Who Takes the Lead? Model Ability and Task-Conditioned Collaboration with Humans

Who Takes the Lead? Model Ability and Task-Conditioned Collaboration with Humans. By Weien Li and five others, from McGill University, MIT, and three other institutions. Published on the nineteenth of October. No code repository is linked to it.

As language-model agents grow more capable, they can carry more of a task without asking anyone for help. Falling reliance on a human partner is easy to read as a failure of collaboration. It may instead be a correct division of labor, and the rate of human involvement on its own cannot tell those two apart. The finding here is that, in a controlled task, stronger agents involve people selectively: rarely when the person has no edge, and often when the person does. A weaker agent shows no such selectivity.

To get that control, the authors built a shared-control maze. An agent-controlled avatar and a human-controlled avatar share one objective: reaching a common target in as few team turns as possible. On each team turn, the agent decides between acting itself and handing the move to the person. Initial placement sets the human's shortest-path advantage, and everything else about the task stays unchanged. That yields three conditions. The human has a clear advantage, neither side does, or the agent does. Because only the starting positions change, any difference in how often the agent hands over can be read against who was actually better placed, rather than against a changed task.

Four models were evaluated: GPT five point six Sol, Command A plus, Qwen three point eight at twenty-seven billion parameters, and Command R. Each was first measured for solo capability, with a single avatar and full-map observability, and then run in co-play with human participants across the three advantage conditions. The solo runs establish how capable each model is alone, which is what the co-play behaviour is then compared against.

Across the four, stronger solo performance goes with selective, route-conditioned human involvement. The models that solve the maze reliably on their own act for themselves almost without exception when the human holds no advantage. When the human route is shorter, they involve the human on roughly eighty percent of initial decisions. Command R, the model with substantially weaker solo performance, does not show that pattern.

The setting is a single kind of task, a shared-control maze, where advantage is set by placement and measured as shortest-path length. Four models were tested, and the contrast in selectivity rests on one weaker model, Command R. The result is stated as an association between solo performance and selective involvement, not as a cause, and the eighty percent figure applies to initial decisions when the human route is shorter.

The practical lesson concerns how collaboration is measured. A falling rate of human involvement is not, by itself, evidence that an agent has stopped cooperating. In this maze, the capable agents asked for help when the person was better placed to give it, and not otherwise, which is what a sensible division of labor would look like.