Can systems surface ambiguity instead of guessing user intent?

PleaSQLarify's authors argue that natural language database interfaces fail not because users are careless but because "systems typically sample the most probable interpretation and perform the corresponding system action" when a request is ambiguous, collapsing uncertainty into one guess instead of surfacing it. They cite the CoSQL dataset, where across "2'459 interactions where users asked SQL experts to write basic SQL queries conversationally, 31.9% of which contained an instruction requiring clarification of the user's intent" — ambiguity is common, not an edge case. Their proposed fix reframes the problem through pragmatics: users are "economical speakers" who underspecify, while systems hold priors over possible actions "that may not align with the users'." The system they built groups raw query features into "semantically interpretable decision variables," selects among them by expected information gain, and exposes them through a visual interface with three linked views — an Action Space, a Decision Space, and a Predicted Query — so a user can see alternative SQL interpretations, disambiguate specific variables, and watch belief updates persist across turns. In a study with twelve participants, this helped users "recognize alternative interpretations they had not anticipated," and on the AMBROSIA benchmark the algorithm "reduces semantic uncertainty in fewer turns than baselines," converging in "two to eight clarification steps."

The mechanism borrows directly from Clark and Wilkes-Gibbs's collaborative model of reference: pragmatic repair proceeds incrementally, with speakers starting from a minimal description and adding "just enough new information to distinguish the intended meaning from remaining alternatives," following a "principle of least collaborative effort" that avoids over-specification up front. PleaSQLarify's contribution is making this collaborative grounding process visible and interactive rather than implicit: instead of a single clarifying question, the interface lets users explore the system's action space directly, and the information-gain-ranked decision variables determine which ambiguity gets resolved first. The user study surfaced three distinct ways participants used this: "broad exploration, targeted prefiltering, and output-driven assembly," which the authors read as evidence that interfaces should support multiple repair strategies rather than enforcing one clarification path.

This sits close to When should AI agents ask users instead of just searching?, which also grounds clarification in conversational pragmatics rather than intent classification; PleaSQLarify adds a concrete algorithmic mechanism (expected-information-gain selection over grouped decision variables) and a visual interface for exposing the action space, where the insert-expansion framework stays at the level of when to probe versus proceed. It also parallels Do spatially-anchored previews stabilize VR geometry editing tasks? in pairing visual preview with clarification to cut interaction rounds, but in a structured-query domain rather than VR geometry, and with an explicit pragmatic-theoretic (RSA) justification for which variable to ask about rather than relying on previews alone. Against Which clarifying questions actually improve user satisfaction?, PleaSQLarify's information-gain ranking is a different mechanism for the same underlying problem — choosing which ambiguity to surface first — grounded in formal pragmatics rather than empirical satisfaction testing.

The study does not establish that this generalizes beyond text-to-SQL or beyond users who already hold a well-formed intent: the authors note the experimental setup "assumes the user has a well-formed intent that can be reasonably mapped to a valid SQL query," explicitly excluding cases "when a user is unsure about their intent or changes their intent" that they observed but did not study. The candidate pool was also fixed at "around 50 language model generations" rather than updated dynamically, and several participants (P5, P7, P9) reported being unable to find their exact intended query within that fixed set. The twelve-participant sample and single domain mean the design principle — surface the action space, rank clarifications by information gain, keep updates traceable — is demonstrated for narrow, intent-clear disambiguation, not yet shown to extend to intent formation or to natural language interfaces generally.