Researchers are applying a new approach to neural quantum state (NQS) optimization, leveraging the 1.5 billion parameter RWKV-7 model to extend quantum many-body wavefunction approximations beyond previous limits. The team, which spans the Mira-Québec AI Institute, Applied Quantum Algorithms and Delft University of Technology, is tackling challenges with existing optimization methods such as stochastic reconstruction, which are described as costly and numerically fragile for large-scale models. To overcome these hurdles, they frame variational energy minimization as a favorable policy gradient problem and draw connections between reinforcement learning and quantum physics. This theoretical approach led to the development of close-in wavefunction optimization (PWO), a trust-region algorithm that avoids explicit matrix inversion and reuses samples, improving stability and convergence across complex spin systems.
A 1.5 billion parameter language model typically used for text generation is being repurposed to tackle difficult problems in quantum physics. Researchers have demonstrated the application of the RWKV-7 architecture to optimize neural quantum states (NQS), potentially changing the way quantum systems are modeled and understood. Existing methods such as stochastic reconstruction become computationally prohibitive as the size of the system increases. To circumvent this, they developed a new trust-region algorithm, proximal wavefunction optimization (PWO), that avoids explicit matrix inversion and reuses samples across multiple updates. This provides significant advantages over techniques hampered by autocorrelation and slow mixing. The researchers say that PWO avoids explicit matrix inversion, reuses samples across multiple updates, and combines the scalability and theoretical guarantees of first-order optimization. Across simulations of Ising and flat-last spin systems, compared to established optimization algorithms, PWO clearly improved both stability and convergence speed, allowing exploration of more complex quantum systems with higher accuracy and efficiency.
Current approaches to approximating quantum many-body wavefunctions using neural quantum states (NQS) increasingly rely on autoregressive models. Autoregressive models are prized for their ability to perform accurate and independent sampling from the Born distribution and avoid the limitations of Markov chain Monte Carlo methods. The researchers demonstrated the relationship between variational energy minimization and policy gradient reinforcement learning and revealed that NQS gradients can be expressed as an advantage-weighted form over the Born distribution. They then fine-tuned the 1.5 billion-parameter RWKV-7 model and demonstrated NQS optimization at a scale three orders of magnitude greater than previous work, highlighting the growing synergy between machine learning and quantum physics and leveraging techniques from different fields to address fundamental challenges in computational quantum mechanics.
Researchers have demonstrated NQS optimization at a scale of more than three orders of magnitude, exceeding previous studies. The team, based at institutions such as Mira Québec AI Institute and Leiden University, identified gaps in NQS optimization principles. Although autoregressive NQS offers advantages in sampling efficiency, its optimization remains a major hurdle. Existing approaches present a trade-off between speed and accuracy. Although first-order optimizers like Adam are scalable, they ignore the geometry of the variational wavefunction, which can lead to unstable or inaccurate convergence in NQS applications, according to the researchers. This allows PWO to clip probability ratio changes, effectively control step size, and prevent rapid updates that can destabilize training. The 1.5 billion parameter RWKV-7 model, a large language model reused for NQS optimization, was fine-tuned with a one-dimensional Ising model. Comparative tests with Adam, minSR, and SPRING on standard benchmarks including 1-D and 2-D Ising and J1-J2 spin systems revealed that PWO consistently improves both stability and convergence speed, suggesting a promising path toward more efficient and reliable NQS training.
Variational energy minimization as policy gradient RL
The pursuit of simulating complex quantum systems is taking an unexpected turn as researchers leverage the architecture of large-scale language models to optimize neural quantum states (NQS). This approach is not just about applying more computing power. This is a fundamental reformulation of the optimization problem itself, and shows similarities between quantum mechanics and reinforcement learning. By fine-tuning the 1.5 billion-parameter RWKV-7 model, researchers demonstrated NQS optimization at a scale that exceeds previous work by three orders of magnitude. This progress stems from new theoretical insights. The researchers demonstrate that the core process of NQS, variational energy minimization, can be understood mathematically as a favorable policy gradient problem for the Born distribution. This connection allows trust region optimization techniques commonly used in reinforcement learning to be applied to NQS training, resulting in Proximity Wavefunction Optimization (PWO), a new algorithm designed to address the limitations of existing techniques.
This suggests a potential path to tackling increasingly complex quantum systems and unlocking new insights in materials science, drug discovery, and fundamental physics. The team’s research highlights a growing trend of interdisciplinary innovations that combine seemingly disparate technologies to push the limits of scientific computing.
Scaling of neural quantum state optimization is now possible using the 1.5 billion parameter RWKV-7 model. This was not possible before. Researchers at institutions such as Mira-Québec AI Institute, Université de Montréal, Leiden University, Delft University of Technology, and CIFAR have demonstrated a significant increase in our ability to model complex quantum systems. This work culminated in the successful fine-tuning of the 1.5 billion parameter RWKV-7 model, demonstrating NQS optimization at a scale that exceeds previous work by three orders of magnitude.
The pursuit of scalable methods to approximate quantum many-body wavefunctions has increasingly focused on neural quantum states (NQS), but significant challenges exist in the optimization of these models. We tested PWO on both Ising and frustrated J1-J2 1-D and 2-D spin systems and revealed significant improvements in both stability and convergence compared to established optimizers such as Adam, minSR, and SPRING. The team successfully fine-tuned the 1.5 billion parameter RWKV-7 model and demonstrated NQS optimization at a scale that exceeds previous work by three orders of magnitude.
