2609.20543R·cs.AI18 de set. de 2026Language-model groups overstate consensus when replaying human deliberation on a reasoning task Shao