Heterogeneous Consensus-Progressive Reasoning for Efficient Multi-Agent Debate
New multi-agent debate method cuts token use by dynamically scaling collaboration based on task difficulty.
A research team led by Yiqing Liu and Hantao Yao has published a new paper introducing Heterogeneous Consensus-Progressive Reasoning for Efficient Multi-Agent Debate (HCP-MAD). This framework tackles a major inefficiency in current Multi-Agent Debate (MAD) systems, where multiple AI agents collaborate through iterative reasoning and critique cycles. Existing methods often apply the same intensive debate process to all tasks, leading to prohibitively high token costs (and thus expense) regardless of whether a problem is simple or complex.
HCP-MAD's core innovation is a three-stage, adaptive reasoning mechanism. First, Heterogeneous Consensus Verification uses a pair of diverse agents to quickly check for agreement, allowing easy tasks to be solved immediately. For moderately complex tasks, the system moves to a Heterogeneous Pair-Agent Debate, where two agents critique each other's reasoning with a dynamic stopping criterion. Only the most difficult, unresolved problems escalate to a final stage of Escalated Collective Voting, which aggregates perspectives from additional agents. This progressive approach ensures computational resources are matched to problem difficulty.
The paper, submitted to arXiv, reports that HCP-MAD achieves higher accuracy across multiple benchmarks while substantially reducing the token consumption typical of MAD systems. By making collaborative AI reasoning more cost-effective, this work paves the way for more practical and scalable applications of multi-agent frameworks in complex problem-solving domains.
- Uses a three-stage process (consensus check, pair debate, collective vote) to adapt computational effort to task difficulty.
- Dynamically scales collaboration, using lightweight 2-agent debates for simple tasks and expanding only for complex ones.
- Demonstrated in experiments to significantly improve accuracy while drastically cutting token usage and associated costs.
Why It Matters
Makes advanced multi-agent AI collaboration viable for real-world use by slashing the computational cost of complex reasoning tasks.