ToolBench
An AI agent evaluation benchmark focused on tool-calling tasks, using single-task success rate as the core metric to assess a model's ability to invoke external tools.
Timeline (last 90 days)
Agent directories are an infrastructure component of the AI agent economy, cataloging tools, services, or marketplaces that AI agents can invoke, with common examples including the OpenAI plugin store, the Anthropic MCP server directory, and third-party projects such as AgentHub and ToolBench.
Most test scenarios in AgentBench and ToolBench are still pre-scripted, with a significant gap from the uncertainty of tool return values and diversity of errors encountered in real production environments
WebArena, ToolBench, and AgentBench are the current mainstream AI Agent benchmarks, focusing on web navigation, tool invocation, and multi-environment comprehensiveness, respectively.
UC Berkeley 研究团队在 ToolBench 基准上发现，模型在复杂工具调用链中的幻觉率可达 15-30%
All Facts (5)
Agent directories are an infrastructure component of the AI agent economy, cataloging tools, services, or marketplaces that AI agents can invoke, with common examples including the OpenAI plugin store, the Anthropic MCP server directory, and third-party projects such as AgentHub and ToolBench.
50%UnverifiedMost test scenarios in AgentBench and ToolBench are still pre-scripted, with a significant gap from the uncertainty of tool return values and diversity of errors encountered in real production environments
50%UnverifiedWebArena, ToolBench, and AgentBench are the current mainstream AI Agent benchmarks, focusing on web navigation, tool invocation, and multi-environment comprehensiveness, respectively.
50%UnverifiedUC Berkeley 研究团队在 ToolBench 基准上发现，模型在复杂工具调用链中的幻觉率可达 15-30%
50%UnverifiedToolBench专注于工具调用能力，包含16000余个真实API
50%