IBM's Q&D trains proactive agents to ask better questions, beating a 15x larger model

ibm · hf · 2026-09-30

IBM studies an overlooked axis of agent proactivity: what information an agent should pursue beyond the user's explicit request.

Defines horizontal proactivity (pursuing unstated info already implied by context) and vertical proactivity (pursuing needs only revealed by earlier evidence).
A need graph recovered from a benchmark's own decomposition enables scoring both forms—and whether the agent stops at the right time—without a model judge.
Q&D (questioner and drafter) trains a questioner to prefer questions whose continuation retrieves more required evidence, with no reward model.
At equal retrieval spend, the trained questioner beats the prompted same-size model and outperforms a 15x larger model on two of three multi-hop QA benchmarks.
In a simulated customer-service agent, it completes more tasks while asking fewer questions, and in retail beats the 15x larger model with fewer follow-up turns.

More from coding & agent