A newly posted paper on Hugging Face proposes Generalized Agent Iteration (GAI), a formal framework intended to describe iterative policy improvement and recursive self-improvement (RSI) as two instances of a single learning paradigm.
The authors frame the motivation around a definitional gap: RSI is increasingly claimed at many scales, yet no single framework formally describes these emerging instances. Its classical counterpart, iterative policy improvement, is captured by generalized policy iteration (GPI), a framework with well-understood theoretical properties — but only where the update principle and the evaluation base lie outside the agent.
GAI defines the agent as a configuration of modifiable components within a system and models learning as a cycle of agent evaluation and agent improvement. Two "dials" then distinguish instances of the paradigm: whether the improving mechanism is part of the agent, and whether the standard it is measured against is grounded outside it.
According to the paper, the first dial marks the boundary between GPI and RSI, while the second determines a system's polarity as anchored, goal drift, or fully self-referential. The authors use these coordinates to place existing systems on the same two axes and to state the defects of recursive self-improvement one condition at a time.
The paper positions itself as a first step toward a formal characterization of RSI that builds on the classical account, makes existing systems comparable, and offers a principled basis for analyzing and designing new ones. The authors also note that GPI is not general enough to enclose RSI, motivating the broader framework.