IBM Paper Splits Agent Proactivity Into Horizontal and Vertical

An agent that uses tools typically responds to what the user explicitly asks. The new IBM and Weizmann paper 'Asking for What Was Never Requested' studies what an agent should chase when the user has not asked but the task requires it.

The authors split that behavior along two axes. "Horizontal proactivity pursues unstated information that the current context already identifies, and vertical proactivity pursues needs that only earlier evidence reveals," they write. Both, they argue, can be scored from a transcript without an LLM judge, using a need graph "recovered from a benchmark's own decomposition."

Their training method, called Q&D for questioner and drafter, dispenses with the reward model entirely. The questioner learns "to prefer the question whose continuation retrieves more of the required evidence." No judge, no human labels on individual asks.

On three multi-hop question-answering benchmarks (MuSiQue, StrategyQA, and 2WikiMultiHopQA), the trained questioner "improves both forms of proactivity over the same model, prompted, and outperforms a prompted model" roughly fifteen times its size on two of the three sets, at equal retrieval spend. On 2WikiMultiHopQA it trailed the larger baseline by 3.5 points. Depth-weighted recall gains across the three sets fell between 5.4 and 12.5 percentage points.

Proactivity is a busy lane right now. Our tracker logs 413 agent stories in the last 90 days, and a separate paper posted the same week takes its own swing at foundations of proactive agents. What separates this one is the attempt to make 'asked the right question' measurable without paying an evaluator model on every step.