Asking for What Was Never Requested: Horizontal and Vertical Proactivity in Agents
Paper defines horizontal and vertical agent proactivity and trains a questioner that beats a 15x larger prompted model.
The paper introduces two axes of proactivity content: horizontal (pursuing unstated information implied by context) and vertical (pursuing needs revealed only by earlier evidence), scored via need graphs recovered from benchmark decompositions without model judges. The proposed Q&D method trains a questioner to prefer questions whose continuations retrieve more required evidence. On held-out multi-hop QA splits, it outperforms a prompted model 15 times larger on two of three benchmarks, and transfers zero-shot to simulated customer-service and retail agent tasks.
- Defines horizontal and vertical proactivity as content dimensions of agent behavior
- Need graphs score proactivity from transcripts without a model judge
- Q&D trains questioner via retrieval outcomes, no reward model needed
- Beats 15x larger prompted model on two of three multi-hop QA benchmarks
- Zero-shot transfer to customer-service and retail agent simulations
From the original article
1 min read at huggingface.co
An agent that uses tools typically responds to what the user explicitly asks, yet completing the task may require information the user never requested. Work on proactive agents mainly studies whether and when an agent should act on its own, not what information it should pursue. We study a distinct axis of proactivity: its content.
By Ido Levy, Asaf Yehudai, Segev Shlomov · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.