Most approaches to optimizing swarms of large language models focus on the final utility of the task, according to a new arXiv paper. That means the way information is processed, integrated, and relied upon within individual interactions remains largely hidden. The authors argue this is a gap: optimizing only for outcomes can obscure why a swarm succeeds or fails.

The paper proposes "behavioral probes" as a way to make these internal information flows observable. By probing how individual agents use and pass along information, researchers could gain insight into the swarm's dynamics. The abstract is truncated, so specific methods are not yet available, but the motivation is clear: move beyond black-box utility and toward a more transparent view of multi-agent reasoning.