Information elicitation is an open‑ended information‑seeking task where an interaction can branch into many potentially valuable directions, requiring the elicitor to continuously decide which pieces of information to pursue as new data appear. This work examines delegating those decisions to foundation models (LLMs), a topic that has received little systematic study. We first compare 11 LLMs spanning several model families and parameter scales on a shared set of information items and elicitation objectives, measuring how each model judges information value. Then we build a controlled elicitation simulation: different models encounter the same information space and follow the same selection rule, isolating model‑specific value judgments from question generation and respondent behavior. Using this framework we characterize the emerging breadth‑depth behavior—whether a model tends to explore widely or dig deeply during elicitation. We also investigate how interaction history reshapes value evaluation and subsequent selection. Sensitivity analyses and ablations across the number of opportunities, response label definitions, presence of history, and explicit redundancy relevance test the robustness and limits of our findings. The project code, data, and trajectory files are released at https://github.com/infosenselab/open-elicitation.
Review