Researchers have proposed a framework for combining large language models (LLMs) with reinforcement learning agents. They formalized the hybrid LLM-planner and RL-controller architecture as a Goal-Augmented Markov Decision Process. The team showed that using an LLM's per-state progress score as a bounded potential function preserves the optimal policy set, even when the LLM scores are inaccurate. This is a stronger guarantee than what general LLM-as-reward approaches provide. The researchers numerically verified their result on a small MDP under four configurations. --- Why it matters: This work matters to engineers and researchers in AI because it provides a more robust way to combine LLMs with reinforcement learning agents, which can improve the performance of hybrid RL systems. Source: https://arxiv.org/abs/2608.18008