A new arXiv paper tackles the question of when a language model should ask for clarification. The authors note that ambiguous inputs can admit multiple interpretations, each potentially yielding a different answer. But they argue that existing work largely treats this ambiguity as a property of the input alone, ignoring the fact that for a particular language model, only some of those alternatives are actually relevant.
The paper's central move is to make ambiguity model-dependent. Instead of asking whether an input is ambiguous in the abstract, the authors propose estimating how ambiguous it is for a given model, based on the alternatives that model itself can generate. This shifts the focus from the text to the interaction between text and model.
Because the abstract is truncated, the specific estimation method and experimental results are not fully detailed here. But the core claim is clear: clarification needs are not universal, and what one model finds ambiguous another may not. This has practical implications for building more efficient and reliable question-answering and dialogue systems. The preprint is available on arXiv under identifier 2610.20052.