Found first: a primary source the press has not covered yet.
A 26-author team introduces PersonaCUABench, a benchmark that holds 74 tasks constant across 11 domains and changes only who is doing the asking. Across nine backbone models and 3,330 runs, mean task completion falls 36% relative to the same tasks under standard benchmark conditions.
What the source says
PersonaCUABench deploys a persona agent playing one of 44 personas across four groups. That agent rewrites each task instruction in the persona's voice and answers the CUA's clarifying questions; the CUA sees only the conversation, never the underlying persona profile or the original instruction. Among test cells that lose points, 59.9% involve the agent correctly identifying what the user asked for and still executing it wrong. A further 11.2% are stops the user requested that the benchmark scorer marks as failures. Even when a persona request is fully completed, users assign the top satisfaction rating in only 62.7% of cases, against 84.1% for the original benchmark instructions. Authors include Xiaomin Li and Yuexing Hao among 26 contributors; the paper was submitted October 16, 2026.
Why it matters
The benchmark isolates a single variable: replace the fixed task specification with the same task as a real person would phrase it, across 44 personas and nine backbone models. That substitution alone removes more than a third of task completions on average. The dominant failure pattern is specific. Agents are not simply tripped up by unfamiliar phrasing; most of the lost ground comes from agents that demonstrably engaged with the user's request and still got the execution wrong. The satisfaction data adds a second dimension: completion is not the same as approval. A task the system marks finished does not reliably earn the top user rating even when it was fully executed. Evaluations built on fixed specifications are measuring a condition that real user interactions do not reproduce.