Who Takes the Lead? Model Ability and Task-Conditioned Collaboration with Humans
Stronger models handed control to humans on roughly 80% of opening decisions when the human had a shorter route.
Researchers tested four models in a shared-control maze where an agent decided whether to move its own avatar or give the human a turn. Starting positions changed whose route to the common target was shorter. Models with reliable solo performance almost always acted themselves when the human had no advantage. The substantially weaker Command R showed no such selectivity, supporting the paper’s argument that human involvement alone is insufficient to judge collaboration.
HF Daily Papers' note
Researchers tested four models in a shared-control maze where an agent decided whether to move its own avatar or give the human a turn. Starting positions changed whose route to the common target was shorter. Models with reliable solo performance almost always acted themselves when the human had no advantage. The substantially weaker Command R showed no such selectivity, supporting the paper’s argument that human involvement alone is insufficient to judge collaboration.
HF Daily Papers' note
score 4