PersonaCUABench: every top CUA loses 28-45% to real-user asks

Nine frontier computer-use agents each lose between 28 and 45 percent of their task completion when the same instruction is rewritten in a real user's voice. That is the central finding of PersonaCUABench, a benchmark from a multi-institution team including Harvard, MIT, Carnegie Mellon and USC that holds 74 tasks across 11 domains fixed and varies only who asks.

A persona agent playing one of 44 users rewrites each request in that person's own words and answers the CUA's follow-up questions, while the agent sees only the dialogue, never the profile or the original instruction. Across 3,330 runs, mean task completion falls by 36% relative to the original requests, dropping from 0.689 to 0.441. The slate tested is current: GPT-6 Astra, Claude Opus 5.5, Grok 4.6, Gemini 3.8 Flash, DeepSeek V4.1 Flash, Kimi K3, MiniMax M3, Qwen 3.8 27B and Muse Glimmer 30B. "Each model's relative loss lies between 28% and 45%," the paper reports.

The failure split implicates evaluation infrastructure as much as the agents. In 59.9% of cells that lose points, the agent engages with what the user asked for and still gets it wrong. Another 11.2% are stops the user explicitly requested that the automated checker scores as failures. Even when a persona request is fully completed, the user gives the top satisfaction rating in only 62.7% of cells, against 84.1% for the original requests.

The authors' prescription is direct: "CUAs should therefore be evaluated not only on task completion but also on whether they meet the requirements and constraints users express."