π-Bench: Evaluating Proactive Personal Assistant Agents in Long-Horizon Workflows

Researchers introduced π-Bench, a benchmark for proactive personal assistant agents.

It includes 100 multi-turn tasks across five user personas, testing hidden intent detection, task dependencies, and cross-session continuity.

The goal is to measure whether assistants can anticipate user needs, not just answer explicit requests.

The benchmark offers a more realistic test for everyday and work-oriented long-horizon assistant behavior.