论文:AI Agent易盲信陈旧记忆,大模型无法可靠修复
AI agents can trust stale memory over fresh evidence, and bigger models do not r…
Agent 记忆可靠性是落地关键痛点,这篇论文用 Qwen3 实测给出了不同规模模型的差异化修复方案,做 Agent 架构的同学值得参考。
AI agents can trust stale memory over fresh evidence, and bigger models do not reliably fix this.
AI 智能体可能会信任过时的记忆而非最新证据,而更大的模型并不能可靠地解决这一问题。
Persistent memory can make an agent confidently wrong even when current evidence is available, so stale facts should be resolved before they reach the model.
持久化记忆可能导致智能体在拥有当前证据的情况下依然自信地犯错,因此应在这些过时事实进入模型之前予以解决。
The paper tests Qwen3 models from 0.6B to 8B on tasks where stored memory is outdated.
该论文测试了 Qwen3 系列从 0.6B 到 8B 的模型在存储记忆已过时任务上的表现。
When memory is needed, they follow the stale value 92%–100% of the time.
当需要调用记忆时,它们有 92%–100% 的概率遵循过时的值。
And making an old note look newer can fool larger models even more.
而且,让旧笔记看起来更新可以进一步欺骗更大的模型。
Their fix depends on model size.
他们的修复方案取决于模型大小。
For 4B and 8B models, timestamps and source metadata are enough to recover most accuracy.
对于 4B 和 8B 模型,时间戳和来源元数据足以恢复大部分准确性。
The 0.6B and 1.7B models need the stale conflict resolved before they see it.
0.6B 和 1.7B 模型需要在看到冲突之前就解决过时信息的矛盾。
So the paper suggests: treat agent memory as untrusted input.
因此,该论文建议:将智能体记忆视为不可信的输入。
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