Reproduce “Knowing but Not Showing: LLMs Recognize Ambiguity but Rarely Ask Clarifying Questions” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jinyan Su, Claire Cardie

Large language models (LLMs) demonstrate an internal capacity to recognize ambiguous user queries when explicitly asked to judge them, often with 60-80% accuracy. However, in standard question-answering scenarios, these models overwhelmingly provide direct answers over 95% of the time, rarely asking clarifying questions or refusing to respond, a behavior that is further suppressed by the provision of retrieved context.

Key insights

  • There is a notable "knowing but not showing" phenomenon where LLMs possess internal ambiguity awareness but do not consistently translate it into external clarification behavior.
  • While retrieved context generally boosts QA accuracy, it amplifies the difficulty gap between ambiguous and unambiguous queries and, counter-intuitively, does not consistently improve explicit ambiguity recognition.
  • LLMs demonstrate a bias towards over-predicting ambiguity when asked to judge it, yet almost universally default to providing direct answers in a question-answering context.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.