PARE: A Stateful Benchmark for Proactive AI Assistants

Building proactive AI assistants that anticipate user needs and act without explicit commands is an appealing goal, but current evaluation methods are fundamentally mismatched to the problem.

Existing benchmarks model applications as flat tool-calling APIs, which strips away the stateful, sequential nature of real digital environments.

Without realistic user simulation that captures navigation history and context-dependent action spaces, researchers cannot reliably test whether an agent correctly infers goals, chooses the right moment to intervene, or coordinates across multiple apps.