Source record

Published by arXiv cs.HC (T1 source). The original is at https://arxiv.org/abs/2610.22990.

Pipeline notes

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

PersonaCUABench evaluates computer-use agents on 74 fixed tasks across 11 domains by varying only the persona (44 personas, four groups) who rewrites each instruction. Across 3,330 runs on nine backbones, mean task completion drops 36% versus original instructions; top user satisfaction falls from 84.1% to 62.7%.

Quantifies a real gap between benchmark task completion and what diverse users actually want. 3,330 runs, code released, method is reproducible.

Cited by

No citations on record.