Asking for What Was Never Requested: Horizontal and Vertical Proactivity in Agents
The paper tests agents on whether they seek the missing information a task actually needs, not just whether they act proactively.
It separates proactivity into “horizontal” needs already visible in context and “vertical” needs exposed only after earlier evidence is found. The authors score this with a need graph recovered from benchmark decompositions, avoiding a model judge. Their Q&D training method teaches a questioner to choose follow-ups that recover more required evidence, improving performance on held-out multi-hop QA tasks and in simulated customer-service settings. HF Daily Papers' note
It separates proactivity into “horizontal” needs already visible in context and “vertical” needs exposed only after earlier evidence is found. The authors score this with a need graph recovered from benchmark decompositions, avoiding a model judge. Their Q&D training method teaches a questioner to choose follow-ups that recover more required evidence, improving performance on held-out multi-hop QA tasks and in simulated customer-service settings. HF Daily Papers' note
score 5