Primary sourcearXiv cs.AI
View sourceEvent ·Single source
E-Bench:多步工具使用智能体的真实产品基准
E-Bench是一个完全合成的基准,包含323个状态改变任务,覆盖Honor of Kings、QQ Music和Tencent Meeting三个产品领域。它通过图引导数据库填充创建可复用环境,并利用生成器-求解器不对称设计迫使智能体发现隐藏信息并组合多次工具调用。结果通过数据库状态差异确定性评分。评测11个前沿LLM,最强模型的Pass^3低于60%,在E-Bench-Code扩展下仍低于70%,说明多步工具使用对当前模型依然困难。
Current read
想知道多步工具调用到底多难?E-Bench测了11个模型,最强不到60%通过率,还特意选了王者荣耀、QQ音乐和腾讯会议场景。
2 sources65° AI heatLast updated 2026/7/26 15:38:28
Evidence
2 sourcesRelated source 1Pandaily
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.