Model Curiosity
The behavioral characteristic of large language models proactively posing epistemic questions, that is, asking questions in order to uncover knowledge gaps, implicit assumptions or alternatives. It is distinguished from functional questions that advance a process or confirm information, and is an annotatable, statistically measurable indicator of interaction behavior in human-machine dialogue.
Timeline (last 90 days)
A study titled 'What Makes a Question Curious? Measuring Model Curiosity in Human-LLM Interactions' systematically defines 'model curiosity' for the first time and measures its distribution and role in real human-machine conversations.
The study constructs a theory-grounded taxonomy that distinguishes curiosity-driven questions from general model questions: the former point toward unknown domains, while the latter are closer to customer-service-style confirmation or task clarification.
The study analyzed four conversation datasets and found that curiosity-driven questions accounted for only 17%–28% of all model questions.
When models ask curiosity-driven questions, users' subsequent responses are more exploratory and richer in epistemic information.
The authors conducted a human study to evaluate the practical help of model curiosity to users, and the results show that curious questions can encourage users to explore, with more pronounced effects in the three scenarios of advice, creativity, and research.
In fact-lookup tasks, excessive questioning may instead be seen as interference.
The research points toward a more refined goal—teaching models to 'ask selectively'.
From an engineering perspective, the evaluation system needs upgrading: in addition to answer accuracy, a dimension of question quality should also be included.
The pattern differences across datasets suggest that scenario adaptation remains a difficulty.
All Facts (9)
A study titled 'What Makes a Question Curious? Measuring Model Curiosity in Human-LLM Interactions' systematically defines 'model curiosity' for the first time and measures its distribution and role in real human-machine conversations.
50%UnverifiedThe study constructs a theory-grounded taxonomy that distinguishes curiosity-driven questions from general model questions: the former point toward unknown domains, while the latter are closer to customer-service-style confirmation or task clarification.
50%UnverifiedThe study analyzed four conversation datasets and found that curiosity-driven questions accounted for only 17%–28% of all model questions.
50%UnverifiedWhen models ask curiosity-driven questions, users' subsequent responses are more exploratory and richer in epistemic information.
50%UnverifiedThe authors conducted a human study to evaluate the practical help of model curiosity to users, and the results show that curious questions can encourage users to explore, with more pronounced effects in the three scenarios of advice, creativity, and research.
50%UnverifiedIn fact-lookup tasks, excessive questioning may instead be seen as interference.
50%UnverifiedThe research points toward a more refined goal—teaching models to 'ask selectively'.
50%UnverifiedFrom an engineering perspective, the evaluation system needs upgrading: in addition to answer accuracy, a dimension of question quality should also be included.
50%UnverifiedThe pattern differences across datasets suggest that scenario adaptation remains a difficulty.
50%