Description
The paper by Rulin Shao, Mike Lewis, Luke Zettlemoyer, Pang Wei Koh and colleagues introduces Context Language Models, which manage their own context by treating it as a file they can update, so the model decides which information is worth keeping. The authors report 11.4% higher accuracy with 21.5% fewer FLOPs on BrowseComp-Plus, show that the approach works in multi-agent systems, and cover both in-context and parametric learning of context-management strategies. They also present an online reinforcement learning method and Suffix Cache Reuse, which cuts server-side compute by 35% compared with standard methods. A good read for anyone working on long-horizon agents and context engineering.
