AI 研究

Anthropic and Accenture extend evaluation beyond model outputs to training, deployment, monitoring, and incident processes, with independence unresolved.

Anthropic 与 Accenture 的 embedded evaluator 计划把评测范围从模型输出扩展到训练、部署、监控和事故流程，也暴露出独立性边界。

Use When Agents Slow Down, Reflexion, Hyperband, and agent evaluation to compare continued revision with independent attempts under reproducible budgets.

结合 When Agents Slow Down、Reflexion、Hyperband 和 Agent 评测，说明怎样比较连续修改与独立尝试，并设计可复现的预算分配实验。

The Anthropic incident review shows why automated monitors and LLM graders need coverage, independence, and tests for their own failure modes.

从 Anthropic Agentic Search 事故复盘看自动监控和 LLM grader 的漏检问题，说明 Agent Eval 也需要覆盖率、独立性和失效模式测试。