Meta等提出上下文语言模型CLM：让模型自管上下文
原标题：Context Language Models: Self-Managing Context to Improve Performance and Reduce Compute Costs
AI 摘要
Meta、MIT 和华盛顿大学的研究者提出 Context Language Models（CLMs），让语言模型把上下文当作可自由编辑的文件自行管理，而非依赖预定义的摘要、压缩和检索机制。研究探索了零样本、上下文学习和强化学习三种方式，在 BrowseComp-Plus、EdgeBench 等基准上报告了准确率提升与 FLOPs 下降。研究者也承认存在信息丢失、提示注入等安全风险，社区评论则质疑其任务特定性和缓存失效等生产可用性问题。
正文节选
Researchers from Meta, MIT, and the University of Washington introduce Context Language Models (CLMs), a new approach that enables language models to manage and edit their own context rather than relying on predefined mechanisms for summarization, compression, and information retrieval, reporting substantial gains in both performance and computational efficiency. The researchers implemented the mechanism allowing LLMs to manage their own context by treating it as a file that they can update with