Stopping AI reasoning early when answers agree is unsafe

Original: Stable Answers, Unfinished Reasoning: Why Self-Consensus Is Not a Safe Early-Exit Signal

Source: arxiv.org ↗