模型58
Arena 推出六个面向真实任务的智能体指数
原标题:Alongside the Agent Arena Index and Alignment Index, we’re introducing six new focused indexes built from real tasks across millions of a...
Arena 在 Agent Arena Index 和 Alignment Index 之外新增六个聚焦指数,基于数百万次智能体会话中的真实任务构建,分别回答知识工作、编码、日常聊天、任务完成、工具使用和跟随用户纠正各用哪个模型。
中文译文尚未完成,先为你展示原文。
Alongside the Agent Arena Index and Alignment Index, we’re introducing six new focused indexes built from real tasks across millions of agentic sessions on Arena. Each index answers a specific question to help you pick a model for your use case:
- Work: Which agent is the best at knowledge work?
- Coding: What’s the best coding agent?
- Chat: Which model is the best everyday chat agent?
- Task Outcome: Which model is the best at completing user tasks?
- Tool Use: Which model is the best at using tools?
- Steering: Which model is the best at following user corrections?
The Work, Coding, and Chat Indexes use all four agent performance signals. The Task Outcome Index focuses on Task Completion and Praise vs. Complaint, the Tool Use Index focuses on Bash Reliability, and the Steering Index focuses on Steerability. In the future, we’ll continue to add signals to these indexes to measure the frontier as it evolves.
As of today, Claude Fable 5.2 is #1 in the Agent Arena Index as well as the Work, Coding, Chat, Task Outcome, and Steering Indexes. Claude Sonnet 5 takes the top spot in the tool use index, and GPT 6.1 Terra is #1 in the Alignment Index.
同一件事的其他报道3 篇
Arena 推出 Agent Arena Index 智能体排行榜
Arena.ai · 原文
Arena.ai 推出 Agent Arena Index 及六个细分指数
Arena.ai · 原文
