PersonaCUABench shows 36% performance drop for diverse personas
October 15, 2026
PersonaCUABench reveals that computer-use agents see a 36% drop in task completion when instructions are filtered through different user personas. The benchmark tests how agents handle implicit constraints and persona-specific communication styles.
HOW THIS AFFECTS YOU
●
builderYou need to move beyond fixed task specifications to make agents robust for diverse real-world users.
●
researcherThis introduces a critical new dimension of evaluation: persona-congruent task completion.