Speaker
Description
Predicting the real-time dynamics of quantum many-body systems is a central challenge in quantum physics, where the exponential growth of Hilbert space makes direct simulation rapidly intractable. In this contribution, we explore attention-based models as a framework for learning and extrapolating quantum dynamics.
Attention mechanisms are the core building blocks underlying modern large language models: in language, they identify which previous words are relevant for predicting the next one; here, we use the same idea to identify which previous quantum states are relevant for predicting future time evolution.
First, we introduce a variational quantum self-attention mechanism in which future elements of a sequence are generated from overlap-weighted combinations of past data. The required nonlinearity arises naturally from quantum interference between state overlaps, and the training objective can be expressed directly as an observable expectation value. This provides a quantum-native route to sequence prediction and yields a quadratic speedup in sequence length relative to classical attention, while remaining applicable to classical data and transverse-field Ising model trajectories.
Second, we present ongoing work showing that classical attention models can learn the structure of quantum time evolution from datasets of trajectories generated within a given Hamiltonian class. Given a short sequence of evolved states up to time t, the model predicts the continuation up to times T≫t. Across TFIM, Fermi-Hubbard, XXZ, Heisenberg, and 3-MAX-SAT Hamiltonians, we obtain high test fidelities and accurate predictions of observables such as magnetization.
We discuss key challenges for generative quantum dynamics, including long-time stability, exposure bias, and the representativity of training ensembles. Overall, our results suggest that attention models, both classical and quantum, provide flexible inductive biases for learning and extrapolating complex quantum many-body dynamics.
References:
https://arxiv.org/abs/2602.06699