Salesforce研究：自改进Agent记忆易累积错误导致性能不稳定
The big problem with self-improving agents is that memory can compound mistakes…
Agent记忆机制的脆弱性被量化验证，对构建可靠自改进系统有直接参考价值，建议Agent开发者关注。
The big problem with self-improving agents is that memory can compound mistakes just as easily as it compounds useful lessons.
自我改进代理面临的一个大问题是，记忆既会像积累有用经验一样，轻易地累积错误。
Salesforce tested 2 memory-based agent methods and found a pretty uncomfortable pattern.
Salesforce 测试了两种基于记忆的代理方法，发现了一个相当令人不安的模式。
With WebArena’s default task order, ReasoningBank improved performance by 1.5 points.
在使用 WebArena 的默认任务顺序时，ReasoningBank 使性能提升了 1.5 分。
Shuffle those tasks, and it dropped by 4.5 points instead.
但如果打乱这些任务的顺序，性能反而下降了 4.5 分。
Why? The default order tended to put easier tasks first, so the agent learned cleaner lessons early.
为什么？默认顺序倾向于将较简单的任务放在前面，因此代理在早期就能学到更清晰、更干净的经验。
But memory works both ways.
但记忆的作用具有两面性。
Agents also saved bad lessons, like recommending APIs in an environment where APIs were impossible, then kept pulling those memories back into future tasks.
代理也会保存错误的经验，例如在一个根本不可能使用 API 的环境中推荐 API，随后又在后续任务中反复调用这些记忆。
Results became more unstable in 71% of cases.
在 71% 的案例中，结果变得更加不稳定。
Even giving the memory system better task details and environment feedback recovered only 31% of the drop.
即使为记忆系统提供更详细的任务信息和环境反馈，也仅能恢复 31% 的性能损失。
So “learning from experience” is only useful if the agent is learning the right thing.
因此，“从经验中学习”只有在代理学到正确内容时才有效。
– arxiv. org/abs/2608.18066
– arxiv.org/abs/2608.18066
Title: "On the Fragility of Self-Improving Agents: Variance, Task Order, and Underspecification"
标题：《论自我改进代理的脆弱性：方差、任务顺序与规范不足》
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力