AI上下文压缩丢失用户指令，小型模型插件可修复
原标题：AI systems quietly drop user instructions when they compress context
AI 摘要
宾夕法尼亚州立大学的研究人员发现，AI系统在压缩上下文以释放空间时，平均只有17%的用户会话约束能保留下来，导致AI可能忽略用户明确设定的规则，如“未经批准不得发送邮件”。他们推出了名为COMPINT的评估套件和一个基于Qwen3.5-9B的小型附加语言模型，该模型能提取并保留用户约束，在测试中实现了超过90%的保留率，且无需训练或修改压缩系统。该工具已在GitHub上发布。
正文节选
AI systems quietly drop user instructions when they compress context When AI systems summarize their context to free up space, user constraints get lost along the way. On average, only 17 percent of instructions survive compression. A small add-on LLM can fix most of the problem. The context window of AI models becomes a bottleneck when users run long conversations without starting a new chat. We've covered why context management matters for output quality, but that's out of reach for everyday u