What Does Multi-Agent LLM Debate Actually Change? A Layered Analysis of Disagreement and Answer Quality

Multi-agent debate, in which several LLMs exchange arguments before producing an answer, is widely assumed to improve answer quality by surfacing genuine disagreement. That disagreement is hard to verify, and no single signal can settle it, so we organize the analysis around four questions: (A) does