A new study finds that language models can make fewer unsupported commitments when they first identify what information an answer requires and decide whether that evidence is available. The prompt-only Chain-of-Self-Questioning method does not retrain the model; it makes answering conditional on an explicit self-assessment.
Researchers tested three variants under 17 conditions on the 817-question TruthfulQA multiple-choice validation set using 11 hosted and open-weight model families. In the final balanced-option protocol, Grounded-CoSQ at a 0.90 threshold reduced the average unconditional wrong-commitment rate from 13.1% with chain-of-thought prompting to 8.9%, a 32.1% relative reduction. Answered-question accuracy rose from 86.9% to 89.7%, while the models still answered 87.6% of questions. The improvement held for all 11 models and every tested threshold.
Other variants offered nearby trade-offs, with coverage of 88.6% and 86.5%. A Natural Questions short-answer experiment provided additional open-form evidence. The method does not establish that a model truly knows what it knows, and benchmark abstention may behave differently in specialized or adversarial settings. Its practical value is a configurable answer-or-review decision for applications where a referral or human check is cheaper than a confident but unsupported response.