EOPA framework boosts proactive AI timing accuracy by 19.8 points

Most AI assistants are reactive—they wait for you to speak first. Proactive assistants instead decide when to jump in based on your activity context, but timing is deeply personal and hard to pre-program. Now researchers from multiple institutions (including Yufeng Wang and Mingkui Tan) propose EOPA, or Evidence-driven Online Preference Adaptation, in a paper on arXiv (2608.04416). EOPA grounds interaction-timing preferences in two evidence carriers: temporal preference anchors (when users tend to accept interruptions) and evidence-bearing activity prototypes (what user states signal a good moment). At each polling step, EOPA estimates evidence from these carriers using user-prior-smoothed estimation and uncertainty-guided scaling, then fuses them to decide between interaction or silence. When it does interact, an LLM generates a context-aware response using high-quality historical responses as examples—no LLM reasoning or retraining needed for updates.

EOPA's big win is efficiency and personalization from sparse feedback. On the ProPerSim-based benchmark, it beat the strongest baseline by 19.80 points in interaction-timing F1 score, while dramatically reducing inference latency for both silence and interaction steps. Even more striking: the average daily adaptation cost dropped from 11.41 seconds to just 0.39 seconds. That means EOPA can continuously learn a user's preferences nearly instantly, without expensive model updates. The framework handles the challenge of sparse interaction-worthy moments by using measurable contextual evidence rather than raw trial-and-error. For product teams building proactive assistants, this suggests a practical path toward assistants that know when to speak up—and when to stay quiet—without needing massive training runs or constant user corrections.