Recently, researchers from Tsinghua University, Ohio State University and University of California, Berkeley designed a test tool - AgentBench, to evaluate the reasoning and decision-making ability of LLM in a multi-dimensional open generation environment . The researchers conducted a comprehensive evaluation of 25 LLMs, including API-based business models and open-source models.
They found that top commercial LLMs exhibit strong capabilities in complex environments, and that top models like GPT-4 can handle a wide range of real-world tasks, significantly outperforming open-source models. The researchers also said that AgentBench is a multi-dimensional dynamic benchmark test, which currently consists of 8 different test scenarios, and will cover a wider range in the future to conduct a more in-depth systematic evaluation of LLM.
▷ Source: arXiv official website