A new paper on arXiv investigates how language-model agents decide when to collaborate with humans. As models become more capable, they tend to carry more of a task independently, asking for help less often. The title, "Who Takes the Lead?", points to the central question of how roles are allocated between model and human.
The abstract argues that falling reliance on a human partner is easy to misread as a failure of collaboration. Instead, it may be a correct division of labor, where the model takes the lead on parts it can handle and defers only when needed. This reframing shifts attention from how often agents ask for help to whether they ask at the right times.
Because the source is a single arXiv abstract, the findings are presented without comparison to other work. The paper's contribution appears to be a framework for evaluating collaboration that accounts for changing model capabilities and task conditions, rather than treating interaction frequency as the sole measure of success.