UniMind: A Middleware Framework for Bridging Classical AI Workloads and Quantum Computing Backends

Abstract

As quantum computing transitions from theoretical research to practical deployment, a compatibility gap emerges between classical AI software stacks and probabilistic quantum hardware. Existing quantum machine learning frameworks operate at the application level and lack system-level abstractions for hardware adaptation, workload routing, and dynamic code generation. This paper presents UniMind, a middleware framework designed to reduce the abstraction gap between classical AI workloads and heterogeneous quantum-classical computing backends. UniMind introduces two components: (1) an LLM-driven orchestration layer that interprets high-level task specifications and generates topology-aware execution plans through constrained sampling, operating exclusively in user space with sandboxed validation and rule-based fallback; and (2) Unibit, a classical preprocessing representation based on sliding-window frequency estimation and sinc-based smoothing that produces parameters for standard quantum angle encoding. We implement a multi-backend mapping engine and validate the framework with reproducible experiments.