Multi-Agent Debate

Multiple LLM instances propose their own answers and critique each other's reasoning processes. Key finding: even when all agents are initially wrong, debate can still produce the correct answer.

2024-2025 advances include A-HMAD, DMAD (ICLR 2025), sparse topology (41% token cost reduction), S2-MAD (94.5% token cost reduction). 17 related papers published in 2024 alone.