Stronger models defer maze turns only when humans hold the advantage
Original title: Who Takes the Lead? Model Ability and Task-Conditioned Collaboration with Humans
Reliable solo solvers almost never deferred without a human route advantage, but deferred on about 80% of initial decisions when the human path was shorter.
The paper asks whether falling human involvement by capable agents is failed collaboration or a correct division of labor. A shared-control maze lets an agent assign each team turn to itself or a person, while initial placement sets the human's shortest-path advantage. GPT 5.6 Sol, Command A+, Qwen 3.8 27B, and Command R were measured solo, then co-played with people. Stronger solo models were selective; Command R, the weaker solo model, was not.
Unconfirmed: The selectivity pattern is the authors' experimental finding and is not independently replicated here.
- Conditions were human advantage, neither advantage, and agent advantage.
- Solo tests used one avatar and full-map observability.
- Strong models acted themselves when the human had no route advantage.
- They involved the human on roughly 80% of initial decisions when the human route was shorter.
- Command R showed no such route-conditioned selectivity.
From the original article
1 min read at huggingface.co
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 when it may instead be a correct division of labor, and the rate of human involvement alone cannot tell the two apart.
By Weien Li, Haochen Liu, Zeyu Li · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.