A Prompt-Only Abstention Loop Cuts Wrong Commitments From 13.1% to 8.9% Across Eleven Model Families

arXiv 2609.17516·medium signalXBlueskyLinkedInCopy link

Chain-of-Self-Questioning (arXiv 2609.17516, submitted 15 Sep 2026) makes answer commitment conditional on an explicit assessment of what information the question requires, with no training involved. On the 817-item TruthfulQA multiple-choice validation set across eleven open-weight and hosted model families, Grounded-CoSQ at tau=0.90 cut mean unconditional wrong-commitment rate from 13.1% under chain-of-thought to 8.9%, a 32.1% relative reduction, while raising answered accuracy from 86.9% to 89.7% and still answering 87.6% of questions. The improvements held for all eleven models at every evaluated threshold, with Critical-CoSQ and Adaptive-CoSQ offering 88.6% and 86.5% coverage operating points.