ProDVI: Using LLM-Generated Code to Initialize RL Value Networks

A new framework called ProDVI (Programmatic Dynamics Priors for Value Network Initialization) has been proposed to address sample inefficiency in deep reinforcement learning (RL). The method leverages large language models (LLMs) to generate executable Python functions that encode coarse hypotheses about environment dynamics, which are then used to produce synthetic transitions for initializing RL agents. This approach eliminates the need for pre-collected datasets, high-fidelity simulators, or meta-learning over related tasks, which are often difficult to access or unavailable. The paper, arXiv:2608.06015, was announced as a cross-type submission. The framework is designed to improve the sample efficiency of RL agents by providing them with informative initializations based on commonsense and domain knowledge encoded in LLMs.