Unverified50% confidenceFactExact time
结果层的核心指标是任务整体成功率（Task-level Success）
1
Sources
50%
Confidence
Long-term
Relevance
7/17/2026
First Seen
Sources
Related Claims
UnverifiedExecution-based Reward signals derived from unit test execution results are significantly effective in improving code correctness rates.69% similarUnverifiedThe focus of evaluation is shifting from the quality of model text output to the ability of agents to complete end-to-end tasks, with evaluation dimensions expanding from a single accuracy metric to multi-dimensional metrics including task completion rate, step efficiency, error recovery capability, and time overhead.69% similarUnverifiedOpenAI内部测试数据显示，经过思维链分解的提示词在结构化任务上的准确率平均提升约30%-40%66% similarUnverified评估指标选择需与业务目标对齐，客服Agent优先任务完成率，代码生成Agent则更关注语法正确率和可运行率65% similarUnverified将反馈数据与具体执行上下文精确关联是定位问题根源和针对性优化提示词的关键65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/541551API
curl https://kongchang.com/api/v1/knowledge/claims/541551MCP
get_claim(id=541551)