Source record

Published by arXiv cs.HC (T1 source). The original is at https://arxiv.org/abs/2610.27148.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

A controlled study of how language-model agents allocate turns between acting autonomously and deferring to a human partner in a shared maze task, with three advantage conditions. Stronger solo-performing models showed selective human involvement (roughly 80% when the human held the shorter route), while the weakest model showed no such selectivity.

Separates a correct division of labor from a collaboration failure in human-AI teams — a useful framing for designing agent workflows that hand off to people.

Cited by

No citations on record.