Can we predict when users need clarification questions?

The paper argues that deciding whether an AI assistant should ask a clarifying question is not a binary ambiguity-detection problem but a question of estimating the distribution over a user's possible intents. As the excerpt states, "the distribution over intended interpretations... is often dominated by a single most-likely interpretation. In such cases, it may be preferable to forego clarification and respond to the user's request directly." To operationalize this, the paper introduces INTENT-SIM, an "uncertainty estimation approach... that determines the utility of asking a clarifying question by estimating the entropy over user intents," reporting that it "consistently outperforms existing uncertainty estimation approaches at identifying predictions that will benefit from clarification" across question answering (QA), machine translation (MT), and natural language inference (NLI).

Existing uncertainty-estimation methods, per the excerpt, "generally rely on estimating entropy over the output distribution, using high entropy to identify erroneous outputs" but "fail to identify why predictions are incorrect" — they cannot distinguish an error caused by genuine ambiguity (clarification would help) from an error caused by the model simply being wrong (clarification would not help). INTENT-SIM instead samples multiple continuations from the LLM and runs a natural-language-inference (NLI) model over them to cluster outputs by inferred intent, then measures entropy across those intent clusters rather than across raw outputs. The paper notes the added cost is dominated by the sampling step itself, with "the overhead from running our NLI model or clustering algorithm... negligible in comparison."

This sharpens the critique in Do standard NLP benchmarks hide LLM ambiguity failures?: where that note shows benchmarks erase ambiguity by assuming one gold interpretation, this paper treats the distribution over interpretations as the object to be measured and acted on directly. It also supplies a candidate fix for the failure documented in Can language models recognize when text is deliberately ambiguous? — rather than asking a model to self-report ambiguity, INTENT-SIM infers it externally from sampled outputs. Where Which clarifying questions actually improve user satisfaction? addresses how to phrase a clarifying question once the decision to ask has been made, this paper addresses the earlier decision of whether to ask at all — and, like How do users actually form intent when prompting AI systems?, treats user intent as a distribution rather than a fixed target.

The excerpt's own limitations section cautions that the true distribution over intents is highly dependent on a variety of extralinguistic factors including time, location, and individual preferences and that the paper's datasets approximate this distribution rather than measuring it directly from real users; it also does not address clarifying-question generation itself or interactions beyond a single clarification round. The claim that INTENT-SIM "outperforms" alternatives is reported as a within-paper benchmark comparison across QA, MT, and NLI test sets, not a result validated in deployed, user-facing systems.