Code underlying the publication: Bayesian Meta-Reinforcement Learning with Laplace Variational Recurrent Networks

Laplace-VRNN: Bayesian Meta-Reinforcement Learning with Laplace Variational Recurrent Networks — JAX implementation for the RLC 2025 paper (De Vries, He, de Weerdt, Spaan). Applies a post-hoc Laplace approximation over RNN hidden states to add uncertainty quantification to memory-based meta-RL agents without retraining or architecture changes. The task posterior is Gaussian centered at the hidden state with inverse covariance from Jacobian outer products. Matches full variational baselines in returns with fewer parameters, and reveals that standard RNN meta-RL agents produce overconfident, inconsistent posteriors.