DEAR framework cuts token use while curbing blind conformity in multi-agent AI debates

Multi-Agent Debate (MAD) has emerged as a popular technique to boost LLM reasoning by having multiple models argue and refine answers. But a known failure mode is blind conformity—agents simply echo the majority or the most confident peer, even when that peer is wrong. Existing fixes that measure individual confidence or perplexity don't capture reasoning correctness and can actually reinforce the herd mentality. In a new arXiv paper (2608.03648), Hao Wu and colleagues at Beijing Jiaotong University take a different angle: look at the group, not the individual.

The team defines 'Debate Relationships' as the mutual referencing patterns between agents. Their framework, DEAR (Dynamically regulating debate relationships), first quantifies consensus and divergence as 'group evidence' to snapshot the debate state. It then runs three steps: 'What'—assess the group's consultation tendency and uncertainty; 'Who'—use a Selection RL-Agent to dynamically pick which peers to reference; and 'How'—a Behavior RL-Agent adjusts generation strategies (e.g., how confidently to argue). These two agents are trained jointly via multi-agent reinforcement learning as a sequential decision-making problem. Results show DEAR outperforms baseline MAD approaches on reasoning tasks while cutting token overhead substantially—making it both more accurate and more efficient for real-world agentic systems.