AgentWorld: Benchmarking Long-Horizon Collaboration of Multi-agent LLMs
Existing multi-agent benchmarks primarily test in competitive settings, short-horizon interactions under 20 steps, or simply aggregate individual performance, failing to isolate and highlight genuine collaboration capabilities of LLM-based agents. We introduce AgentWorld, a benchmark of 100 human-annotated tasks (with 100 augmented variants) for evaluating long-horizon, multi-a
Reading in Plain English mode — toggle below to switch
Existing multi-agent benchmarks primarily test in competitive settings, short-horizon interactions under 20 steps, or simply aggregate individual performance, failing to isolate and highlight genuine collaboration capabilities of LLM-based agents. We introduce AgentWorld, a benchmark of 100 human-annotated tasks (with 100 augmented variants) for evaluating long-horizon, multi-a
As multi-agent systems become more complex, robust benchmarks are crucial for evaluating their collaborative capabilities beyond simple competitive tasks. This work provides a standardized way to assess long-horizon, multi-agent collaboration, which is a key challenge in developing autonomous AI systems.