From Static to Systematic: A Survey on Efficient Inference Across LLM, RL, and Agentic Eras

The paradigm of LLM inference is undergoing a fundamental transformation, evolving from simple sequence generation to complex systematic reasoning and action. In this survey, we argue that efficient inference methods must be understood not as static solutions but as evolving techniques whose applicability shifts with each new computational paradigm. We organize the literature into a three-level taxonomy across three eras---the LLM Era, the RL Era, and the Agentic Era---and introduce the Three-Tax view (KV Tax, Thinking Tax, Action Tax) as a unified lens for characterizing the cumulative efficiency burden each era introduces. Through cross-era analysis, we identify which methods transfer, which require fundamental adaptation, and which optimization dimensions emerge only with new paradigms. Beyond cataloguing existing work, we distill the underlying principles that govern how inference optimizations evolve, offering a forward-looking roadmap for anticipating where new bottlenecks will emerge as the paradigm continues to shift.