Morini & Terna's Cross Targets RL lets agents self-define behavior

In a new arXiv paper (2608.03445), researchers Matteo Morini and Pietro Terna explore how artificial agents can develop their own behavior strategies without predefined rewards or external supervision. Their key innovation is a technique called Cross Targets (CT), which trains a learning agent by crossing data between its guesses about the action to take and its guesses about the following results. This unsupervised approach aligns with the bounded rationality hypothesis, where agents use simple, efficient learning rather than exhaustive optimization.

The authors demonstrate CT through a 'blind' strategy application: random walkers on a graph learn to stay within homogeneous regions, which in turn solves a node classification problem—without ever being explicitly told what 'correct' classification means. The results are described as 'unexpected,' and the authors argue that these emergent behaviors are largely effects of observer interpretation. The paper builds on earlier work published in the Carlo Alberto Notebooks (No. 638, January 2021), and it sits at the intersection of reinforcement learning, social information networks, and agent-based modeling. While the immediate use case is graph-based classification, the broader implication is that agents can self-organize coherent behaviors with minimal external guidance, a potentially valuable insight for multi-agent systems, network analysis, and autonomous decision-making in complex, unstructured environments.