τ²-bench

What it tests: Tool-calling agent evaluation in realistic customer-service domains (airline, retail, telecom). The agent must use APIs, follow policies, and handle multi-turn user interactions.

考什么:仿真客服场景的工具调用评测(航空/零售/电信):agent 必须正确调 API、遵守业务规则、处理多轮用户对话。

Why it counts: The most practical agent benchmark — it predicts whether a model will reliably drive tools in production, which raw chat benchmarks can't tell you.

为什么算数:最实用的 Agent benchmark——能预测模型在生产环境调工具的可靠性,这是聊天类 benchmark 看不出来的。