We develop a score-based variational inference method that learns a product-of-t-experts model via a Feynman-identity latent-variable formulation, reducing inference to a sequence of convex quadratic programs with provable convergence.