Problem

Preparing ground states in quantum computing and classical quantum simulation presents a significant optimization challenge. Existing protocols, such as Variational Quantum Eigensolver (VQE), Density Matrix Renormalization Group (DMRG), and Auxiliary-Field Quantum Monte Carlo (AFQMC), typically require extensive manual tuning of hyperparameters. This manual process is often inefficient and lacks a systematic, automatic feedback mechanism for guiding improvements. The central problem this research tackles is how to automate and optimize these ground-state preparation protocols using AI coding agents, by allowing candidate solutions to be automatically executed and evaluated, thus enabling more efficient and objective algorithmic enhancements.

Method

The authors introduce an autoresearch framework that leverages AI coding agents, specifically using GPT-5.4, to optimize ground-state preparation protocols. This method operates by allowing the agent to generate code edits or configuration changes, execute these modified protocols on specific benchmarks, and then obtain a scalar score based on physical metrics, such as energy. The framework prioritizes retaining changes that lead to improved scores within predefined computational budgets, ensuring an auditable optimization trace. This approach was applied to three distinct quantum simulation methods:

For Variational Quantum Eigensolver (VQE), the agent was tasked with modifying the ansatz and optimizer parameters for molecular active-space problems. The optimization objective for VQE was to minimize the Rayleigh quotient, representing the variational energy. In the Density Matrix Renormalization Group (DMRG) context, the agent tuned one-dimensional DMRG protocols for spin-chain Hamiltonians, focusing on choices like one-site versus two-site updates, initialization strategies, bond-dimension schedules, and truncation cutoffs to manage entanglement-driven bond growth. Lastly, for Auxiliary-Field Quantum Monte Carlo (AFQMC), the agent optimized trial-states and propagation parameters under a fixed computational budget. The AFQMC objective function was designed to favor candidates exhibiting lower post-equilibration energy and reduced fluctuations at a consistent stochastic cost.

Results

The autoresearch framework demonstrated significant improvements across all three ground-state preparation methods. For Variational Quantum Eigensolver (VQE), the AI agent's 100-iteration campaigns successfully transformed initial weak baselines into sophisticated protocols, often incorporating UCCSD-style warm starts and enhanced local refinements. This optimization reduced the final absolute energy error by 5.5 to 12.9 orders of magnitude, achieving chemical accuracy for molecules such as BH, LiH, BeH2, and H2O.

In the Density Matrix Renormalization Group (DMRG) experiments, the optimization campaigns evolved from basic random-start protocols to more advanced DMRG2 or staged DMRG1→DMRG2 schedules. This yielded lower final energies and a noticeable improvement in the correlation patterns across all four tested critical spin chains at length L=64. The results were visually confirmed through improved mutual-information error patterns. For Auxiliary-Field Quantum Monte Carlo (AFQMC), the autoresearch consistently optimized parameters like walker population, imaginary-time step, and block geometry across H2, LiH, H2O, and N2. This led to lower live scores and significantly reduced post-equilibration energies while adhering to the same fixed computational budgets, indicating enhanced statistical stability and accuracy.

Collectively, these findings confirm that the AI agent could effectively mutate simple baseline protocols into more complex and efficient ones, consistently achieving improved energy proxies within specified computational constraints across diverse quantum simulation settings.

Implications

This research demonstrates that ground-state preparation protocols can be successfully reframed as an automated search problem, effectively tackled by advanced AI coding agents. The methodology presented is broadly applicable and has the potential to be extended to various other quantum simulators, compilers, and hardware environments, provided that a reliable scalar objective can be derived from execution. A significant future implication lies in optimizing TT-gates within fault-tolerant quantum algorithms, which could drastically reduce the computational resources needed to achieve quantum advantage.

The authors suggest that future work should focus on scaling this approach to prepare larger many-body quantum states, possibly by using performance extrapolation from smaller systems as an optimization proxy. Furthermore, the framework could be enhanced by allowing the mutation policy's underlying model parameters to be dynamically tuned against the autoresearch score itself, drawing inspiration from reinforcement learning with verifiable rewards and self-improving agent concepts. Such advancements are anticipated to lead to the discovery of ground-state preparation protocols that surpass current state-of-the-art capabilities.