新智元 (微信公众号)T2·78 pts
AgentWorld Benchmark Reveals Top Models Achieve Only ~50% Success in Multi-Agent Collaboration Tasks
- AgentWorld focuses on actual collaborative performance over dozens of turns with distinct roles and resource constraints, rather than just dialogue quality.
- Experimental data reveals a major bottleneck: even the strongest models achieve only ~50% task completion rates in scenarios requiring active information exchange and coordinated division of labor.
- The study suggests future breakthroughs lie not just in enhancing individual model intelligence, but in optimizing inter-agent planning, memory sharing, and communication protocols.