PersonaCUABench shows computer-use agents fail real user personas across 74 tasks
Haoming Li · hf · 2026-10-30
Researchers from multiple universities introduce PersonaCUABench, a benchmark for computer-use agents (CUAs) serving diverse user personas. It shows CUAs frequently fail to satisfy real user requirements and constraints, reducing task completion and satisfaction versus fixed benchmark instructions.
- Holds 74 tasks across 11 domains fixed and varies only who asks: a persona agent plays one of 44 personas in four groups, rewriting each instruction in that user's own words and answering the CUA's questions while the CUA sees only the public interface.
- Key insight: users omit requirements, add constraints, or require approval before consequential actions — completing fixed benchmark instructions doesn't guarantee serving the person who asked.
More from coding & agent
- VibeGame from Nanjing University builds complete 2D games from prompts with adversarial agent teams — nanjinguniv · 2026-10-30
- WebAnything Auto-Rebuilds Real Websites into Scalable Web Agent Environments — Tianci Xie · 2026-10-30
- WorkFactory Scales Professional Task Generation for Agent Training, Yields WorkPilot — IQuestLab · 2026-10-30
- Kimi launches 3D Make: connects to Blender and CAD tools to build models, scenes, and games — Kimi_Moonshot · 2026-10-30
- Kimi team shares its internal 3D workflow: parametric vs mesh modeling with K3.1 Pro — Kimi_Moonshot · 2026-10-30
- Animated short made with ComfyUI: ChatGPT poses, Qwen restyle, MiniMax H3 — bseoan · 2026-10-30