Researchers steer language models away from unnecessary clarification questions

Researchers report that steering internal computations flipped 72-100% of unnecessary clarification questions in held-out tests across five eligible language models. Their preprint studied a controlled agent environment in which requests were already unambiguous, using intervention labels from a rule-based overseer validated against blind human judgment. The method located where each model decided whether to ask or act, then applied steering when an internal probe indicated a problem. Legitimate questions survived while simulated oversight needs fell. Only five of thirteen models qualified for the study, and attempts to embed the correction into model weights produced different outcomes across model families.