Asking for What Was Never Requested: Horizontal and Vertical Proactivity in Agents

This work formalizes two distinct axes of agent proactivity: horizontal proactivity for unstated context-identified information, and vertical proactivity for evidence-revealed needs. To achieve this, the authors propose Q&D (questioner and drafter), a method that trains a questioner to retrieve required evidence without relying on a reward model or judge. Evaluated on multi-hop QA benchmarks, the trained questioner enhances both forms of proactivity, outperforming a prompted model 15times larger on two out of three tasks. When deployed in simulated customer-service and retail agents, the method completes more tasks while requiring fewer follow-up turns.

The trained questioner outperforms a prompted model 15times larger in the same role on two of the three benchmarks.

In an interactive customer-service agent, it completes more tasks while asking fewer questions.

The performance gain persists after controlling for question volume and length.