Can models learn to ask better clarifying questions through self-improvement?

Users leave important aspects unsaid, and asking questions could resolve the ambiguity — but models ask poor questions. STaR-GATE applies self-improvement (STaR) to question-asking itself: generate a synthetic dataset of 25,500 persona-task prompts simulating a Questioner conversing with a Roleplayer whose preferences are hidden; the Questioner asks questions to elicit preferences, and is then iteratively finetuned on the questions that increased the probability of high-quality responses (responses generated by an Oracle with access to the Roleplayer's latent preferences). After two iterations of self-improvement, the Questioner asks better questions and produces responses preferred over the initial model on 72% of tasks.

The keeper is that eliciting preferences is a trainable skill, improvable by self-play against simulated users — reward the questions that lead to better downstream answers, and question-asking improves without human-written question supervision. It targets the elicitation half of personalization that prompt-stuffing and persona-assignment skip.

This is a strong fit for Adrian's clarification/proactivity thread. It pairs with Can models learn to ask clarifying questions without explicit training? (emergent vs explicitly-rewarded question-asking) and addresses the deficit named by Why can't advanced AI models take initiative in conversation? — STaR-GATE trains the initiative that passive next-turn optimization suppresses.