Unifying Goal-Conditioned RL and Unsupervised Skill Learning via Control-Maximization

A new theoretical paper on arXiv (2605.06145) addresses the gap between goal-conditioned reinforcement learning (GCRL) and mutual information skill learning (MISL). The authors identify three canonical GCRL formulations and prove they are fundamentally inequivalent, potentially inducing incompatible optimal policies in the same environment. They propose a unified framework called control maximization that treats both GCRL and MISL as instances of the same principle. The paper aims to explain why skills learned through unsupervised MISL can support downstream goal-reaching, a phenomenon that lacked theoretical grounding. The work is purely theoretical and does not include experimental validation.