The paper "Aligning Language Models to Explicitly Handle Ambiguity" addresses the challenge of managing ambiguous queries in large language models (LLMs). The authors propose a novel pipeline called Alignment with Perceived Ambiguity (APA), which leverages the model's intrinsic knowledge to handle ambiguous inputs more effectively. The key idea is to quantify the model's perceived ambiguity using information gain (INFOGAIN) and use this metric to guide the alignment process. The APA pipeline consists of four stages: initial prediction assessment, perceived ambiguity detection, response construction, and supervised fine-tuning. Experimental results on various question-answering (QA) datasets demonstrate that APA significantly improves the model's ability to handle ambiguous queries while maintaining its performance on unambiguous queries. The paper also introduces three new datasets—AmbigTriviaQA, AmbigWebQuestions, and AmbigFreebaseQA—to provide a comprehensive framework for evaluating models' robustness in managing ambiguity. The authors conclude by discussing the limitations of their work and future directions, including extending the methodology to broader domains and more complex types of ambiguities.The paper "Aligning Language Models to Explicitly Handle Ambiguity" addresses the challenge of managing ambiguous queries in large language models (LLMs). The authors propose a novel pipeline called Alignment with Perceived Ambiguity (APA), which leverages the model's intrinsic knowledge to handle ambiguous inputs more effectively. The key idea is to quantify the model's perceived ambiguity using information gain (INFOGAIN) and use this metric to guide the alignment process. The APA pipeline consists of four stages: initial prediction assessment, perceived ambiguity detection, response construction, and supervised fine-tuning. Experimental results on various question-answering (QA) datasets demonstrate that APA significantly improves the model's ability to handle ambiguous queries while maintaining its performance on unambiguous queries. The paper also introduces three new datasets—AmbigTriviaQA, AmbigWebQuestions, and AmbigFreebaseQA—to provide a comprehensive framework for evaluating models' robustness in managing ambiguity. The authors conclude by discussing the limitations of their work and future directions, including extending the methodology to broader domains and more complex types of ambiguities.