伯克利新论文发现 LLM 常能察觉用户偏好或截止时间已变，却仍沿用旧值，因为注意力持续漂回上下文中的旧提及。在 5 个开源模型中，无需重训练、仅将注意力引导至最新值即可修复大部分此类错误；GPT-5.6 Sol 在长智能体日志上仅答对 40 题中的 9 题，给出当前状态后则 40 题全对。论文建议智能体追踪可变信息时，直接把当前状态写进提示词，而非让模型翻查历史。
New Berkley paper: LLMs often know you changed your mind but still use your old choice, so agents need the current state spelled out.
When a preference or deadline changes, the old version stays in context. The model still holds the new one, but its attention keeps drifting back to older mentions.
In 5 open models, nudging attention toward the newest value fixed most of these mistakes without retraining. Even top-tier GPT-5.6 Sol got only 9 of 40 questions right on long agent logs, but 40 of 40 when given the current state.
If your agent tracks anything that changes, keep the current state in the prompt instead of making the model dig through history.
– arxiv. org/abs/2609.38866
Title: "When Context Changes: Understanding Update Failures in LLMs"
来源：Rohan Paul · x.com