Semantic-Driven Large Model Scheduling for Distributed Systems via Unified Representation and Policy Generation

Abstract

This study develops a unified semantic representation and policy generation framework for large model-driven intelligent scheduling in distributed systems. The goal is to address the limitations of traditional scheduling methods, which often suffer from increased latency, reduced throughput, and unstable policies under highly dynamic, heterogeneous, and large-scale workloads. The method first constructs a high-dimensional semantic fusion representation through system state encoding, task semantic encoding, and topology modeling. This allows the scheduler to capture both local resource contention and global structural dependencies. A multi-step semantic prediction module is then introduced to model future system states and support proactive policy planning in dynamic environments. During policy generation, a differentiable cost estimator provides fine-grained evaluation of action quality under different resource configurations. A self-correction module is further applied to maintain consistency of scheduling behavior in semantic space and ensure long-term stability. Experiments across multiple dimensions, including scheduling step size, policy network depth, peak load intensity, and node heterogeneity, show consistent advantages in key metrics such as Makespan, Scheduling Latency, Throughput, and Scheduling Stability. The results demonstrate that semantic-driven and prediction-driven mechanisms work effectively together in distributed scheduling. They highlight the potential of large models in multi-constraint and multi-state scheduling tasks and provide a feasible direction for building intelligent scheduling architectures for future computing infrastructures.