Researchers proposed a formal basis GAI for self-teaching agents

Researchers have published a work Generalized Agent Iteration (GAI), offering a unified mathematical structure for describing iterative policy improvement and recursive self-improvement (RSI). The authors aim to combine classical learning methods with concepts of autonomous intelligence development within a single paradigm.

The basis of GAI is a model of an agent as a configuration of modifiable components, where the learning process is represented as a cyclic algorithm of evaluation and subsequent optimization. The system uses two key parameters: the arrangement of the improvement mechanism (inside or outside the agent) and the presence of an external coordinate system for evaluating results. These variables allow classifying systems by their polarity: from rigidly defined goals to fully self-referential structures.

The proposed approach allows formalizing the differences between classical iterative policy improvement (GPI) and recursive self-improvement. The authors use GAI to systematize existing architectures, which makes it possible to identify defects in self-learning processes under individual conditions. This allows comparing heterogeneous systems on a single coordinate scale.

The work is aimed at creating a fundamental basis for analyzing and designing future autonomous agents. Formalizing the conditions under which an agent is capable of self-modification helps avoid the problem of "goal drift" and ensures predictability of system behavior in the long term. The study is available under the CC BY-NC-ND 4.0 license.