on representative sequence generation tasks demonstrate consistent improvements in generation coherence, linguistic consistency , and robustness with minimal computational overhead. These properties make the framework suitable for reliability‐critical natural language generation in networked intelligent systems, including industrial automation platforms, IIoT data reporting pipelines, and emerging Industry 5.0 intelligent manufacturing environments, particularly in English‐dominated technical documentation and industrial reporting scenarios where semantic accuracy is critical.