- Multi-agent debate now runs leaner and smarter.
Heterogeneous Consensus-Progressive Reasoning for Efficient Multi-Agent Debate (HCP-MAD) was unveiled on arXiv. The authors replace the usual static debate topology with three stages: a quick consensus check by two heterogeneous agents, an adaptive pair‑agent critique that stops early when agreement is reached, and an escalated collective vote that brings in extra agents only for the hardest cases. Experiments on six standard benchmarks show higher accuracy than prior MAD systems and a noticeable drop in token usage. The code has been released on GitHub.
The change matters because token budgets dominate the cost of large‑language‑model deployments. By letting simple queries resolve after a brief pair‑wise exchange, HCP-MAD avoids the blanket overhead of full‑scale multi‑agent loops. For tougher problems it still gathers diverse opinions, preserving the original reasoning benefits. This adaptive scaling could make debate‑style AI assistants viable in real‑time settings where latency and cost are tight.
If the method holds up across more domains, it may push the field away from one‑size‑fits‑all debate architectures toward a more economical, task‑driven collaboration model.