Human Feedback Guides Generative Meta-Learning for Robust Generalization.

A major challenge in machine learning is ensuring models generalize effectively to environments that differ from their training data, especially when target domain data is scarce or unavailable. Researchers propose Generative Meta-Learning with Human Feedback (GMHF), a novel framework designed to overcome this "domain gap" by incorporating expert human intuition to guide the synthesis of new data. GMHF is built on a theoretical foundation that shows aligning generated data distributions with human beliefs about underlying physics can significantly reduce generalization error. The framework operationalizes this by using a Conditional Neural ODE (cNODE) as a generative digital twin, paired with a Reinforcement Learning (RL) agent. This agent iteratively refines the latent physical parameters of generated trajectories based on human feedback, effectively steering the meta-learner towards the unobserved target distribution. Empirical results on a nonlinear Duffing oscillator and a non-dynamical probabilistic model confirm that GMHF substantially reduces deployment loss as expert reliability increases, demonstrating human-AI collaboration as a rigorous catalyst for robust generalization.