In this talk, we explore three contributions in QML relevant for near-term quantum devices. First, we introduce the Quantum Generative Adversarial Autoencoder (QGAA), a novel architecture that fuses quantum autoencoders with adversarial training to learn compact latent representations of quantum states. We use the model to generate energy profiles of simple molecules. Next, we shift to the problem of thermal state preparation with Meta-VQT and NN-Meta VQT—two meta-learning variational algorithms that generalize across many-body Hamiltonians. We achieve up to 30× speedups over traditional methods such as VarQITE in few qubit systems simulation. Finally, we present HELIA, a hardware-efficient ansatz grounded in Lie algebra dynamics, paired with a dual training strategy that combines classical simulation and quantum gradients. This hybrid approach mitigates barren plateaus and reduces quantum resource usage by up to 60%, while maintaining high accuracy in ground-state estimation and phase classification tasks. These works lay the groundwork towards next milestones of QML that is deployment at scale, is HW tested, and has a potential for quantum advantage.
