Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2608.18554.

Pipeline notes

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

CentaurBench evaluates LLMs on their ability to augment rather than automate real-world work tasks. Across seven economically grounded tasks, automation rankings and augmentation rankings correlate only modestly; the automation winner loses on augmentation in five of seven tasks, and assistance is not reliably positive.

Challenges the assumption that stronger automation implies stronger assistance, offering a framework for selecting models for human-AI and multi-agent collaboration where standard benchmarks mislead.

Cited by

No citations on record.