Classical State Preparation for Variational Quantum Algorithms via Reinforcement Learning

A new framework called CRiSP uses reinforcement learning to build better starting points for variational quantum algorithms, a class of hybrid quantum-classical methods that researchers hope will deliver practical quantum advantage [1]. The framework, detailed in a paper posted to the arXiv preprint server on 22 May 2026, addresses a central obstacle for variational quantum algorithms (VQAs): their optimization landscapes are plagued by barren plateaus and numerous local minima, which stall convergence [1]. VQAs operate by outsourcing computationally difficult subroutines to a quantum device while a classical optimizer tunes the circuit parameters, a hybrid approach that has drawn wide interest in quantum machine learning [3]. CRiSP — short for Clifford Reinforcement Learning agent for State Preparation — formulates the selection of initial Clifford gates as a sequential decision-making problem. It employs Neural-Guided Monte Carlo Tree Search driven by a Transformer-based policy trained via self-play to insert learned Clifford gates before fixed parameterized rotations [1]. Because Clifford circuits can be simulated efficiently on classical hardware, the entire initialization is constructed through polynomial-time classical stabilizer simulation without changing the underlying circuit architecture [1]. To handle deep circuits, the authors integrate a curriculum learning strategy that progressively expands the search horizon, allowing the agent to scale to larger problem sizes [1]. On Quantum Approximate Optimization Algorithm (QAOA) benchmarks reaching 22 qubits and 1,370 parameters, CRiSP outperformed state-of-the-art Clifford initialization methods by a mean factor of 3.17× in average energy accuracy, with a maximum improvement of 45.02× [1]. For best-achieved energy accuracy, the mean improvement was 2.44× and the maximum reached 16.01× [1]. The paper also reports that assessments on Variational Quantum Eigensolver (VQE) tasks demonstrated the framework’s robustness and generalizability [1]. Variational quantum algorithms are considered a leading candidate for near-term quantum devices because they require shallower circuits than fully quantum approaches. However, their performance depends heavily on the quality of the initial parameters. Previous Clifford initialization methods relied on heuristic rules that often failed to navigate the large combinatorial search spaces of practical circuits [1]. By recasting the problem as one of sequential decision-making and training a reinforcement learning agent to explore the space of Clifford prefixes, CRiSP provides a systematic alternative that the authors argue scales more reliably [1].