Reading up on Jeff Auriemma

The paper proposes a four‑category taxonomy—Initialization, Reasoning, Orchestration, Synthesis—for LLM inference in agents and introduces Reasoning Yield to quantify useful versus mechanical token usage.