Average-reward reinforcement learning poses a known difficulty: its Bellman operator contracts in span only when the underlying Markov chain is mixing. The new arXiv paper "Mixing Regularization for Average-Reward Q-Learning and Q-Planning" takes this problem as its starting point, noting that existing work typically assumes a contraction—an assumption the abstract suggests is restrictive.
The paper's title points to the proposed remedy: adding a mixing regularization term to average-reward Q-learning and Q-planning. Because the abstract is truncated, the precise form of the regularization and the theoretical results are not available from the source alone.
Where the source is clear, it frames the contribution as addressing the gap between the mixing requirement and the contraction assumption common in prior work. No other sources were provided, so there is no independent account to compare against.