Entropy Regularization

Entropy regularization adds an entropy-based term or constraint to an objective to control how concentrated or diverse a model’s distributions and representations become. Methods use Shannon, Rényi, or Tsallis entropy in settings such as reinforcement-learning exploration, test-time adaptation, language-model training, latent-variable learning, and entropically regularized optimal transport solved with Sinkhorn methods. Some approaches minimize prediction entropy for confident decisions, while others preserve or increase entropy to encourage exploration, prevent collapse, or maintain diverse outputs; adaptive and class-, token-, or perceptual-level controls aim to apply these effects selectively. The central challenge is balancing useful confidence or sparsity against bias, lost diversity, and unstable optimization, alongside developing guarantees for convergence and statistical performance.