上下文语言模型技术框架提升模型效能与计算效率

研究团队提出了上下文语言模型（CLMs）技术框架，使语言模型能够原生管理自身上下文1。该方法通过将上下文视为可更新的文件来实现，允许模型对其进行无限制修改1。

在多个基准任务的测试中，CLMs框架展现出显著优势。在BrowseComp-Plus基准上，模型精度提升11.4%的同时计算量减少21.5%1；在12小时EdgeBench任务中准确率提高5%，计算量下降59%1；在24小时多仓库智能体集群任务上性能提升65%，计算成本保持不变1。特定模型方面，Qwen3.5-9B在BrowseComp-Plus上性能提升47.6%，计算量减少12%1。

研究还引入了在线强化学习方法和缓存复用技术以进一步优化性能。团队开发的Suffix Cache Reuse技术使服务端计算减少35%1，而通过自然语言指令进行的技能优化使上下文管理任务准确率提升35.9个百分点1。

Researchers have introduced Context Language Models (CLMs), a novel framework that enables language models to natively manage their own context 1. Unlike traditional approaches that treat context as static data, CLMs view context as updateable files, allowing models to modify and maintain information dynamically throughout processing 1.

The framework demonstrates significant improvements across multiple benchmark tasks while substantially reducing computational requirements 1. On BrowseComp-Plus, CLMs achieved an 11.4% improvement in accuracy while reducing computational load by 21.5% 1. For the 12-hour EdgeBench benchmark, the approach increased accuracy by 5% with a 59% reduction in computation 1. In a 24-hour multi-repository agent cluster task, performance improved by 65% while maintaining the same computational cost 1. The research also introduced online reinforcement learning methods and a Suffix Cache Reuse technique that decreased server-side computation by 35% 1.

Additional innovations include context management optimization through natural language instructions, which enhanced task accuracy by 35.9 percentage points 1. When applied to Qwen3.5-9B, the CLM approach yielded a 47.6% performance improvement with a 12% reduction in computational requirements 1. These results suggest that native context management represents a more efficient alternative to existing context handling strategies in language models 1.