Reproduce “Aligning Language Models to Explicitly Handle Ambiguity” on OpenResearch

About this paper

Hyuhng Joon Kim, Youna Kim, Cheonbok Park, Junyeob Kim, Choonghyun Park, Kang Min Yoo, Sang-goo Lee, Taeuk Kim

The paper introduces Alignment with Perceived Ambiguity (APA), a four-stage pipeline that enables language models to explicitly identify and manage ambiguous queries by either requesting clarification or providing direct answers. APA demonstrates consistent performance gains of up to 6 points in ambiguity detection F1 across various LLMs and datasets while enhancing general question-answering accuracy.

Key insights

• LLMs can implicitly perceive query ambiguity, and this perception can be effectively quantified using INFOGAIN from a self-disambiguation task, providing a robust signal for targeted alignment.
• Explicitly training LLMs to handle perceived ambiguity not only improves their ability to request clarification but also enhances their accuracy on unambiguous questions, indicating a broader refinement of core QA capabilities.
• Data quality, particularly identifying and utilizing samples the model itself perceives as ambiguous, is more impactful than data quantity for effective ambiguity handling alignment, enabling strong performance with smaller datasets.

Summary by alphaXiv.